Human-written content outperforms AI-generated articles by up to 2.5 times in search visibility

A Russian digital agency ran an eight-month search experiment designed to answer one question cleanly: when a search engine has nothing else to go on but the text itself, does it treat human-written copy differently from AI-generated copy, and if so, by how much and for how long. The agency, AINET, built ten brand-new websites, gave five of them articles written by professional human experts and five of them articles generated by AI systems, and tracked how each one performed in Yandex and Google for eight months straight.

The headline finding, stated plainly

By the end of the experiment, the human-written pages held 71.4% of search visibility in Yandex against 28.6% for the AI-generated pages, and 67.3% in Google against 32.7%. At points during the test the gap reached 2.5 times in Yandex’s favour for human content. The AI-written pages did not fail outright — one of them, generated by the agentic AI system Manus, finished the experiment ranking first in Yandex and third in Google. But as a group, the AI-generated texts lost ground steadily after an opening period in which they performed competitively, and in one case briefly ahead.

That arc — a fast AI start, a false peak, then a slow and fairly thorough reversal — is the part of this study worth taking seriously, because it is not what either side of the AI-content argument usually predicts. People who dismiss AI writing outright expect it to underperform from day one. People who argue AI content is “just as good” point to early performance data and stop looking. This experiment ran long enough to catch both claims being half right and half wrong at different points in the same eight months.

The experiment controlled for almost everything except the text itself. Ten domains, same age, same hosting, same design, same single invented keyword, launched on the same day. That design choice is what makes the result worth more than the dozens of looser AI-versus-human content comparisons that circulate in SEO forums, most of which compare existing pages with different backlink profiles, different domain histories and different publication dates, and then draw conclusions about content quality that the data cannot actually support.

The numbers belong to one market, one query, one content vertical and one eight-month window, and this analysis treats that scope honestly rather than inflating it into a universal law about artificial intelligence and search. But the pattern it documents — AI content that starts competitively and decays, against human content that starts strong and holds — lines up with larger, separately conducted web-scale research that this piece also examines later. Two independent measurement efforts, built differently and run by different organisations, converge on the same shape of result. That convergence, more than any single percentage, is the reason this story matters to anyone who publishes content for a living.

Who ran this experiment and what they were trying to isolate

AINET is a Russian digital agency with an SEO/GEO practice, and the study came out of its own department rather than from an outside research lab. The agency’s head of SEO/GEO, Yaroslav Dombrovsky, led the analysis, and its founder and chief executive, Ruslan Minyazhetdinov, supplied the business framing when the results were shared with the press. Russian marketing outlets ppc.world, CNews and Comnews carried the findings at the end of October 2026, and ppc.world’s writeup, sourced directly from AINET’s press office, is the most complete public account of the methodology.

The question AINET set out to answer was narrower than “is AI content good or bad.” Plenty of studies already compare AI and human writing on grammar, factual accuracy or reader preference, and plenty of agencies already have informal data on how their own AI-assisted pages perform. What none of that data can cleanly separate is the effect of the text itself from everything else that normally differs between an AI-written page and a human-written one: which domain it sits on, how old that domain is, what links point to it, what else is published nearby, and how search engines have already learned to trust or distrust the publisher.

Isolating the writing from the publishing context required removing every other variable AINET could think of. A page written by a seasoned expert but published on a site search engines already treat as low-quality will underperform regardless of the prose. A thin AI-generated page sitting on a decade-old, well-linked domain with strong existing rankings will often perform fine regardless of how it was produced, at least for a while, because domain-level trust signals can paper over page-level thinness. Neither comparison tells you anything about the writing. AINET’s answer was to build the publishing context from scratch, identically, ten times over, so that the only variable left standing between any two of the ten sites was who or what wrote the words.

This is, in effect, a blinded trial design borrowed from a different discipline and applied to search ranking — the kind of controlled comparison that is common in clinical research and almost never attempted in SEO, where most published “studies” are correlational analyses of existing search results rather than prospective experiments with a built-in control group. The rarity of that design is itself worth noting. Search engine optimisation produces an enormous volume of commentary and very little experimental evidence, because running a true controlled test means spending real money launching domains and waiting months for data, with no guarantee the result will be clean or will arrive before someone else publishes a similar finding. AINET committed eight months and real infrastructure to get an answer it could stand behind.

Building a controlled environment: ten domains, zero history

The experiment’s backbone was ten freshly registered domains with no prior content, no inbound links and no digital footprint of any kind — specifically, no trace in the Internet Archive’s Wayback Machine, which search engines and SEO tools alike treat as a rough proxy for a domain’s age and prior use. A domain that shows up in Archive.org snapshots from years earlier, even under a different owner or topic, carries residual signals that can distort a ranking test. AINET screened that out at the registration stage.

All ten sites were built on the same server, launched on the same day, and designed identically: same layout, same HTML and CSS, no content management system styling differences, no theme variation that might itself carry an SEO signal through faster load times, different schema markup or accidental technical advantages. The only engineered difference between any two of the ten domains was the article published on it. Five got text from professional human writers with subject expertise in the chosen topic; five got text generated by AI systems working from an identical prompt.

That level of control matters because search ranking algorithms weigh dozens of signals beyond the text on the page — site speed, mobile usability, structured data, internal linking, HTTPS configuration, and more — and any uncontrolled difference in those signals becomes a confound that undermines the comparison. A faster-loading AI-content site beating a slower human-content site would prove nothing about writing quality. AINET’s design closes off that possibility by making the technical layer identical across all ten properties, which is the methodological detail that separates this from the far more common practice of comparing two unrelated websites and attributing the ranking gap to content type alone.

The sites were optimised identically for search as well. Same title tag structure, same meta description approach, same on-page formatting conventions applied uniformly whether the underlying text came from a person or a model. This removes a second common confound in amateur content comparisons: the tendency for whoever favours one approach to also, consciously or not, put more SEO polish into the pages that support their preferred conclusion. AINET applied one optimisation standard across all ten properties before letting the text itself do whatever work the text was going to do.

The nonsense keyword that made the test possible

The single cleverest design choice in the experiment was the target query itself: “lipihydroration” (транслитерация: “липигидродация”), a term that does not exist in the Russian beauty and skincare vocabulary, has no established meaning, no competing content, and no search history. AINET invented it specifically for this test.

A made-up query solves a problem that would otherwise be unsolvable with real keywords. Any existing, searched term already has a search engine results page populated by established sites with their own domain authority, backlink profiles and ranking history. Launching ten new domains to compete for a real keyword means competing against that entrenched field as well as against each other, which muddies any comparison between the ten test pages themselves — some of the variance in their performance would come from how they stack up against outside competitors, not from how they stack up against each other.

With a nonsense keyword, there was no outside competition. The ten pages were, for all practical purposes, the entire search results page for that query. Whatever ranking order Yandex and Google produced among those ten pages reflected a pure, closed comparison: ten pieces of content, identical technical wrapper, nothing else to rank against. The query’s artificiality is precisely what makes the resulting data trustworthy rather than a limitation of it.

This approach has precedent outside SEO. Researchers testing search or recommendation algorithms often use synthetic items or invented terms for exactly this reason — it is the only way to study ranking behaviour without the confound of pre-existing competitive context. The tradeoff, which this analysis returns to later when discussing the study’s limits, is that a made-up query without real search demand behind it is not identical to a real commercial keyword with genuine user intent, click-through behaviour and engagement data feeding back into the ranking systems. AINET bought methodological cleanliness at the cost of real-world messiness, which was the right trade for the specific question it was asking, and a trade that needs to be named rather than glossed over when reading the headline numbers.

Five professional writers against five AI systems

The human side of the test came from five professional beauty-industry experts, writing original articles on the invented topic under the same brief. The AI side came from five of the most widely used generative systems on the market at the time of the test: Manus, ChatGPT, Perplexity, Gemini and DeepSeek, each generating its article from a single shared prompt designed to produce the same kind of content the human writers were asked for.

That list of five AI systems spans a genuinely varied set of architectures and products, which matters for how far the result generalises. ChatGPT and Gemini are general-purpose conversational assistants from OpenAI and Google respectively, built primarily for dialogue and broad task completion. Perplexity is positioned specifically as an answer engine that blends retrieval with generation, built around citing sources as it answers. DeepSeek is a Chinese-developed open-weight model family that gained wide attention in global markets for matching frontier performance at lower training cost. Manus is different in kind from the other four — it is an agentic AI system designed to complete multi-step tasks autonomously rather than a single-turn chat assistant, which likely explains why its output behaved differently from the other four AI-generated texts throughout the experiment, a pattern this analysis examines in detail further down.

Using five different systems rather than one is a real strength of the design. A test that only compared human writing against, say, ChatGPT alone would tell you something about one model’s output quality, but nothing about whether the pattern holds across the current field of generative tools people actually use for content production. By spreading the AI side across five systems with different training approaches, different retrieval capabilities and different output styles, AINET reduced the risk that its result was really a story about one underperforming model rather than a story about AI-generated text as a category.

It also means the five AI texts were not identical to each other in style or structure, even though they answered the same prompt — which is visible in how differently they performed over the eight months. A single-model test would have produced a tighter, more homogeneous AI-text group and a less informative result about the category as a whole.

The first weeks: an immediate and lopsided human lead

In Yandex, the opening weeks produced an almost complete sweep: human-written pages took all five of the top positions, and the visibility split was 72.7% for human content against 27.3% for AI-generated content. Google told a different story at the same point in time — the two groups were nearly even, with AI-generated pages holding a narrow edge at 50.9% against 49.1% for human content.

That split between the two search engines in the opening phase is itself informative. Yandex and Google use different ranking architectures, different quality-signal weightings, and in Yandex’s case a search engine built and tuned specifically around the Russian-language web, where a beauty-industry prompt executed by five different AI systems might carry more easily recognisable linguistic tells than in a market with a larger base of AI-generated training and output data. Whatever the mechanism, Yandex identified a quality gap between the two content types within weeks, while Google’s early ranking treated the two groups as roughly interchangeable.

This divergence matters because it undercuts any simple story about AI detection. If search engines simply ran AI-generated text through a classifier and demoted anything that scored above a threshold, both engines should have produced similar early results, since the underlying text was the same regardless of which search engine was indexing it. The fact that Yandex and Google started from different places suggests the early-stage ranking gap, where it existed, came from something closer to ordinary quality signals — engagement, structure, the match between content depth and reader intent — than from an explicit AI-content filter running at the point of indexing.

The early Google numbers are also the detail that gives AI-content advocates their best talking point, and it is worth stating plainly rather than minimising: for a genuine stretch of real time, on a real search engine, AI-generated text in this experiment performed exactly as well as human-written text, and briefly better. Anyone who had stopped measuring at week three would have reasonably concluded that the writing source made no difference in Google. The entire value of running the experiment for eight months rather than three weeks is that this early read turned out to be incomplete rather than wrong — a snapshot of behaviour that changed as the test went on, not a final verdict.

Four months in: the reversal nobody at the agency expected

By the four-month mark, the pattern from week one had inverted in both search engines, in opposite directions from where each had started.

In Yandex, the initial human dominance had nearly dissolved. The gap between the two groups shrank close to even, with AI-generated content’s visibility share climbing to 45.5%, and two of the five AI-written pages — generated by Manus and Gemini — moved into the top two ranking positions for the test keyword, ahead of every human-written page. In Google, the movement went further still: AI-generated content’s visibility share rose to 63.6% against 54.5% for human-written content, and by this checkpoint only one of the five human-written pages remained in the top five results at all.

That is a genuinely striking mid-experiment result, and it is the point in the data where, if AINET had ended the test, the conclusion would have run in the opposite direction from the one the full eight months eventually supported. AI-generated content, roughly four months after publication, was outperforming professionally written expert content in both of the search engines under study. Any honest account of this experiment has to sit with that fact rather than skip past it on the way to the eventual reversal, because it is exactly the pattern that explains why so many publishers and agencies, looking only at early-to-mid performance data from their own AI content pushes, have concluded that AI-generated articles work just as well as human ones.

Dombrovsky’s own explanation for why this mid-point exists, discussed in more detail later in this piece, centres on a lag between when AI-generated text is published and when ranking systems accumulate enough signal to classify it as secondary or derivative. In the gap between those two points, AI content can genuinely outperform — not through trickery, but because nothing in the available signal yet distinguishes it from original material, while it may simultaneously benefit from being cleanly structured, keyword-matched, and free of the small inconsistencies that sometimes creep into rushed human drafts. The four-month data point in this experiment is the clearest documented instance of that window actually showing up in live ranking data, across two independent search engines, rather than being argued from theory.

