In June 2024 Perplexity’s chief executive told Lex Fridman that footnoting every sentence would make AI answers trustworthy and leave Google stuck with its links. Two years of tests, crawler disputes and lawsuits show the footnote solved provenance, not accuracy, and that the citation itself has become the evidence publishers use against answer engines.
Table of Contents
A footnote in an AI answer tells you where a sentence came from. It does not tell you the sentence is right, and it does not pay the person who wrote the original. That gap now sits at the centre of AI search. The best-scoring tool in a March 2025 Tow Center for Digital Journalism test still misidentified the source of 37% of news excerpts, and the publishers being cited are in federal court arguing that a wrong answer under their masthead is a misattribution.
The clearest statement of the opposite bet runs for three hours. On 19 June 2024, Lex Fridman published episode #434 of his podcast, a conversation with Aravind Srinivas, chief executive of Perplexity. Srinivas described an answer engine built on an academic rule, argued that Google’s link-based ad business would stop it from following, and said Perplexity was in no hurry to sell ads. The interview is friendly: Fridman opens by recommending the product to listeners, and the episode carries six sponsor reads, none from Perplexity. Each of those claims can now be checked.
An answer engine built on the academic footnote
Srinivas lays out the pipeline early in the episode. Conventional search retrieves links, the system pulls the relevant paragraphs, and a large language model writes a short answer with a footnote on every sentence. The rule came from his first research paper, where senior co-authors told him that “every sentence you write in a paper should be backed with a citation”. Wikipedia’s sourcing standards were the second model.
He was open about the trade-off. The team chose not to give links the prominent spot on the page, accepting some wrong answers, on the bet that hallucinations would fall steeply as models got cheaper and the index got fresher, leaving a long tail that would keep shrinking. Later he describes crawling as a matter of manners: respect robots.txt, respect crawl delays, skip what a site disallows. He also explains what he calls answer engine optimization, where a site hides invisible text instructing an AI what to say, and admits that defending against it is a reactive game.
The Tow Center test of 1,600 cited answers
In February 2025, Tow Center researchers took excerpts from 200 articles across 20 publishers and asked eight AI search tools to name each article’s headline, publisher, date and URL. They picked excerpts that returned the original in Google’s top three results, so the task was easy for a classic search engine. Collectively the tools answered more than 60% of the 1,600 queries incorrectly. Perplexity did best at 37% wrong; Grok 3 was wrong on 94%. Perplexity’s citation-first design beat every rival tested, and that part of the 2024 pitch holds.
The test, though, measured exactly what a footnote exists to do: find the source and point to it. Perplexity Pro answered more prompts correctly than the free version yet had a higher error rate, because it answered confidently where it should have declined. Despite a partnership with the Texas Tribune, Pro cited syndicated copies of Tribune stories in 3 of 10 cases. A miss rate above a third, eight months after the promise of steeply falling errors, is far from a long tail. The researchers note that each excerpt was queried once and outputs vary between runs.
Robots.txt manners against the stealth crawler finding
The same study found that free Perplexity correctly identified all ten paywalled National Geographic excerpts, although the publisher blocks Perplexity’s crawler and has no deal with it. The authors allowed that the text might have reached the tool by other routes. Cloudflare went further on 4 August 2025. Customers had blocked PerplexityBot and Perplexity-User in robots.txt and in firewall rules, yet Perplexity still answered questions about their pages. Cloudflare registered fresh, unindexed test domains that disallowed all bots, and Perplexity still described their content.
Cloudflare counted 20 to 25 million daily requests from the declared agent and 3 to 6 million from an undeclared agent posing as Chrome on macOS, rotating through IP addresses and networks outside Perplexity’s published ranges across tens of thousands of domains. It removed Perplexity from its verified-bot list. Perplexity’s rebuttal, as reported by Search Engine Journal, argued that its assistants fetch pages on demand for a specific user question, store nothing, train nothing, and should not be judged by crawler rules. The two accounts still disagree on who did the fetching. For a publisher the practical result is the same: a robots.txt line no longer settles whether an answer engine reads a page.
Google moved the ad inside the answer
The strongest argument in the episode concerns Google’s margins. Srinivas reasoned that any answer unit which discourages a link click takes money from Google’s most profitable product, so Google would not push hard on answers, much as it left cloud computing to Amazon. The diagnosis was correct. In Pew Research Center data from 900 US adults in March 2025, users who saw an AI summary clicked a traditional result on 8% of visits, against 15% without one. They clicked a link inside the summary on 1% of visits and ended the session outright on 26% of pages, against 16%.
The prediction was wrong. Google introduced AI Overviews in 2024, put ads into them on US mobile that October, and on 21 May 2025 began testing ads below and inside AI Mode responses while expanding AI Overview ads, TechCrunch reported. With $66.89 billion in ad revenue in the first quarter of 2025 alone, Google solved the margin problem by moving the ad into the answer instead of protecting the link.
Perplexity’s own ad test ended where the interview hinted
On ads for Perplexity itself, Srinivas was cautious: subscriptions meant no desperation, a Netflix-style hybrid might work, and any format had to avoid costing user trust. Five months later, on 12 November 2024, the company began showing sponsored follow-up questions and side placements in the US. Its blog post promised advertisers would not shape answers and said subscriptions alone could not fund a sustainable revenue share for publisher partners.