Eight months in: the swing back toward human text

By the end of the experiment, the ranking order had reversed again, and this time more decisively than at any earlier checkpoint.

In Yandex, human-written pages occupied positions two through six, taking 71.4% of search visibility with an average position of 4. The AI-generated group fell to 28.6% visibility with an average position of 8.2 — meaning the typical AI-written page by this point was sitting almost a full page further down the results than the typical human-written one. In Google, human-written content reached 67.3% visibility at an average position of 3.6, while AI-generated content held 32.7% at an average position of 7.4, with only one of the five AI-generated pages still holding a top-six spot.

The magnitude of the swing is the headline number, but the shape of the swing is the more useful finding. This was not a steady decline for AI content from week one onward, which would support a simple “AI content never worked” narrative, nor was it a steady climb for AI content that the experiment simply cut off too early, which would support the opposite narrative. It was a trough-to-peak-to-trough pattern for human content and the mirror image for AI content, concentrated in the same eight-month window, measured identically throughout.

That U-shaped recovery for human-written content, after four months of underperformance, is the data point most likely to get lost in any quick summary of this study, because “human content wins by 2.5x” is a cleaner headline than “human content lost the middle of the race and still won by the final lap.” But the middle section is exactly what makes the ending credible rather than foreordained. If human-written content had simply led from start to finish, a skeptic could reasonably argue the test had an undetected confound favouring human pages from the outset — something about the writers’ prompt adherence, keyword density or formatting that happened to align with ranking signals regardless of authorship. The fact that AI content took a real, measured lead in the middle of the test and then lost it on the same even footing is harder to explain away as an artefact of the setup, and easier to explain as an actual property of how search engines evaluate the two content types over time.

Average position as the number that matters more than visibility share

Visibility share — the percentage of total available visibility a group of pages captures — is the figure that made headlines, but average ranking position is the number that translates most directly into real-world consequence, because click-through rate on a search results page drops sharply and non-linearly as position moves down the page. A page averaging position 3.6, as the human-written group did in Google by month eight, sits in territory that still captures a notable share of clicks from anyone scanning the results. A page averaging position 7.4, where the AI-generated group landed in the same market at the same time, sits well past the point where most searchers stop scrolling on a typical results page, particularly once AI Overviews, featured snippets and other result-page features push organic listings further down the visible fold.

The gap between position 4 and position 8.2 in Yandex, or between 3.6 and 7.4 in Google, is not a gap between “doing well” and “doing slightly less well.” It is closer to a gap between “capturing most of the available organic traffic for this topic” and “capturing a small fraction of it, if any.” For a commercial website, that is the difference between a page that earns back its production cost and a page that functions as digital inventory nobody finds.

This is also where the raw visibility-share percentages can understate how lopsided the real-world outcome was. A 71.4% versus 28.6% split in visibility share sounds like the AI-generated group still captured more than a quarter of the available attention. An average position of 8.2 against 4 tells a sharper story: the AI-generated pages were not evenly distributed just below the human-written ones, clustered at positions 5 through 9. They were scattered further down the results, likely with at least one or two pages buried on a second results page entirely, dragging the average down while a lone strong performer — Manus’s text, discussed next — pulled the AI group’s visibility share up from what it would otherwise have been. Reading visibility share and average position together, rather than either number alone, is the right way to interpret what actually happened to these ten pages over eight months.

Manus, the one AI text that never lost its footing

Every account of this experiment, from AINET’s own press materials through to the trade coverage that followed, singles out one exception to the overall AI-content decline: the article generated by Manus, which finished the eight-month test in first place in Yandex and third place in Google — ahead of every other AI-generated text, and ahead of most of the human-written ones as well.

Manus is architecturally distinct from the other four AI systems tested. Where ChatGPT, Gemini, Perplexity and DeepSeek in their standard modes generate a response in a single pass from a prompt, Manus operates as an agentic system designed to plan and execute multi-step tasks, which can include browsing, drafting, revising and structuring content in a more deliberate, iterative sequence than a one-shot chat completion. That operational difference is the most plausible explanation for why its output behaved differently from the other four AI-generated texts in this experiment, though the study itself did not isolate the specific mechanism, and AINET’s public materials do not go further than naming Manus as the exception.

What this single data point should and should not be read as is worth being precise about. It is not evidence that “AI content works fine, you just need the right tool” as a blanket conclusion — four of the five AI systems tested, including some of the most widely used generative tools on the market, produced content that declined in the same window where the Manus output held its ground or improved. One success case out of five AI systems is informative about variance within the category, not a refutation of the category-level pattern. It is, however, a genuine and useful finding in its own right: it suggests that whatever property search engines are detecting and penalising in the four underperforming AI texts is not an immutable feature of “AI-generated content” as an undifferentiated class, but something closer to a production-process variable — depth of iteration, structural planning, perhaps something resembling editorial revision built into the generation pipeline — that varies between AI systems and could, in principle, be replicated deliberately by anyone using AI tools with more care than a single prompt-and-publish pass. The Manus result is the strongest piece of evidence in this entire study that “AI-assisted, carefully produced” and “AI-generated, quickly produced” are functionally different categories for ranking purposes, even though both fall under the same casual label.

Yandex and Google told almost the same story on different schedules

Running the experiment across two search engines simultaneously, rather than just one, lets AINET’s data answer a question that a single-engine study could not: is this a property of one company’s ranking algorithm, or something closer to a general feature of how large-scale search ranking systems evaluate text at all.

The headline pattern — fast AI start, mid-experiment AI advantage, late human recovery — showed up in both Yandex and Google, which is the strongest argument in the entire study against the idea that the result is some quirk specific to one search engine’s particular implementation. Two independently engineered ranking systems, built by separate companies with separate priorities and separate histories of algorithm development, converged on broadly the same trajectory over the same eight months, even though they started from different initial positions — Yandex favouring human content heavily out of the gate, Google starting the two groups almost level.

The differences in timing and magnitude between the two engines are also worth noting rather than averaging away. Google’s mid-experiment swing toward AI content was more pronounced than Yandex’s — a 63.6% visibility share for AI-generated pages in Google against 45.5% in Yandex at the same four-month checkpoint — which suggests Google’s ranking systems may have been somewhat slower, during this test window, to recognise the AI-generated content’s weaknesses than Yandex’s were, or alternatively that Google’s algorithm weighted whatever signals favoured the AI-generated pages (structure, keyword matching, formatting cleanliness) more heavily during that middle period. By month eight, Google had not just caught up to Yandex’s pattern but modestly exceeded it in the share of visibility reclaimed by human content, 67.3% against Yandex’s 71.4% — close enough to call the two outcomes comparable, but with Google’s recovery arriving from a deeper mid-experiment hole.

None of this amounts to evidence that Yandex or Google runs an explicit “detect AI text and demote it” classifier as a standalone ranking factor, a claim neither company has confirmed and this study cannot test directly. What it supports is a more defensible and, frankly, more interesting claim: that the accumulated, multi-signal evaluation both search engines perform over time — engagement data, comparison against the wider corpus of similar content, pattern recognition across the kind of secondary or derivative material both companies have spent over a decade building systems to identify — eventually sorts AI-generated and human-written text into different outcome buckets, even when nothing in the initial on-page signal distinguishes them. The mechanism is almost certainly not identical between the two companies, but the end state that mechanism produces, over this particular eight months, was close enough to call convergent. AINET’s visibility and ranking data at three points in the eight-month test

The table compresses eight months of ranking movement into the three points AINET reported publicly, and the shape it shows — a lead, a loss of that lead, and a larger recovery — is the finding this whole study rests on. Average position, shown for the final checkpoint, matters more for real traffic than the visibility-share percentages alone, since click-through rate falls steeply once a page drops past the first few organic results on either search engine.

Dombrovsky’s explanation: recognition, not instant detection

Yaroslav Dombrovsky, who led the study for AINET, framed the central finding in a single sentence to ppc.world: search engines understand where a text is generated and where it is written by an expert, but not immediately. That distinction — recognition that develops over time rather than detection that fires at the point of publication — is the mechanism AINET is proposing to explain the entire shape of its data, and it fits the trajectory better than either a simple “AI content never ranks” story or a simple “search engines can spot AI text instantly” story, neither of which matches what the four-month checkpoint actually showed.

Dombrovsky also drew the connection from this search-ranking finding to the newer discipline of generative engine optimisation, arguing that the two are linked rather than separate problems: language models construct their own answers from sources that search engines already treat as reliable, so a page that loses visibility in traditional search simultaneously loses its chance of being pulled into an AI-generated answer. “If a text loses position in Search, it also loses the chance of appearing in a neural answer,” he told ppc.world, tying the study’s organic-search findings directly to the AI-answer-visibility concerns that have become a parallel preoccupation for SEO practitioners over the past two years.

That linkage is worth taking seriously on its own terms, independent of whether one accepts every detail of AINET’s experimental design, because it describes a compounding mechanism rather than a static one. If AI-generated content tends to decay in organic search rank over a period of months, and if AI answer engines draw disproportionately from pages that already rank well in organic search — a pattern this analysis examines in more depth later, with independent data from a separate research effort — then a page’s slow organic decline does not just cost it search clicks. It progressively removes that page from the pool of candidate sources an AI system might cite when answering a related question, compounding the original visibility loss into a second, AI-specific visibility loss on top of it. Dombrovsky’s framing suggests this is not two separate problems that content teams need to solve independently, but one underlying problem — content that search engines eventually classify as secondary — showing up in two places where businesses measure their online presence.

Minyazhetdinov on the AI content factory temptation

Ruslan Minyazhetdinov, AINET’s founder and chief executive, used the study’s results to address a decision he says he sees businesses making with increasing frequency: as organic search traffic growth slows or declines, the instinct is to compensate with volume, and AI text generation makes volume cheap enough that building what amounts to an AI content factory looks, on paper, like a rational response to a shrinking channel.

His assessment of that instinct, delivered in the same ppc.world piece, is blunt: cheap generated texts produce a visibility spike lasting several months, after which the investment resets to zero, and the right posture toward generative AI in content production is as a tool an expert uses, not as a replacement for the expert. “Our data says the opposite — cheap generated texts give a burst for a few months, after which the investment resets to zero. Neural networks should be used as a tool in the hands of an expert,” Minyazhetdinov told the outlet.

That framing, from the person running the agency that commissioned the study, carries an obvious note of self-interest — a digital agency has a direct commercial reason to argue that expertise and human oversight remain indispensable, since that is the service it sells. The framing is worth stating alongside the data rather than instead of it, which this analysis does throughout rather than treating any single source’s interpretation as the final word. But the underlying economic logic Minyazhetdinov describes — a multi-month visibility spike followed by a reset to zero — matches the shape of the Manus-excepted AI content’s trajectory in AINET’s own measured data closely enough that the business argument and the experimental result reinforce rather than merely echo each other. A business that reads only the four-month data point and commits budget to an AI content factory on that basis is, according to this study’s full eight-month arc, making a decision based on a temporary condition that the same data shows reversing by month eight.

Search visibility as a metric: what it measures, and what it doesn’t

Visibility, as AINET and most SEO tools use the term, is a composite metric built from ranking position across a set of tracked keywords, typically weighted so that a top-one ranking contributes far more to the score than a position-nine ranking, and then expressed as a percentage share relative to the maximum theoretically achievable visibility for that keyword set. In this experiment, with only one keyword in play and ten competing pages, visibility share reduces to something close to a direct readout of relative ranking position between the human and AI-generated groups — which is part of why the single-keyword design, for all its artificiality, produces unusually interpretable numbers.

What visibility share does not directly measure is click-through rate, conversion rate, time on page, or any other behavioural signal a real visitor would generate by actually engaging with the content. AINET’s published figures describe where these ten pages sat in the search results, not how real searchers responded to them once there — largely because the invented keyword has no organic search volume, so there were no real searchers generating that behavioural data to measure in the first place. This is the central tradeoff built into the nonsense-keyword design discussed earlier, and it resurfaces here: the study can speak with unusual confidence about ranking position, because the controlled setup removes confounds that would normally muddy that measurement, but it cannot speak at all to reader engagement, satisfaction or commercial outcome, because the keyword’s artificiality means no real audience ever searched for it.

This matters for how the study’s findings should and shouldn’t be extended. A business reading this analysis and concluding that AI-generated content underperforms specifically because readers find it less engaging, less trustworthy, or less useful than human-written content is reaching a conclusion the data does not support, however plausible it might sound and however much it aligns with other research on reader perception of AI content. What the study supports is narrower and, for SEO purposes, sufficient on its own: ranking algorithms, evaluating these two content types on an identical technical and publishing foundation, over an identical eight months, with identical topical input, sorted them into measurably different ranking outcomes by the end of the test. Why the algorithms did that — reader behaviour signals the invented keyword never actually produced, direct textual analysis, pattern-matching against known AI output characteristics, or some combination the ranking systems’ owners have not disclosed — is a mechanism question the study’s design cannot answer, and this analysis is careful not to claim it does.

A short history of content farms, before AI needed inventing

The problem AINET’s experiment documents — cheap, fast, mass-produced content capturing short-term search visibility before search engines adjust and demote it — is not new to the AI era. It is, in large part, a faster and cheaper repeat of a cycle the search industry already lived through once, roughly fifteen years earlier, under the label “content farm.”

In the late 2000s and into 2010, a business model emerged around producing enormous volumes of low-cost articles, often written by freelancers paid per piece at rates that left little room for research or revision, targeted at search queries identified through keyword research tools as having commercial value and low existing competition. The goal was never to be the best answer to a query; it was to be an adequate, cheaply produced answer that could rank well enough, for long enough, to carry display advertising. The most prominent and most discussed example of this model was Demand Media, whose property eHow became, in the trade press of the period, a shorthand for the entire category, producing on the order of several thousand new articles per day across its network at the model’s peak.

The economics were strikingly similar to the ones Minyazhetdinov describes AI content factories chasing today, even though the production technology could not have been more different: human freelancers rather than language models, but the same underlying bet that volume, produced cheaply enough, would outrun any quality signal the search engine could bring to bear before the investment paid for itself. Google’s own public commentary from the period, including from the engineer then leading its webspam team, Matt Cutts, described the company’s deliberate effort through January and February 2011 to build an algorithmic response specifically targeted at “shallow or low-quality content” produced at scale for search traffic rather than for readers — language that reads, with the obvious substitution of “generated” for “written,” almost identically to Google’s current scaled content abuse policy, discussed in detail later in this piece.

The parallel is not an accident of framing. It reflects a recurring structural feature of commercial search: whenever a production technology makes volume genuinely cheaper than quality, a wave of publishers tests whether search ranking systems can be outrun by scale, and search engines have, roughly every decade, eventually built a countermeasure. AI text generation is simply the latest and by far the cheapest production technology to trigger this cycle, and the AINET experiment is best read as one carefully measured data point inside a repeating historical pattern rather than as the discovery of an entirely new phenomenon. Google’s response to the content farm era arrived under the name Panda, launched on 23 February 2011 and described at the time as affecting roughly 12% of U.S. English-language search queries — an enormous fraction for a single ranking change, and a measure of how large the problem had grown before Google acted on it directly inside its core ranking algorithm rather than through manual enforcement alone.

The eHow case is the part of this history that complicates any tidy “search engines always catch low-quality mass content eventually” narrative, and it is worth telling honestly because it maps onto exactly the kind of mid-experiment ambiguity AINET’s own four-month data point documents. When Panda first rolled out, eHow — by that point the most frequently cited example of a content farm in the industry’s own trade press — did not lose visibility. Several SEO tracking firms reported that eHow’s search visibility actually increased in Panda’s immediate aftermath, prompting a round of commentary in outlets including Search Engine Land and Search Engine Journal questioning whether the update had worked at all, or whether it had simply rewarded one of its intended targets by accident. Demand Media’s own executives, including chief executive Richard Rosenblatt, publicly pushed back on the “content farm” label throughout this period, arguing the characterisation was unfair to a legitimate publishing operation built around matching writers to searcher demand.

It took a second wave of the Panda update, rolled out internationally later in 2011, before eHow’s search visibility declined substantially, and even then the decline unfolded gradually rather than as a single dramatic cliff — Demand Media continued operating the site for years afterward, restructuring its content review process and eventually spinning the property off under new ownership, rather than disappearing from search overnight. The full correction, in other words, took not weeks but the better part of a year, playing out across multiple algorithm updates rather than a single decisive one.

The relevance to AINET’s study is direct. A content farm operator reading only the first Panda rollout’s effect on eHow in early 2011 would have reasonably concluded the update had failed, in exactly the way a publisher reading only AINET’s four-month checkpoint might conclude AI-generated content ranks fine. In both cases, the real verdict took longer to arrive than the initial signal suggested, and in both cases the eventual correction, once it did arrive, was substantial rather than marginal. The pattern recurring across a fifteen-year gap, with human-written low-effort content in one case and AI-generated content in the other, is evidence that the delay itself — not just the eventual demotion — is a structural feature of how large search ranking systems respond to new categories of mass-produced content, rather than a coincidence specific to either episode.

Google’s scaled content abuse policy, read closely

Google’s current framework for exactly this problem is the scaled content abuse policy, introduced on 5 March 2024 alongside a core ranking update and two companion spam policies — expired domain abuse and site reputation abuse. The policy explicitly replaced an older, narrower rule against “spammy automatically-generated content,” and Google’s own announcement of the change was unambiguous about why the replacement was necessary: the earlier policy, written before generative AI became widely accessible, could no longer noticeably distinguish between the production methods a determined publisher might use to generate low-value content at scale.

The policy’s own wording is worth quoting precisely, because it is frequently paraphrased in ways that either overstate or understate its actual scope: scaled content abuse applies to “producing content at scale to boost search ranking, whether automation, humans or a combination are involved.” That phrasing is deliberate and important. Google does not define the violation by production method. A human-written article produced cheaply and in volume specifically to capture search traffic, with no substantial attempt to serve a reader’s actual need, falls under the same policy as an AI-generated article doing the same thing. The determining factor, as Google states it, is purpose and effort rather than authorship — which means the policy’s text does not, on its own, predict or explain AINET’s finding that AI-generated content specifically underperformed human-written content when both were produced under the same brief and published through the same channel.

What the scaled content abuse policy does establish is the enforcement framework into which AINET’s finding fits as one data point rather than a contradiction. If, in practice, AI-generated text at the current state of the technology more often exhibits the characteristics the policy targets — low originality, minimal demonstrated expertise, content indistinguishable from what a dozen other generators would produce from the same prompt — then a policy that is nominally production-method-neutral can still produce outcomes that correlate strongly with production method, simply because the two groups of content differ systematically on the actual criteria the policy evaluates. AINET’s five AI-generated texts, answering an identical prompt, were likely to share far more structural and stylistic DNA with each other than the five human-written texts, produced independently by five different experts, shared with one another — and that kind of convergent, template-like similarity across competing pages is close to a textbook description of what scaled content abuse enforcement is built to catch, regardless of whether the policy’s authors had this specific scenario in mind. Google’s Search Quality Rater Guidelines are the internal manual used to train the thousands of human evaluators — a figure Google has previously put at roughly 16,000 — who assess sample search results and feed quality judgments back into the systems engineers use to tune ranking, without those individual ratings directly changing any specific page’s rank in real time. The guidelines are not the ranking algorithm itself, but they describe, in unusual detail for a company that is otherwise guarded about ranking mechanics, exactly what Google’s own quality standard looks like in practice.

The January 2025 revision to these guidelines restructured the section on low-effort content specifically to address generative AI, replacing older language about “automatically-generated content” with a broader standard built around four attributes: effort, originality, talent or skill, and accuracy — a framework search commentators have since taken to abbreviating as EOTA or EOT/SA. The guidelines state directly that the use of generative AI tools alone does not determine a page’s effort or quality rating, a clarification Google added specifically to head off the reading that any AI involvement automatically triggers a penalty. What does trigger the lowest possible rating, under section 4.6.6 of the guidelines, is content — whatever its origin — that is “copied, paraphrased, embedded, auto or AI generated, or reposted from other sources with little to no effort, little to no originality, and little to no added value for website visitors.”

Google’s own public communication around this update, through spokespeople including search advocate Danny Sullivan and senior search analyst John Mueller, has repeatedly made the same point in slightly different words at public events: it is not whether AI touched the content, it is whether the content shows genuine effort and adds something a reader could not already find elsewhere. A further revision Google made to its “Creating helpful, reliable, people-first content” guidance in October 2026 extended this same framework, formally defining “main content” and listing the same four rater attributes, while also adding an explicit warning against fabricating author identities — invented names, AI-generated headshots, falsely implied credentials — a problem that has grown alongside the growth in AI-assisted publishing, as sites look for ways to simulate the appearance of human authorship and expertise without the substance.

Read against AINET’s data, this framework offers a plausible, if unproven, account of why the eight-month trajectory looked the way it did. A text generated from a single AI prompt, without the kind of expert knowledge, specific examples or original observation a working professional naturally brings to a topic, is close to a textbook case of content with “little to no originality” under Google’s own stated standard — not because it was AI-generated, but because the production method, in practice, often produces exactly the low-originality, low-added-value output the standard describes. The guidelines do not punish the tool. They punish, deliberately and by explicit design, the thing the tool is frequently used to produce when used quickly and without expert oversight.

Telling secondary content from primary content inside a ranking system

Neither Google nor Yandex has published the specific technical mechanism by which their ranking systems might distinguish AI-generated text from human-written text, and AINET’s experiment, whatever its other strengths, was not designed to isolate that mechanism either — it measured the outcome, not the cause. But enough is publicly known about how large-scale ranking systems evaluate content quality in general to sketch out the plausible ingredients, without overstating certainty about which of them actually did the work in this particular test.

Duplication and near-duplication detection is one of the oldest and most mature tools in a search engine’s arsenal, originally built to catch scraped or copied content rather than AI output, but directly applicable to it: five AI systems answering one shared prompt are likely to converge on similar structures, similar phrase patterns, and similar factual framing, in a way that five independently researched human experts, drawing on their own distinct professional experience, are not. A ranking system comparing a page against the broader corpus of similar content — which is exactly the kind of comparison Google’s die-to-die language around duplicate and near-duplicate content describes — would plausibly flag convergent AI output as less original even without running any dedicated “AI detector” at all.

Engagement signal aggregation over time is the second plausible mechanism, and it is the one that best explains the multi-month delay AINET’s data shows rather than an instant judgment at publication. Search engines do not rank purely on the words on a page; they incorporate, to a degree neither company discloses precisely, how real users interact with a page once they reach it — whether they stay, whether they return to the results page quickly to try something else, whether the page satisfies the query well enough that the search session ends there. These signals accumulate gradually, require real traffic to generate, and cannot much distinguish content quality within the first days or weeks after publication, when traffic volume is low and statistically noisy. A quality gap that only becomes measurable once enough aggregate user behaviour has accumulated would produce exactly the kind of delayed divergence AINET observed — present in the data by month eight, not yet visible in the data at month one.

Direct linguistic pattern analysis, the mechanism most people assume when they hear “AI detection,” almost certainly plays some role as well, particularly given how aggressively AI text detection as a technology category has matured since 2023, but it is the least well-evidenced of the three mechanisms as an explanation for this specific study’s results, and the one this analysis is most cautious about asserting. What can be said with more confidence is that these mechanisms are not mutually exclusive, that search engines have strong incentives to combine multiple weak signals into a more reliable composite judgment rather than relying on any single detector, and that a multi-month delay before a quality judgment solidifies is consistent with a system that waits for enough corroborating evidence across several of these channels before it genuinely adjusts a page’s position — exactly the pattern both the AINET experiment and the historical Panda rollout against eHow independently show.

The web-scale picture: AI writing has reached parity in volume, not in rank

AINET’s experiment is a controlled test on ten domains built around one invented keyword in one market. The strongest reason to take its conclusions seriously beyond that narrow scope is that a separate, much larger and methodologically very different research effort, run by the SEO research firm Graphite, has independently reached a strikingly similar conclusion by studying the live, uncontrolled web at massive scale.