By February 2026 the experiment was over. The Decoder, citing the Financial Times, reported on 18 February that Perplexity had been pulling ads since late 2025 because an executive said they made users question every answer. The company now calls itself an accuracy business, sells plans from $20 to $200 a month, and claims more than 100 million users at an $18 billion valuation. The 2024 caution proved right, but the asymmetry ran the other way: the incumbent pushed ads into answers and the startup retreated. The sources reviewed do not say what now funds the publisher revenue share that ads were meant to pay for.
Hallucinations now arrive with a publisher’s name attached
Dow Jones and the New York Post sued Perplexity in 2024, and on 21 August 2025 Judge Katherine Polk Failla refused to dismiss the case or move it to California, Bloomberg Law reported. On 4 and 5 December 2025, the Chicago Tribune and The New York Times filed their own suits in the same Manhattan court. Besides copyright claims over both inputs and outputs, both papers brought trademark claims alleging that Perplexity generates hallucinations and shows them alongside their marks, according to Law360. The Times had sent a cease-and-desist letter in October 2024. Encyclopaedia Britannica, Merriam-Webster and Reddit were pursuing claims of their own at the time.
This is the turn the 2024 interview did not anticipate. The footnote converts a wrong answer into a claim about a named brand, and the citation that was meant to make an answer checkable becomes the exhibit. These are allegations; none of the sources reviewed reports a ruling on the merits.
Footnote habits for publishers and answer-engine readers
For publishers, a citation is visibility, not traffic. The Pew figure of one click per hundred summary visits makes that plain. Access decisions now belong at the network level as well as in robots.txt; Cloudflare says its managed rule blocking AI crawlers is open to all customers, free plans included. Because false attribution is now a litigated harm, checking what assistants say under your name deserves the same routine as rank tracking.
For readers, the footnote is a pointer to open before acting on anything that matters. The Tow results show a paid tier is no guarantee, and the hidden-text manipulation Srinivas himself described means a cited page can carry instructions written for the machine, not for you.
The measurement that would settle the 2024 bet
The evidence supports one part of the episode: sourcing every sentence gave Perplexity better provenance than its rivals. Accuracy and publisher economics did not follow. The uncertainty that matters most is legal, whether courts treat answers built on retrieved articles as substitutes for those articles. Two developments would move the judgment: a rerun of the Tow test showing miss rates well below a third, which would vindicate the promise of falling errors, and a merits ruling in one of the Manhattan cases. Until then, a footnote proves where a sentence came from, and little else.
Author:
Jan Bielik
CEO & Founder of Webiano Digital & Marketing Agency
This article is an original analysis supported by the sources cited below
Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet | Lex Fridman Podcast #434
The episode under review, published on 19 June 2024, with its sponsor list and chapter outline.
Transcript for Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet | Lex Fridman Podcast #434
The human-generated transcript used for the citation rule, the hallucination bet, the robots.txt remarks, the Google margin argument and the ads stance.
Why we’re experimenting with advertising
Perplexity’s 12 November 2024 announcement of sponsored questions, its promise of uninfluenced answers and its link between ads and publisher revenue share.
Perplexity is using stealth, undeclared crawlers to evade website no-crawl directives
Cloudflare’s 4 August 2025 test results, request volumes for declared and undeclared agents, and its decision to de-list Perplexity as a verified bot.
AI Search Has a Citation Problem
The Tow Center’s 1,600-query test of eight AI search tools, including Perplexity’s 37% error rate, the Pro tier’s confident errors and the National Geographic finding.
Google users are less likely to click on links when an AI summary appears in the results
Pew browsing data showing 8% versus 15% click rates, 1% clicks on cited links and higher session-ending rates with AI summaries.
Google is bringing ads to AI Mode
Report of Google’s May 2025 test of ads inside AI Mode, its expansion of AI Overview ads and its first-quarter 2025 ad revenue.
Perplexity AI Can’t Escape Dow Jones Training Copyright Suit
Report of the 21 August 2025 order keeping the Dow Jones and New York Post case in Manhattan federal court.
NYT, Chicago Tribune Sue Perplexity Over ‘Verbatim’ Outputs
The December 2025 suits, their trademark claims over hallucinations shown with the papers’ marks, and the other pending publisher claims.
Perplexity pulls advertising from its search engine, calling itself an “accuracy business”
Report of Perplexity’s February 2026 exit from ads, its trust rationale, subscription pricing and claimed user numbers and valuation.
Perplexity Says Cloudflare Is Blocking Legitimate AI Assistants
Summary of Perplexity’s rebuttal that user-triggered assistants fetch on demand, store nothing and should not face crawler rules.
Validátor prešiel bez chýb aj varovaní. Telo článku má 1 396 slov, 8 sekcií a 11 zdrojov z 11 domén. Stav súdnych sporov som uviedol len tak, ako ho potvrdzujú otvorené zdroje: po zamietnutí návrhu na zrušenie konania v auguste 2025 už žiadny z nich nepotvrdzuje rozhodnutie vo veci samej. Mám článok nahrať do webiano.digital ako koncept?
| Citing this article? Brief excerpts are welcome. Please credit Webiano.digital, name the author where stated, and include a link to https://webiano.digital and to this original article. Full or substantial republication requires prior written permission. Read our Copyright and Content Use Policy. |
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.