Graphite’s research tracks what share of newly published English-language articles across the open web are classified as primarily AI-generated, using an averaged score across three independent AI-detection systems — Pangram, Copyleaks and GPTZero — applied to large samples drawn from Common Crawl, a publicly available archive of web page snapshots used widely in web-scale research. Graphite’s most recent published figures, covering data through the first quarter of 2026, found that primarily AI-generated articles made up 35.9% of sampled content within twelve months of ChatGPT’s November 2022 public launch, rose further to 49.6% by the first quarter of 2025, briefly crossed the halfway mark at 50.9% in the fourth quarter of 2025, and then settled back to 49.9% in the first quarter of 2026 — a plateau that Graphite’s own researchers describe as having held roughly steady, fluctuating narrowly around the 50% line, for roughly a year and a half by the time of the most recent measurement.

That plateau is itself a real finding, and Graphite’s own interpretation of it lines up directly with AINET’s experimental result. Graphite’s researchers explicitly flag that the growth in AI-generated publishing appears to have stalled once it reached roughly half of newly published content, and they point to a companion study — examined in the next section — on how that content actually performs in search as a plausible explanation: publishers and content operations may be learning, empirically and the hard way, that heavily AI-generated content underperforms in the channels that matter for traffic, and are adjusting production volume accordingly, even without having access to a controlled study like AINET’s that spells the mechanism out explicitly.

The scale difference between the two research efforts is worth stating plainly, because it is also their complementary strength. AINET’s ten domains and single keyword trade real-world scale for experimental cleanliness, producing a result that is unusually trustworthy about causation within its narrow scope but says nothing directly about the wider web. Graphite’s tens of thousands of sampled live web pages trade experimental control for scale, producing a result that describes the actual current state of the web with far more breadth but cannot claim the same causal clarity a controlled experiment provides, since none of the variables other than AI-detection score were held constant across the sampled pages. Two studies with opposite strengths and opposite weaknesses, run independently by unrelated organisations in different countries using different methods, arriving at the same underlying conclusion — AI-generated content is common in production but underrepresented in search success — is a considerably stronger evidentiary position than either study offers alone.

The sources ChatGPT and Perplexity actually cite when they answer

Graphite’s companion research, published separately from its production-volume tracking under the title “How Does AI-Generated Content Perform in Search and Answer Engines,” addresses exactly the question Dombrovsky raised about the link between organic search ranking and AI-answer visibility, using real query data across both Google Search and two major AI answer engines.

The findings are consistent across every surface Graphite measured. In Google Search, the research found that AI-generated content makes up only a small share of pages ranking in the organic results — roughly 14% — while 86% of ranking pages are primarily human-written, with the AI-generated share concentrated disproportionately toward lower ranking positions and the share of AI content at position one measured at a markedly lower 7%. The same study examined which articles ChatGPT’s web-search feature and Perplexity actually cite when answering user questions, and found an almost identical split on both platforms: 82% of articles cited by ChatGPT with web search are human-written, against 18% AI-generated; Perplexity’s citations split 82% human against 18% AI-generated as well.

The consistency of that roughly 80-to-20 human-to-AI split across three functionally different systems — one traditional search engine and two generative answer engines built on entirely different retrieval-and-synthesis architectures — is the single most important corroborating data point this analysis can offer for AINET’s narrower finding. It would be a reasonable methodological objection to AINET’s study to suggest its result might be an artefact of the Russian-language market, the beauty-content vertical, or some quirk of Yandex’s specific algorithm. Graphite’s data, covering English-language content and American-dominant AI platforms, removes every one of those specific objections while reproducing the same directional finding: across language, market, vertical and underlying technology, human-written content substantially outperforms AI-generated content at the point where a search or answer system decides what to actually surface to a user.

Graphite’s data also lines up with the compounding mechanism Dombrovsky described. The study found that a large share of pages an AI answer engine retrieves while researching a response never make it into the final cited answer at all — AI-detectable content was measurably less likely to survive that retrieval-to-citation filtering step than human-written content covering the same topic. An article that fails to rank well in organic search is, by this evidence, simultaneously less likely to be pulled into the candidate pool an AI system considers citing in the first place, and less likely to survive the final selection even when it is pulled in — a double penalty that matches exactly the dynamic Dombrovsky described from the organic-search side of AINET’s own experiment.

GEO and why a ranking drop becomes an AI-answer drop

The academic grounding for the link Dombrovsky describes predates both the AINET and Graphite studies by roughly two years. In November 2023, a team of researchers from Princeton University, Georgia Tech, the Allen Institute for AI and IIT Delhi — Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande — published a paper titled “GEO: Generative Engine Optimization,” later accepted at the 2024 ACM SIGKDD conference, which coined the term that has since become the standard shorthand across the SEO industry for optimising content specifically to be cited inside AI-generated answers rather than merely to rank in a traditional list of links.

The paper’s central empirical contribution was a controlled benchmark of roughly 10,000 queries, called GEO-BENCH, used to test nine distinct content-modification strategies against a generative-engine pipeline modelled on retrieval-augmented systems like Bing Chat. The strategies that moved the needle most were adding cited sources, adding direct quotations, and adding specific statistics — concrete, verifiable, attributable content, in other words, rather than generic or vague assertion — which together lifted a page’s visibility inside generated answers by as much as 30 to 40% in the study’s measurements, with some lower-ranked source pages gaining substantially more.

The throughline connecting the Princeton research to both AINET’s and Graphite’s findings is evidentiary specificity. A generative engine, as the Princeton researchers describe the category, retrieves candidate documents and then synthesises a single prose answer from them, in effect fact-checking and weighing competing sources as it writes rather than simply listing links in ranked order. That process rewards content carrying the markers of genuine, specific, checkable expertise — named sources, real numbers, direct quotation — and by the same logic, it is poorly suited to rewarding the kind of generic, templated output a single AI prompt tends to produce when nothing forces it to cite a specific source or supply a number it cannot verify. A model generating an article about an invented beauty-industry term, with no real external reference material available to draw from because the topic does not actually exist, is structurally incapable of producing the kind of attributed, evidence-dense content the GEO research found generative engines reward most — which may be a contributing reason, beyond organic search ranking signals alone, that the AI-generated texts in AINET’s experiment struggled to hold their position once the initial novelty of a fresh page wore off.

This is also the mechanism underlying Dombrovsky’s warning about AI answers specifically. If an AI answer engine’s synthesis process favours attributed, specific, verifiable content, and if that kind of content correlates strongly with genuine expert authorship rather than single-pass AI generation, then the same quality gap documented in organic search rankings by both AINET and Graphite should reproduce itself, independently, inside AI-generated answers — which is precisely what Graphite’s ChatGPT and Perplexity citation data, examined in the previous section, actually shows. Properties search and answer engines evaluate, and how the two content types compare across the evidence cited in this piece

This table pulls together AINET’s controlled eight-month result and Graphite’s independent web-scale measurement into one comparison. The two data sources were produced separately, in different countries and different markets, yet converge on the same direction and a similar order of magnitude, which is the strongest reason to treat the underlying pattern as more than a coincidence of one agency’s single experiment.

Business impact by sector: beauty, e-commerce, finance, legal, local services

AINET’s test ran specifically in the beauty and skincare content vertical, chosen presumably because the agency had ready access to professional expert writers in that field and because beauty content is a well-established, commercially genuine category in both Russian search markets. Extending the finding to other sectors requires judgment about which industries share the structural features that made the human-versus-AI gap show up here, and which might behave differently.

Beauty, wellness and personal care content sits close to what Google’s quality framework treats as a sensitive category requiring higher evidentiary standards, given the direct connection between the advice given and a reader’s health, appearance and spending decisions. Sectors with a similar profile — nutrition, fitness, skincare, cosmetic procedures — are plausible candidates for a comparable or even larger gap between expert-written and AI-generated content, since these are exactly the categories where Google’s rater guidelines apply a stricter accuracy bar for what the company calls Your Money or Your Life topics, and where genuine first-hand experience is both harder for an AI system to fabricate convincingly and more useful to a reader when genuinely present.

Financial services, legal information and medical content sit even further along that same spectrum, carrying the heaviest accuracy and trust requirements of any content category under Google’s published standards, and these are sectors where the professional and regulatory cost of an AI-generated factual error is also highest — a wrongly stated tax threshold, an outdated legal deadline or a mischaracterised drug interaction carries consequences well beyond a lost search ranking. Businesses in these verticals have the strongest case for treating this study’s findings as directly applicable, and arguably understating the real risk, since AINET’s beauty-content test did not even attempt to measure factual accuracy or liability exposure, only ranking position.

E-commerce product and category content, by contrast, may behave somewhat differently, because a notable share of commercial product content — specifications, size charts, shipping policy summaries — carries less of the kind of distinctive expert voice that separated the human and AI writers in AINET’s test, and more of the kind of structured, factual, low-variance information that AI systems may reproduce reliably from accurate source data without the same originality gap showing up. Local service businesses — plumbers, dentists, contractors — occupy a related middle ground: content describing a specific local business’s own services, written with real operational detail a generic AI prompt cannot invent, should behave more like AINET’s human-written group, while generic “how to choose a plumber” advice content generated at scale across thousands of local business websites, with only the business name swapped between versions, is close to a textbook description of the kind of templated output both Google’s scaled content abuse policy and this study’s findings warn against. The underlying variable in every sector appears to be the same one the study isolated directly: not the industry label, but whether the content carries specific, non-generic, demonstrable expertise that a shared AI prompt structurally cannot supply.

The economics of an AI content factory, worked through

Minyazhetdinov’s warning about businesses building “AI content factories” in response to declining organic traffic deserves a closer look at the actual arithmetic behind that decision, because the appeal of the approach is not irrational — it follows directly from how much cheaper AI-generated text is to produce than professionally written, expert-reviewed content, and from how visible that cost difference is compared to the far less visible, multi-month-delayed cost of the eventual ranking decline AINET’s study documents.

A professional expert writer, paid at rates that reflect genuine subject knowledge, research time and the kind of first-hand detail that separates expert content from generic advice, represents a real and immediate cost per article — one that scales linearly, or close to it, with the number of articles a business commissions. Generating an equivalent volume of articles through an AI system costs a small fraction of that per-article price, and the marginal cost of producing additional volume falls close to zero past the initial setup. For a business measuring return on content investment over a short time horizon — a quarter, or even the four months at which AINET’s own data showed AI content briefly leading — the AI approach looks straightforwardly superior: similar or better visibility, at a fraction of the cost.

The problem, according to both AINET’s measured data and Minyazhetdinov’s framing of it, is that this short-horizon calculation misses the asset’s actual depreciation curve. A piece of content is not a one-time expense that either works or doesn’t; it is closer to a long-lived asset whose value compounds through sustained search visibility, provided the asset holds its ranking position over time. AINET’s eight-month data shows the AI-generated content’s practical value — measured as its contribution to search visibility — declining sharply in the second half of the test period, while the human-written content’s value, after an unexpected dip, recovered and compounded. A business that commissioned AI content specifically because of its low upfront cost, and then measured success only at the three- or four-month mark, would conclude the investment paid off handsomely relative to its cost. The same business measuring at month eight, on AINET’s data, would find the visibility it had been counting on had substantially eroded, while the cheaper production cost had not changed — meaning the true cost-per-unit-of-sustained-visibility for the AI-generated approach was considerably higher than it appeared at the point most businesses would naturally choose to evaluate it.

This is the specific economic trap Minyazhetdinov’s comment about investment “resetting to zero” describes, and it is a trap built directly into how most businesses measure content ROI: by looking at recent performance rather than performance sustained over a timeframe long enough to capture the delayed ranking correction this study — and the historical Panda precedent before it — both suggest is coming.

The point past which the experiment’s conclusions stop generalising

A study this useful deserves an equally honest account of where its evidence runs out, and several limits are worth naming explicitly rather than leaving implicit.

The sample size is small. Five human-written pages and five AI-generated pages is enough to detect a large, consistent effect of the kind AINET found, but it is not enough to rule out the possibility that one or two unusually strong or unusually weak individual pieces of content — on either side of the comparison — are doing disproportionate work in the aggregate numbers. The Manus exception, examined earlier, demonstrates exactly this kind of individual variation within the AI-generated group; it is entirely plausible that comparable variation exists within the five human-written pages as well, with one or two unusually strong expert articles carrying more of the human group’s aggregate visibility than the published figures, which report only group-level averages, can reveal.

The topic is narrow and the market is specific. One beauty-industry prompt, executed in the Russian language, tested against Yandex and Google specifically, is a single data point about one content category in one linguistic and regulatory environment, not a universal statement about every content vertical in every search market. The sector analysis earlier in this piece extends the finding by informed judgment rather than by additional direct evidence, and that extension should be read with appropriately reduced confidence the further a given industry’s characteristics diverge from beauty content’s reliance on demonstrable personal and professional expertise.

The keyword had no real search demand, a design choice this analysis has already flagged as a deliberate and justified tradeoff for experimental cleanliness, but one that means the study cannot speak to how real user engagement — genuine click-through behaviour, genuine satisfaction signals, genuine return visits — might have influenced the ranking trajectory differently than it did for a keyword with actual searcher traffic generating that behavioural data from day one.

Eight months is long relative to most SEO experiments but still short relative to the full lifecycle of a piece of content, and AINET’s own data shows the ranking order shifting substantially between every checkpoint it measured. A study that stopped at any one of the three checkpoints reported here would have reached a materially different conclusion than the full eight-month arc supports; there is no principled reason to assume the trajectory had fully stabilised by month eight rather than continuing to shift further in either direction with more time, and AINET’s public materials do not indicate whether the test has continued beyond the reported eight months or ended there. Readers citing this study — including, this analysis acknowledges, its own citation of it — should treat “eight months” as the longest verified window currently available rather than as evidence of where the trend necessarily settles. It is worth being precise about a distinction that gets collapsed constantly in how studies like this one get summarised and shared: AINET’s experiment is a study of fully AI-generated content produced from a single prompt with no subsequent human editing, review or enrichment, not a study of AI as a writing tool used within an expert-led production process. Those are different objects, and the data from this test cannot be extended from one to the other without additional assumptions the experiment itself does not test.

Every public statement from AINET about its own findings makes exactly this distinction, and makes it emphatically. Minyazhetdinov’s framing — AI as a tool in the hands of an expert, not a replacement for the expert — describes a production model the experiment never actually measured: human-led content creation with AI assistance for drafting structure, summarising research, generating first-pass copy for human revision, or handling mechanical tasks like formatting and metadata generation. None of the five AI-generated texts in this study went through that kind of hybrid process; they were generated and published as AI output, compared directly against separately produced, fully human-written text. The study is evidence against unsupervised, single-prompt AI content replacing expert writers. It is not evidence against AI-assisted workflows in which expert judgment remains the final and controlling input.

This distinction matters because the broader research field, including the Graphite data discussed earlier in this piece, explicitly flags AI-assisted, human-edited content as a category its measurement methods cannot cleanly separate from either pure human writing or pure AI generation, and multiple researchers in this space — Graphite’s own team among them — have noted they expect that hybrid category to be more prevalent than their detection tools currently reveal, and plausibly to perform differently from the unsupervised AI content their studies can actually measure. The honest reading of the evidence assembled in this piece is narrower and more useful than “AI content is bad”: it is that content produced without substantial human expertise, originality or editorial investment underperforms over time regardless of whether a person or a model did the typing, and that current single-prompt AI generation, used without that human layer, tends to produce exactly the kind of low-originality output this describes — which is a conclusion about production process and editorial standards, not a verdict on the technology used to execute any individual step within that process.

A practical workflow for using AI without losing visibility

The evidence assembled across AINET’s experiment, Graphite’s web-scale measurement, the Princeton GEO research and Google’s own published quality standards points toward a fairly specific and actionable production model, rather than a blanket instruction to avoid generative AI in content work altogether.

Use AI for structure, drafting speed and research synthesis, not as the final source of expertise. The Princeton GEO research’s strongest finding — that specific, attributable statistics and named sources drive the largest measurable gains in AI-answer visibility — describes exactly the kind of input an AI model cannot supply on its own for a genuinely novel or specialised topic, because it has no first-hand data to draw from beyond its training corpus. A human expert, asked to review and enrich an AI-generated draft with real examples, specific figures from their own professional experience, and direct observations a generic prompt could not produce, is adding precisely the ingredient both the GEO research and AINET’s result suggest search and answer systems reward.

Treat the first weeks or months of strong performance from AI-assisted content with appropriate skepticism rather than as confirmation of a working strategy. AINET’s own data shows AI-generated content performing competitively, and briefly better than human-written content, for roughly a third of the eight-month test window. A business or agency that measures success only within that window, as the economic analysis earlier in this piece worked through in detail, risks mistaking a temporary condition for a stable one. The appropriate response is not to avoid tracking early performance, but to build review checkpoints at three, six and twelve months specifically to catch the kind of delayed reversal this study documents, rather than declaring victory and scaling up production based on month-one or month-three data alone.

Apply genuine editorial review before publication, not as a formality but as a substantive check against the specific failure mode Google’s quality standards target. The rater guideline language examined earlier — effort, originality, talent or skill, accuracy — functions as a reasonably concrete checklist: does this piece contain something a competing page, generated from the same or a similar prompt, would not also contain; does it reflect genuine first-hand knowledge or observation; is every factual claim checked against a verifiable source rather than accepted on the model’s own authority. Content that can pass that checklist, regardless of how much AI assistance went into producing the first draft, is evidence-based content built to the standard this study’s findings reward. Content that cannot pass it is the category this study’s findings, and Google’s own stated policy, both independently warn against.

The verification habit this study is really an argument for

Step back from the specific numbers, and AINET’s experiment is in the end an argument for a discipline that long predates generative AI and will outlast the current wave of anxiety about it: measuring outcomes over a time horizon long enough to catch delayed effects, rather than trusting whatever signal happens to be available at the moment a decision gets made. That discipline failed the content farm operators of 2010, who built businesses on the assumption that Google’s early tolerance of mass-produced content was permanent rather than a lag waiting to close. It is failing, according to this study’s own mid-experiment data, any business currently measuring its AI content strategy only against the three-to-four-month window where AINET’s own numbers show AI-generated content briefly outperforming expert-written alternatives.

The specific failure mode this study documents is not unique to AI content, and treating it as an AI-specific problem risks missing the more general lesson. Any production shortcut that trades genuine quality for speed or cost, in any content category, in any era of search technology, faces the same structural risk: search and answer engines that evaluate content on accumulated, multi-signal evidence will eventually catch up to a quality gap that initial, thin signal cannot yet detect, and the businesses most exposed to that eventual correction are precisely the ones that scaled the shortcut fastest while the correction was still pending. Generative AI makes the shortcut dramatically cheaper and therefore dramatically easier to scale than the freelance-writer content farms of 2010 ever were, which raises the stakes of the same underlying mistake rather than introducing a fundamentally new one.

For a business or publisher weighing how to interpret this study, the actionable version of the lesson is straightforward: treat any new content-production efficiency gain with the assumption that search and AI-answer systems will eventually evaluate its output on the same substantive standards — originality, demonstrated expertise, genuine value to the reader — that they have always nominally applied, even when early results suggest otherwise. The gap between “early results suggest otherwise” and “the standard actually gets enforced” is where AINET’s four-month data point sits, and it is exactly the window in which a business following only recent performance data would make the decision this study’s full eight months argues against.

The reliability of the tools used to detect AI writing in the first place

A natural question raised by both AINET’s experiment and Graphite’s web-scale research is how confident anyone should be in the underlying detection methods used to classify text as AI-generated versus human-written, since the entire evidentiary chain in the web-scale studies depends on automated classifiers rather than on known ground truth about who actually wrote each sampled article.

Graphite’s own published methodology addresses this concern directly by reporting both false-positive and false-negative rates for the three detectors it used — Pangram, Copyleaks and GPTZero — tested against a dataset of articles with verified human authorship and a separate dataset of articles deliberately generated by recent frontier models including GPT-5, Gemini 3.1 Pro and Claude Opus 4.6. The reported false-negative rates — AI-generated content misclassified as human-written — varied substantially by detector and by the generating model, ranging from close to zero in some configurations to several percentage points in others, with Copyleaks showing a noticeably higher false-negative rate against certain model outputs than GPTZero did in the same test. This variance is precisely why Graphite averages across three independent detectors rather than relying on any single one, a methodological safeguard that reduces, though does not eliminate, the risk that any one detector’s particular blind spot distorts the headline percentages.

AINET’s experiment sidesteps this detection-reliability question almost entirely, and this is an underappreciated strength of its design worth returning to. Because AINET controlled which five pages were AI-generated and which five were human-written from the outset, rather than inferring authorship after the fact from a classifier run against unknown text, its study carries no detection-accuracy risk at all. The ranking outcomes it measured are ground truth about real search engine behaviour toward content of known, verified origin — the single clearest methodological advantage AINET’s small, controlled test has over Graphite’s much larger but classifier-dependent web-scale measurement, and the reason the two studies’ convergence on a similar finding, despite opposite risk profiles on this specific point, strengthens confidence in the shared conclusion rather than simply repeating the same underlying measurement error twice.

The broader implication for any business or researcher trying to measure AI content prevalence or performance going forward is that detector-based studies of the live, uncontrolled web should be read with an appropriate allowance for classification noise, particularly as AI models themselves continue to improve at producing output that is harder to distinguish from human writing through linguistic pattern analysis alone — a trend several of the sources examined in this piece note explicitly as an ongoing and, from the detection industry’s perspective, worsening problem.

The Russian search market as a testbed, and why the pattern likely travels

Running this experiment specifically in Russia, using Yandex alongside Google, offers an unusual methodological bonus that is easy to overlook: it provides a second, independently engineered search ranking system operating in a market where AI-generated Russian-language content may carry different linguistic signatures than AI-generated English-language content, given that the major generative AI systems — including four of the five tested here — were trained predominantly on English-language data before being extended to other languages.

Yandex has built its ranking systems over more than two decades specifically around the Russian-language web, including its own machine learning architecture, known as MatrixNet and its successors, developed independently of Google’s ranking infrastructure. A search engine built from the ground up around one language’s particular structure, idiom and web ecosystem is plausibly better positioned to detect subtle linguistic artefacts of non-native or machine-translated-feeling text than a search engine for which that language is one of over a hundred it ranks content in. This may help explain why Yandex’s initial rejection of the AI-generated content in AINET’s test was so much sharper than Google’s — 72.7% to 27.3% against Google’s near-even 49.1% to 50.9% — even though both engines converged on a broadly similar final outcome by month eight.

If this interpretation holds, it has a useful implication for how readers outside Russia should weight this study’s relevance to their own markets. A pattern that shows up clearly in a market where the dominant search engine is specifically tuned to the content’s native language, and that also eventually shows up, with a longer delay, in a global search engine operating across many languages simultaneously, is more likely to represent a genuine property of how large-scale ranking systems evaluate content quality than an artefact specific to one linguistic environment. English-language markets, where the major AI systems were most heavily trained and where detection tooling and quality signals are most mature, given that Google’s own quality rater guideline updates and scaled content abuse enforcement have focused predominantly on the English-language web to date, arguably have less excuse than the Russian market to expect generative AI content to escape the same eventual correction AINET measured — if anything, the infrastructure built to catch it is more developed, not less, in the markets where most readers of this analysis are likely publishing. One detail buried in Google’s October 2026 update to its helpful-content guidance, noted briefly earlier in this piece, deserves its own treatment because it intersects directly with the kind of content AINET’s experiment examined: the explicit warning against fabricating author identities, including invented names, falsely implied professional credentials and AI-generated headshots used to simulate the appearance of a real human expert standing behind a piece of content.

This is worth flagging because it describes the obvious next move for any publisher who reads AINET’s conclusion — that search engines eventually favour demonstrably expert, human-authored content — and responds not by hiring real experts, but by dressing AI-generated content in the visual and textual signifiers of expert authorship without the substance behind it: a fabricated author bio claiming credentials nobody holds, a generated headshot attached to a byline that corresponds to no real person, an “About the Author” section describing years of professional experience that were never lived. Google’s explicit guidance against exactly this practice signals that the company anticipated this response and built a countermeasure into its quality standard ahead of, or at least alongside, its rollout.

AINET’s experiment did not test this specific evasion, because its five AI-generated texts were tested as AI-generated texts, without fabricated human bylines attached. Whether a convincingly fabricated expert identity could noticeably improve an otherwise-unchanged AI-generated article’s ranking trajectory is an open empirical question this study cannot answer, and one that would be genuinely useful for a future controlled experiment to test directly, given how directly it bears on the practical question many content operations are likely asking in response to studies like this one. What can be said with more confidence is that Google has pre-emptively classified the practice as a guideline violation, which means a publisher pursuing this route is not finding an unaddressed loophole in the standard, but knowingly violating an explicit, recently published policy — a materially different risk position than simply publishing unlabelled AI-generated content that happens to underperform on its own merits, which is the behaviour AINET’s study actually measured. The honest answer for any business genuinely trying to close the gap this study documents is the one Minyazhetdinov already gave: put a real expert’s judgment into the content, rather than simulating that expert’s presence after the fact.

The risk of overcorrecting against AI entirely

A fair reading of this study has to guard against an overcorrection that is just as poorly supported by the evidence as the “AI content works fine” conclusion this piece has already spent several sections dismantling: the idea that any use of generative AI anywhere in a content pipeline is now a liability to be eliminated outright.

Nothing in AINET’s data, Graphite’s data, the Princeton GEO research or Google’s own published policy supports that stronger claim, and several of those same sources explicitly contradict it. Google’s rater guidelines state directly that the use of generative AI tools alone does not determine a page’s quality rating — language the company added specifically because it anticipated, correctly, that publishers would otherwise read AI involvement itself as disqualifying. The Manus result inside AINET’s own experiment demonstrates the same point from the opposite direction: an AI system produced content that outperformed most of the human-written pages in the same test, which is impossible to reconcile with a blanket “AI content cannot rank” rule, however convenient that rule would be to state.

The actual, better-supported position sits between the two extremes this debate tends to collapse into. AI generation used as the entire production process, with no expert review, no added specificity, no genuine editorial investment, tends to produce content that underperforms over a multi-month horizon — that is what AINET measured directly. AI generation used as one tool within a process still anchored by genuine human expertise, fact-checking and original observation is a different production model entirely, one this study did not test and one several of the sources examined here — Graphite’s researchers among them — explicitly flag as the harder-to-measure, likely more prevalent, and plausibly better-performing middle category that current research methods struggle to isolate cleanly.

For a business trying to set policy based on this evidence, the practical distinction is less about banning a tool and more about defending a standard: does a given piece of content, whatever combination of human and AI effort produced it, contain genuine expertise, specific verifiable detail and real editorial investment that a competitor running the same prompt through the same AI system could not trivially reproduce. Content that clears that bar has, on the evidence assembled in this piece, a reasonable claim to durable search and AI-answer visibility regardless of which tools touched it along the way. Content that does not clear it is exposed to exactly the delayed correction AINET’s eight months measured, regardless of whether a human or a model typed the words that failed to clear it.

Metrics publishers should track beyond search rankings alone

AINET’s study measured one outcome variable with unusual precision — search visibility and ranking position — because that was the variable its controlled design was built to isolate. A publisher applying this study’s lessons to real content operations needs a wider dashboard than that, because ranking position is a leading indicator of business outcomes rather than the outcome itself, and because some of the risks this study implies show up in metrics ranking data alone will not capture.

Time-to-decline, not just current ranking, deserves its own tracked metric. Because AINET’s data shows the gap between AI-generated and human-written content opening up gradually over months rather than appearing instantly, a publisher relying only on a snapshot of current rankings risks missing the early signs of the same decline this study documents until it has already cost several months of accumulated traffic. Tracking ranking trajectory — the rate of change over rolling three-month windows, specifically for content identified as primarily AI-generated — would catch the kind of mid-experiment reversal this study found well before it fully plays out, giving a publisher time to intervene with editorial revision rather than discovering the problem only once visibility has already cratered.

Citation presence inside AI answer engines is a second metric most content operations still do not track systematically, despite the direct link Dombrovsky drew between organic ranking and AI-answer visibility, and despite Graphite’s independent data confirming that link empirically. Tools that monitor whether a given page or domain is being cited by ChatGPT, Perplexity, Google’s AI Overviews and comparable systems, for the queries that page is meant to serve, give a publisher visibility into the second half of the compounding penalty this piece has described — the AI-answer visibility loss that follows an organic ranking decline — rather than discovering it only once the organic traffic drop has already registered in analytics.

Content audit flags tied to production method, maintained as an internal record of which pieces were generated primarily by AI, which were human-written, and which followed a genuine hybrid process with real human revision, give a publisher the ability to retrospectively test exactly the kind of comparison AINET ran deliberately, using their own real content portfolio rather than a synthetic experiment. A publishing operation that cannot answer, for any given underperforming page, what production process generated it is not well positioned to diagnose whether this study’s findings apply to its own situation or to correct for them if they do. AINET is itself an SEO and GEO agency, and this study doubles as a piece of evidence-based marketing for the service it sells — expert-informed content strategy rather than volume-based AI content production. That commercial context does not invalidate the study’s methodology, which this analysis has examined in detail and found genuinely rigorous by the standards of an industry that rarely produces controlled experiments at all, but it is worth naming because the study’s publication and the agency’s business incentives point in the same direction, and readers should weigh the evidence on its own methodological merits rather than on the authority of who published it.

The broader SEO and digital marketing industry has spent roughly three years working through exactly the tension this study documents: the same generative AI tools that threaten to commoditise content production also threaten the specific service model many agencies have historically sold, which is human-produced, search-optimised writing delivered at a day rate or project fee. An agency that can demonstrate, through evidence like AINET’s experiment, that the specific thing it sells — expert judgment, genuine research, editorial oversight — produces durably better search and AI-answer outcomes than the cheaper AI-only alternative has a strong commercial reason to invest in exactly this kind of research, independent of whether the underlying finding happens to be true.

The useful test for any reader evaluating this dynamic is whether the study’s methodology would hold up if an agency with the opposite commercial incentive had run it. AINET’s controlled design — synthetic keyword, matched technical infrastructure, multiple AI systems rather than one, eight-month duration rather than a quick snapshot — is the kind of rigour that should produce broadly the same result regardless of who conducts the test, provided the test is run honestly, which is a much different bar than asking whether the agency benefits from the finding. This analysis has applied that test throughout, cross-referencing AINET’s numbers against Graphite’s independently conducted, commercially unrelated web-scale research rather than accepting AINET’s figures on the strength of the agency’s own framing alone — and the two studies’ convergence, documented at length in earlier sections, is the strongest available answer to the fair question of whether this result would survive replication by a party with no stake in the outcome.

Brand trust erodes faster than a single ranking metric shows

A dimension of this study’s findings that neither AINET’s published data nor most of the press coverage around it addresses directly is brand trust — the slower, harder-to-quantify erosion that happens when a business’s published content is perceived by its actual audience as generic, impersonal or unreliable, independent of and in addition to whatever search ranking penalty that content eventually incurs.

Search ranking and reader trust are related but distinct phenomena, and the relationship between them runs in both directions. A page that ranks poorly reaches fewer readers, so it has less opportunity to damage trust simply by being read less often. But a page that does rank, and that a reader does encounter, carries its own separate risk: generic, templated content that reads as interchangeable with a dozen competing pages signals to a human reader, whether or not that reader consciously registers the signal, that the publisher did not invest real effort or expertise in answering their question. That perception compounds across repeated encounters in ways a single ranking snapshot cannot capture — a reader who encounters one unsatisfying, generic page from a brand is less likely to trust the next page from that same brand, creating a reputational drag that outlasts any individual piece of content’s own ranking trajectory.

This matters directly for how a business should weigh AINET’s findings, because the study’s own metric — search visibility — almost certainly understates the full cost of the AI-content strategy it measured. A business that scales AI-only content production, even setting aside the eventual ranking decline this study documents, is simultaneously accumulating reader-trust damage across every page a real visitor does encounter during the months before that ranking decline fully plays out — damage that persists even after the specific underperforming pages have been removed, rewritten or deprioritised, because the reader’s impression of the brand’s overall reliability was formed by direct experience rather than by the search algorithm’s eventual verdict. Google’s own rater guideline framework gestures at this same dynamic from the search engine’s side, treating accuracy and genuine value to the reader as core quality signals precisely because search engines are, in the end, aiming to proxy for the same reader satisfaction a business should care about directly, not merely as an input into someone else’s ranking formula. AINET’s finding does not arrive in a vacuum of prior SEO research on content quality and ranking, even setting aside the AI-specific studies this piece has already examined in depth. Independent correlational research from established SEO tooling companies, built from large-scale crawls of real search results rather than controlled experiments, has for years found a consistent association between content depth, originality and demonstrated expertise on one side, and ranking position on the other — the kind of evidence that long predates generative AI and that this study’s findings slot neatly alongside rather than contradict.

The limitation of that older correlational body of research has always been the same one that limits Graphite’s web-scale AI studies: correlation between content characteristics and ranking position does not, on its own, establish that the content characteristic caused the ranking outcome, since highly ranked pages also tend to sit on more established domains, with more accumulated backlinks and more mature publishing histories that could independently explain both their depth and their ranking. AINET’s controlled design is useful specifically because it closes that gap for the AI-versus-human question in particular, holding domain age, backlink profile and publishing history constant across all ten test pages and isolating the content variable directly — a design advantage genuinely rare in SEO research, and the reason this analysis has returned repeatedly to the control as the study’s central methodological strength rather than treating it as a minor technical detail.

Read together, the newer controlled evidence and the older correlational evidence tell a consistent story rather than two separate ones. Years of large-scale correlational SEO research established that content depth and demonstrated expertise correlate with ranking success, without being able to fully rule out confounding factors. AINET’s controlled experiment, and Graphite’s independent web-scale measurement alongside it, now supply evidence that the relationship holds even when those confounding factors are deliberately removed, and that it holds specifically for the comparison between AI-generated and human-written text that has become the industry’s most urgent open question since generative AI tools became widely accessible to any publisher who wanted to use them.

The specific window search engines seem to need before they act

A detail worth isolating from the rest of AINET’s data, because it has direct operational value for anyone trying to time content review decisions, is the rough duration of the delay between publication and the point at which a ranking system’s judgment appears to stabilise: somewhere between the four-month checkpoint, where AI-generated content was still performing competitively or better in both search engines, and the eight-month checkpoint, where the reversal had fully played out.

This window is consistent, at least directionally, with the historical Panda precedent examined earlier in this piece, where the first rollout in February 2011 initially failed to genuinely demote eHow, and a second, more decisive correction arrived later the same year — a gap of several months between the first algorithmic attempt at a correction and the point at which it visibly worked as intended. It is also broadly consistent with how search engines have historically described their own update cadence for addressing newly identified categories of problematic content: an initial detection and policy response, followed by iterative refinement over subsequent months as the systems involved accumulate more training data and more real-world examples of the pattern they are trying to catch.

Four to eight months is not a number either AINET or this analysis would recommend treating as a precise, universal threshold — the study measured one topic, one query, one pair of search engines, and the actual delay in any other content category, market or search engine could plausibly run shorter or considerably longer depending on factors this experiment did not test, including how quickly real user engagement data accumulates for a given query’s search volume, how aggressively a given search engine’s spam and quality systems have been tuned for that specific content vertical, and how distinctive the AI-generated content’s weaknesses are relative to genuinely strong human-written alternatives competing against it. What the window does usefully establish is an order of magnitude: not days, not weeks, but a genuine fraction of a year — long enough that any business measuring AI content performance on a monthly or quarterly dashboard needs to extend its evaluation horizon well past the point where early results look most encouraging, and long enough that the mid-experiment reversal this study documents is a realistic risk rather than a theoretical one for any publisher currently relying on early AI-content performance data to justify further investment. The strongest performer among the five AI systems in AINET’s test was not a conventional chat-based model but an agentic system, and that detail connects to a broader shift in how AI tools are being built and deployed across the content and software industries during 2025 and 2026 — a shift toward systems designed to plan, execute and revise multi-step tasks autonomously, rather than systems that produce a single response to a single prompt and stop.

The distinction matters for content production specifically because a single-prompt generation, however capable the underlying language model, structurally cannot replicate what a genuine editorial process does: research a topic from multiple angles, draft an initial version, identify gaps or weak claims in that draft, revise specifically to address them, and iterate until the output meets a quality bar the first pass did not reach. An agentic system built to execute exactly that kind of multi-step workflow, even without any human involvement at all, is closer in process — though not necessarily in outcome — to how a genuine expert writer actually works than a single chat completion is, and AINET’s result, while limited to one example within one experiment, is at least directionally consistent with that process difference producing a substantially different output quality.

This is also the reason this analysis has repeatedly resisted treating the Manus finding as evidence that any single “better” AI tool solves the underlying problem this study documents. The relevant variable the Manus result points toward is not which specific AI product a publisher chooses, but whether the production process behind any piece of content — AI-assisted or not — includes genuine iteration, revision and quality-checking before publication, rather than a single pass treated as finished. A single-prompt output from an agentic system used carelessly would plausibly show the same weaknesses as a single-prompt output from a conventional chat model; conversely, a conventional chat model’s output, put through a genuine multi-round human revision process before publication, would plausibly perform closer to how the human-written content in this study performed than to how the four underperforming AI texts did. The tool matters less than whether the process built around it includes the kind of iterative scrutiny that separates a first draft from a finished, expert-reviewed piece of content — a distinction the agentic-AI trend makes newly automatable in principle, but does not make automatic in practice simply by existing.

Every competitor running the same AI prompt changes the comparison

One structural feature of AINET’s test deserves more attention than it typically receives in summaries of the study: the AI-generated texts were not just AI-generated in isolation, they were generated from a shared prompt across five different systems answering the same brief on the same invented topic, which means the AI-content group in this experiment was competing not only against human-written content but, implicitly, against a version of itself produced by four other tools working from in effect the same instructions.

This is close to a worst-case scenario for AI-generated content’s prospects of differentiation, and it is also an increasingly realistic description of what is actually happening across large parts of the commercial web as more publishers adopt similar AI content workflows. If a notable share of businesses in a given industry are generating content using comparable AI tools, comparable prompts drawn from similar competitor research, and comparable publishing templates, the resulting content across that industry converges toward a narrower range of structures, phrasings and coverage angles than independent human writers, each bringing distinct professional backgrounds and idiosyncratic framing, would naturally produce. A search engine evaluating that increasingly homogeneous field of AI-generated competitors has, almost by construction, an easier differentiation task than it would evaluating a field of genuinely varied human-written content — not because the AI content is necessarily worse in any absolute sense, but because its sameness relative to its own closest competitors makes it harder for any individual piece to establish the kind of distinctiveness that both Google’s originality standard and the Princeton GEO research identify as a primary driver of ranking and citation success.

This dynamic plausibly intensifies over time rather than easing, which has a direct bearing on how durable AINET’s eight-month finding should be expected to remain as a predictor of future behaviour. As more publishers across more industries adopt generative AI for content production — a trend Graphite’s web-scale data shows has already reached roughly half of newly published web content — the pool of AI-generated text available for any given topic grows increasingly convergent and increasingly difficult for individual publishers to noticeably differentiate from, which should, all else equal, make the specific competitive advantage of genuine human expertise and original observation more useful over time rather than less, precisely because it becomes rarer relative to an increasingly homogeneous field of competing AI output answering the same underlying prompts. Comparing how different outlets framed AINET’s findings when they broke at the end of October 2026 offers a useful secondary signal about which elements of the study different audiences found most newsworthy, and surfaces a few framing choices worth flagging explicitly for readers relying on secondary coverage rather than the primary ppc.world account this analysis has drawn from directly.

Trade coverage aimed at the Russian digital marketing community, including CNews and Comnews, led with the headline comparative figure and the agency’s own framing around AI content factories, closely mirroring the ppc.world piece that originated most of the public reporting on the study. Coverage aimed at e-commerce sellers, including a writeup from the marketplace-focused outlet Uniseller, reframed the finding specifically around implications for product listing content on Russian marketplaces — an extension of the study’s conclusions into a specific commercial vertical that AINET’s own experiment, run in the beauty-content space rather than e-commerce product listings, did not directly test, and a reminder of how quickly a controlled study’s narrow, specific finding gets generalised into broader industry advice once it enters wider circulation.

This analysis has deliberately tried to avoid that same generalisation drift, which is why the sector-impact discussion earlier in this piece explicitly frames its extension of AINET’s findings to e-commerce, finance, legal and local-service content as informed judgment rather than as additional direct evidence, and why the limitations section catalogues the specific scope boundaries — sample size, topic narrowness, the single invented keyword, the eight-month ceiling on measured duration — that any responsible reading of this study needs to carry forward. The gap between what a controlled experiment actually measured and what trade press coverage subsequently claims it proved is a recurring feature of how research travels through an industry eager for clean, quotable conclusions, and readers encountering this study through secondary summaries rather than the primary reporting would do well to apply the same scrutiny to those summaries that this piece has applied to the study itself.

AI Overviews change the calculation for every page on this list

A factor neither AINET’s experiment nor most of the coverage around it addresses directly, but one that has become increasingly central to how organic search visibility translates into real traffic over the course of 2025 and 2026, is the growing prominence of Google’s AI Overviews feature — an AI-generated summary that appears above traditional organic listings for a large and growing share of search queries, compressing the space available to the ranked results this study measured and, in many cases, providing searchers with an answer that reduces or eliminates the need to click through to any individual ranked page at all.

This matters for interpreting AINET’s findings because it changes what a given ranking position is actually worth in practice, independent of the ranking position itself. A page sitting at position four, as the human-written content group averaged in Yandex by month eight, may capture much less real traffic today than an identical ranking position would have captured several years ago, simply because an AI-generated summary sitting above the organic results has already answered a share of the searches that would otherwise have generated a click. The compounding mechanism Dombrovsky described — organic ranking loss feeding into AI-answer visibility loss — runs through exactly this feature, since Google’s AI Overviews draw on the same underlying evaluation of page quality and trustworthiness that determines organic ranking, meaning a page that loses organic position for the reasons AINET’s study documents is simultaneously less likely to be drawn into the AI Overview summary that increasingly determines whether a searcher clicks through to any page at all.

For publishers evaluating this study’s relevance to their own traffic numbers, this adds a layer of urgency beyond what the raw visibility-share percentages convey on their own. The gap between human-written and AI-generated content’s performance, measured purely in ranking position, is already substantial by AINET’s final-checkpoint data. The real-world traffic gap that ranking difference produces is plausibly larger still, once the additional filtering effect of AI Overviews and similar generative search features — features that are expanding in coverage and prominence across both Google and other search platforms throughout the period this study was conducted — is factored into how much of that ranked, visible traffic ever reaches an actual click. AINET’s experiment answers a specific question well and leaves several adjacent questions open, and naming the most important of those openly is more useful to the field than treating this one study as a closed case. A credible follow-up, building directly on AINET’s design, would be worth running for several reasons this analysis has surfaced along the way.

A larger sample across more topics and keywords would address the single biggest limitation already flagged: five pages per group, on one invented keyword, cannot rule out the possibility that one or two unusually strong or weak individual pieces are driving the aggregate result. Repeating the design across a dozen or more invented keywords spanning several content verticals — not just beauty, but finance, local services and product categories as well — would let researchers test whether the magnitude and timing of the human-versus-AI gap AINET found holds consistently across topics, or varies by vertical in the way the sector-impact analysis earlier in this piece speculated it might.

A genuine hybrid-content arm — AI-generated first drafts substantively revised by the same professional experts who wrote the fully human-written comparison group — would directly test the production model Minyazhetdinov recommended rather than only the two extremes AINET actually compared. This is, by a wide margin, the single most commercially useful experiment the industry does not yet have clean data on, since it is the production model most businesses are actually adopting or considering, and neither AINET’s study nor Graphite’s web-scale research can currently isolate its performance cleanly from either pure human writing or pure AI generation.

Extending the measurement window well past eight months would test whether the reversal AINET documented represents a stable equilibrium or merely the latest swing in a trajectory that continues shifting with more time — a question this piece’s limitations section raised directly and that only additional longitudinal data, ideally from the same researchers using the same controlled infrastructure, can actually resolve. Until that kind of extended, expanded replication exists, AINET’s eight-month window remains the most rigorous single data point available on this specific question, genuinely informative within its scope, and appropriately treated as a strong early signal rather than a final, permanently settled verdict.

Strategic outlook: a drafting tool, not a byline

Pull together everything this piece has examined — AINET’s controlled eight-month result, Graphite’s independent web-scale measurement, the Princeton GEO research on what generative answer engines actually reward, Google’s own escalating policy response, and the historical precedent of the content farm era that preceded generative AI by over a decade — and the strategic picture that emerges is more settled than the ongoing industry argument about AI content usually suggests.

Generative AI, used as the sole and unsupervised production method for content meant to carry a business’s search visibility and reader trust, has a documented and now multiply-corroborated tendency to underperform genuinely expert, original, human-authored content over a multi-month horizon, even in cases where it performs competitively or better in the short term. That is no longer a single agency’s claim resting on a single experiment; it is the consistent direction of every major piece of evidence examined in this analysis, produced by unrelated organisations using different methods in different markets. The debate the industry has been having — AI content versus human content, as though the two were mutually exclusive production strategies locked in a zero-sum contest — is, on this evidence, the wrong debate to be having. The more useful framing, and the one both AINET’s own executives and Google’s published guidance converge on independently, treats generative AI as a drafting and research-acceleration tool operating inside a process still anchored by real expertise, genuine fact-checking and deliberate editorial judgment, not as a replacement for the byline that process produces.

The businesses best positioned for what comes next are not the ones that have banned AI tools from their content operations, nor the ones that have scaled unsupervised AI production to its fastest possible volume. They are the ones treating the speed generative AI genuinely offers as capacity freed up for more expert review, more original research and more rigorous fact-checking per piece of content — rather than capacity spent on producing more pieces of content with less human judgment behind each one. That is the production model AINET’s own experiment never actually tested, but that its findings, read honestly alongside everything else assembled in this analysis, point toward as the one most likely to survive the next eight months of search engine scrutiny rather than being quietly corrected by it. It is worth pausing on why AINET chose beauty-industry content specifically for this test, beyond the practical reason that the agency likely had ready access to professional beauty experts willing to participate. The choice also reflects something structural about the category that makes it a genuinely informative testing ground for the human-versus-AI question, rather than an arbitrary topic selected for convenience alone.

Beauty and skincare content sits at an interesting midpoint between the kind of content where expertise is hardest for an AI system to fake convincingly and the kind where readers have the least tolerance for generic, impersonal advice. Unlike a narrowly technical field such as tax law or clinical medicine, where factual accuracy dominates and a knowledgeable AI system drawing on accurate training data might plausibly produce defensible content, beauty and skincare advice draws heavily on the kind of embodied, first-hand, product-specific and skin-type-specific experience that a professional practitioner accumulates through years of direct client work — the sort of detail that resists generic synthesis precisely because it is anecdotal, situational and hard to source from any single training corpus. At the same time, unlike a highly niche technical field with a small, specialised readership, beauty content serves an enormous, engaged, repeat-visiting audience with strong pre-existing opinions about what credible expertise in the category looks like, which plausibly makes both search engines and readers more attuned to generic-feeling content in this category than in a more obscure or purely factual one.

This combination — hard for AI to fake convincingly, easy for both algorithms and readers to recognise when faked — makes beauty content a demanding test case rather than a soft one, and arguably a stricter test of the underlying hypothesis than an easier category would have provided. If AINET’s human-versus-AI gap had emerged in a category where genuine expertise is harder to define or demonstrate, a skeptic could reasonably argue the result reflected the category’s particular quirks rather than a generalisable pattern. That beauty content, specifically chosen for its reliance on exactly the kind of experiential, person-specific knowledge an AI system struggles hardest to replicate, produced such a clear result is, if anything, a point in favour of the finding’s broader relevance rather than a reason to discount it as a category-specific artefact.

The cost of waiting versus the cost of acting too early

Every business reading this study faces a timing problem that the data itself helps clarify but cannot fully resolve: act on the finding immediately, by overhauling content production processes now, or wait for further confirmation given the acknowledged limitations in sample size and scope this piece has catalogued at length.

The asymmetry in that decision is worth spelling out plainly, because it favours earlier rather than later action for most of the businesses this study’s findings are likely to be relevant to. The cost of acting on this finding and turning out to be wrong — investing more heavily in expert review and original content development than a weaker version of the pattern would have strictly required — is a cost of modestly reduced efficiency in content production, a real but bounded and recoverable cost. The cost of waiting for additional confirmation and the underlying pattern turning out to be real, meanwhile, compounds in exactly the way this piece has described throughout: months of accumulating AI-generated content built on a production model that AINET’s data, Graphite’s independent corroboration, and the historical Panda precedent all suggest carries a substantial and largely predictable risk of delayed ranking decline, with that decline arriving only after the content volume — and the sunk cost attached to it — has already grown considerably larger than it would have been had the correction come earlier.

This asymmetry is the reason this analysis has consistently framed the practical recommendations throughout this piece as reasonable to act on now, rather than as conclusions to hold in reserve pending a larger, more definitive replication study that may be years away. A business does not need certainty that AINET’s exact percentages will reproduce precisely in its own industry and market to reasonably conclude that the underlying mechanism — search and AI-answer systems eventually favouring demonstrated, original expertise over generic, convergent AI output — is worth designing a content strategy around today, particularly given how directly that conclusion aligns with Google’s own explicitly published quality standards, standards the company has stated plainly rather than left for outside researchers to infer. Treating this study as sufficient evidence to act on, while remaining appropriately clear-eyed about its specific scope and limitations, is a more defensible position than either dismissing it as too narrow to matter or treating its headline percentage as a universal law. If one concrete, operational change follows most directly from everything examined in this piece, it is in how a content brief itself gets written before any drafting — human, AI-assisted or otherwise — begins, because the gap AINET’s experiment measured traces back, more than any other single factor, to what the underlying brief did and did not ask for.

A brief that specifies only a topic and a target keyword, leaving the actual substance of the piece entirely to whoever or whatever executes it, produces exactly the kind of generic, convergent output this study found underperforming — because nothing in that brief forces the inclusion of the specific, attributable, original detail that both AINET’s data and the Princeton GEO research independently identify as the differentiating factor between content that holds its search and AI-answer visibility and content that eventually loses it. A brief that instead requires a named, verifiable source for every factual claim, at least one piece of information or observation the writer or system could only know from direct professional experience, and explicit identification of what makes this specific piece different from the dozen most obvious competing answers to the same query builds the originality and specificity standard directly into the production process, rather than hoping it emerges incidentally from whoever happens to be doing the writing.

This reframing has a practical advantage worth stating directly: it works regardless of whether the execution involves a human writer, an AI system, or some combination of both, which sidesteps the increasingly unproductive debate over banning or mandating specific tools in favour of a standard that can be enforced at the brief and review stage regardless of production method. A content operation that rewrites its briefs to demand this kind of specificity, and that builds editorial review around verifying the brief was actually fulfilled rather than merely checking that a word count and keyword target were hit, is applying the single clearest lesson AINET’s experiment offers without needing to resolve every open question this analysis has raised about the study’s precise scope, sample size or generalisability — because the standard the brief enforces is the same standard the evidence, across every source examined in this piece, converges on as the one that actually determines durable visibility.

Taking the study at the scale it actually supports

Every section of this analysis has tried to hold two things in tension that are easy to collapse into each other: taking AINET’s experiment seriously as genuinely rigorous evidence, and refusing to inflate its specific, narrow findings into a universal claim the data cannot support. That tension is the right place to leave this discussion, rather than resolving it in either direction for the sake of a cleaner ending.

What the evidence supports, stated at the scale it actually earns: on one invented keyword, in one content vertical, across two major search engines, over eight carefully measured months, professionally written expert content ended up with roughly 2.5 times the search visibility of AI-generated content answering the same brief, after both content types had passed through an opening period in which the AI-generated group performed competitively and, for several months, better. That same directional pattern — AI-generated content underrepresented in both organic search rankings and AI-answer citations relative to its share of overall web production — shows up independently in large-scale web research covering a different language, a different set of search platforms, and tens of thousands of real pages rather than ten controlled ones. Google’s own published quality standards, developed and revised across more than a decade of responding to exactly this category of production-at-scale problem, describe criteria — effort, originality, demonstrated skill, accuracy — that plausibly explain why content produced primarily through unsupervised AI generation would tend to underperform content produced with genuine human expertise behind it, without requiring any claim that search engines specifically target AI as a category rather than the low-originality output AI generation frequently, though not universally, produces.

None of that adds up to proof that any specific business’s AI content strategy will fail, nor that any specific piece of human-written content will succeed. It adds up to a well-corroborated pattern, worth taking seriously by anyone making content production decisions today, and worth continuing to test rather than treating as permanently settled — which is precisely the posture this analysis has tried to model throughout: specific about what the evidence shows, explicit about where it stops, and honest that the most useful thing a single study like this one can do is change which questions get asked next, not end the conversation.

A closing note on what “written by AI” will keep meaning

The category this entire study investigates — “AI-generated content” as a single, stable thing that can be measured and compared against “human-written content” — is itself a moving target, and it is worth closing on exactly how fast that target is likely to keep moving, because it bears directly on how long AINET’s specific numbers should be expected to remain representative of current AI output quality.

The five systems AINET tested in 2026 — Manus, ChatGPT, Perplexity, Gemini and DeepSeek — represent a snapshot of generative AI capability at one point in a field that has shown rapid, continuous improvement since the public launch of ChatGPT in November 2022. Graphite’s own detection-reliability data, cited earlier in this piece, already shows genuinely different false-negative rates when testing detectors against output from newer frontier models compared with older ones, which is itself evidence that the underlying quality and distinctiveness of AI-generated text has shifted materially even within the roughly three-year span these studies collectively cover. It would be a reasonable expectation, consistent with that trend, that AI systems tested a year or two from now would produce output that is harder to distinguish from expert human writing, on the specific dimensions — originality, specificity, demonstrated first-hand knowledge — that AINET’s experiment found lacking in four of its five AI-generated test articles.

This is not a reason to discount the study’s current relevance, but it is a reason to treat its specific numbers as a measurement of a particular moment in a fast-moving technology, rather than as a fixed law about what AI-generated content will always look like. The Manus result already demonstrates, within this single experiment, that the gap between “AI-generated” and “indistinguishable from expert human writing” is not fixed even today — it varies by system, by production process, and plausibly by how much iterative refinement sits between a raw prompt and a published piece. The honest long-term expectation this analysis can offer is that the underlying standard search and AI-answer engines are evaluating against — genuine effort, genuine originality, genuine demonstrated expertise, genuine accuracy — is unlikely to change even as the tools used to meet or fail that standard keep improving, which means the practical lesson of this study is more durable than its specific headline percentage. Businesses that build their content strategy around clearing that underlying standard, by whatever combination of human and AI effort actually clears it, are positioned to remain visible regardless of how the next generation of AI writing tools narrows or widens the gap AINET measured in late 2026.

Eight months, not eight weeks, was the only honest test

It is worth returning, in closing, to the methodological choice that made this entire study worth writing about rather than another forgettable SEO case study: AINET’s decision to run the experiment for eight full months rather than stopping at the first clear result.

Most content comparisons published in the SEO industry report a result weeks after launch, because that is when a publisher or agency has enough data to write a confident-sounding blog post, and because few organisations have the patience or budget to keep a controlled test running for the better part of a year with no guarantee the eventual result will differ from the early one. AINET’s own data demonstrates, in the starkest possible terms, exactly why that shorter horizon produces unreliable conclusions: a test stopped at week three would have shown AI content competitive with human content in Google and clearly behind in Yandex; a test stopped at month four would have shown AI content winning in both. Only the full eight months revealed the pattern that actually held once the dust settled — and even that endpoint, as this analysis has stressed throughout, is not guaranteed to represent a permanently stable final state rather than one more point on a longer curve.

The discipline of waiting for the slower, less convenient answer, rather than publishing the first answer that arrives, is the actual methodological contribution this study makes to an industry that badly needs more of it. Every number in this piece — AINET’s visibility percentages, Graphite’s web-scale detection rates, the Princeton GEO benchmark’s citation-lift figures, the multi-year arc of Google’s scaled content abuse enforcement — rewards exactly this kind of patience, and punishes exactly the kind of early, confident conclusion-drawing that a faster-moving industry constantly incentivises. A business, an agency or a researcher that internalises nothing else from this analysis should internalise that one discipline: measure long enough to see the reversal before deciding there isn’t one coming.

Common questions about the AINET study and AI content performance

Author:
Jan Bielik
CEO & Founder of Webiano Digital & Marketing Agency

This article is an original analysis supported by the sources cited below

Видимость авторских текстов в Поиске вдвое выше, чем нейросетевых ppc.world’s report on AINET’s eight-month experiment, sourced directly from the agency’s press office, containing the full visibility and average-position figures at each checkpoint along with quotes from Yaroslav Dombrovsky and Ruslan Minyazhetdinov.

Авторские тексты обошли ИИ в поиске до 2,5 раза по видимости Analito’s coverage of the same AINET study, summarising the methodology, the three measurement checkpoints, and the agency’s business recommendations for using AI in content production.

Авторы против нейросетей: экспертные тексты обошли ИИ до 2,5 раз CNews trade coverage of the study’s release, naming AINET’s leadership and corroborating the headline findings reported elsewhere.

Авторы против нейросетей. Экспертные тексты обошли искусственный интеллект до 2,5 раз Comnews’ report on the same study, part of the wave of Russian trade press coverage that followed the study’s late-October 2026 release.

What web creators should know about our March 2024 core update and new spam policies Google’s own announcement introducing the scaled content abuse policy, explaining why it replaced the earlier automatically-generated-content policy and how it applies regardless of production method.

Does Google Penalize AI Content? No – But It Punishes This An analysis of Google’s scaled content abuse policy text and how it defines the violation by purpose and scale rather than by whether AI was involved in production.

Google adds main content, effort and originality criteria and a fake author warning to its helpful content guidance Coverage of Google’s October 2026 update to its helpful content guidance, detailing the four rater attributes — effort, originality, talent or skill, and accuracy — and the new warning against fabricated author identities.

Google Search Quality Rater Guidelines: Key Insights About AI Use A detailed breakdown of the January 2025 update to Google’s Search Quality Rater Guidelines, including the section 4.6.6 language on content created with little to no effort, originality or added value.

Scaling AI Content Is The #1 Enterprise Priority: How Do You Scale Without Penalty? Search Engine Journal’s reporting on Google’s John Mueller discussing the updated Quality Rater Guidelines and how they treat AI-generated content grouped under low-effort, low-originality standards.

A Complete Guide to the Google Panda Update: 2011-21 A history of the Panda update, including its original February 2011 launch, its stated goal of targeting shallow and low-quality content, and Demand Media’s scale at the time as a content producer.

Google Panda Update: eHow Untouched, Says Pleased with Progress on Content Quality Contemporary reporting on eHow’s early escape from the first Panda rollout and Demand Media’s public response, documenting the gap between the update’s intended and actual initial effect.

U.S. Google Panda Update, Round 2: Shrinking Demand for eHow Coverage of the later, more decisive Panda update wave that reduced eHow’s search visibility months after the initial rollout had largely spared the site.

Nearly half of online articles are now AI-generated: Study Search Engine Land’s coverage of Graphite’s web-scale research tracking the share of newly published English-language articles classified as primarily AI-generated from 2020 through early 2026.

Five Percent — More Articles Are Now Created by AI Than Humans Graphite’s own published research and methodology for measuring AI-generated content prevalence across a large Common Crawl sample, including its detector-averaging approach and reported false-negative and false-positive rates.

Five Percent — How Does AI-Generated Content Perform in Search and Answer Engines? Graphite’s companion research measuring how AI-generated content performs specifically in Google Search rankings and in ChatGPT and Perplexity citations, including the 86% human/14% AI organic search split and the 82%/18% citation splits for both answer engines.

Exclusive: AI writing hasn’t overwhelmed the web yet Axios’s reporting on Graphite’s findings, including the plateau in AI content production volume and the search-performance data suggesting publishers are adjusting strategy in response to weaker AI-content ranking outcomes.

GEO: Generative Engine Optimization An overview of the original 2023 Princeton-led research paper that coined the term Generative Engine Optimization, describing its GEO-bench methodology and the content strategies found to most increase visibility inside generative AI answers.

Generative Engine Optimization (GEO): The Full Definition, Explained A summary of the Princeton GEO study’s findings on which content modifications most increase a source’s likelihood of being cited inside a generative engine’s synthesised answer.

What Is Scaled Content Abuse? An explanation of Google’s scaled content abuse policy and supporting data on the share of AI-generated content among top-ranking search pages as of mid-2026.

Google Forecloses On Content Farms With “Panda” Algorithm Update Search Engine Land’s original 2011 reporting on Google’s stated intent to target shallow, low-quality content farms, including the company’s own definition of the problem the Panda update was built to address.

This article was prepared with the assistance of artificial intelligence tools. The content underwent expert human review, and Webiano Digital & Marketing Agency assumes editorial responsibility for its final version and publication.