Conversion intelligence is the advantage no ad platform can sell you

Meta delivered 19% more ad impressions in the first quarter of 2026 than in the same quarter a year earlier, and charged 12% more for the average one. Alphabet reported 13% more paid clicks on Search and a 5% higher average cost per click over the same period. Read those four numbers together and the shape of the market becomes clear: there is more reach available than at any point in the history of digital advertising, and it costs more every year.

The advantage moved from buying attention to knowing which attention pays

That combination is the end of a business model that worked for roughly fifteen years. Buying attention used to be a craft. Someone who understood keyword match types, audience layering, placement exclusions, and dayparting could buy the same inventory for less than a competitor could, and the gap between a good media buyer and an average one showed up directly in cost per acquisition. Anyone who ran accounts between 2012 and 2019 remembers the feeling of finding a cheap audience nobody else had noticed.

That craft has been automated away. Performance Max, AI Max, Advantage+ and their equivalents on TikTok and LinkedIn now decide placement, audience, bid, and increasingly the creative itself. The manual controls that used to separate operators have been folded into model inputs. Google’s own framing at Google Marketing Live in May 2026 was that execution friction is disappearing from advertising, and the honest reading of that framing is that the technical ability to build a campaign no longer differentiates anybody. Two agencies with the same budget, the same feed, and the same product will now buy roughly the same reach at roughly the same price.

What still differs, wildly, is what each of them knows about the outcome.

One team reports 4.2 ROAS from the platform interface and believes it. Another team knows that its branded search campaign produces almost no incremental revenue, that 40% of its reported conversions are modelled rather than observed, that its lead form generates three qualified opportunities per hundred submissions in one segment and eleven in another, and that a €14 first order in one product category is worth €95 over eighteen months while a €70 first order in another is worth €72 and never repeats. The second team pays the same CPMs as the first. It makes better decisions with them, and the compounding effect of better decisions on the same inventory is the only durable edge left in paid media.

This is what conversion intelligence means in practice, and it is a capability rather than a product. No platform can sell it, because the parts that matter are specific to one business: which outcomes count, what each is worth, and which spend actually caused them. Google will sell bidding automation that consumes conversion values. Meta will sell ranking models that consume conversion events. Neither can tell an advertiser whether the value it is feeding them is correct, and both have a structural interest in the number being generous.

The gap has widened for reasons that have nothing to do with marketing fashion. Browser and platform privacy changes broke the observational record of what happened after a click. Consent requirements in Europe removed a large share of users from measurement entirely. Google retired most of the Privacy Sandbox in October 2025 after six years of building it, leaving the industry with no standardised replacement for third-party identifiers. AI answer interfaces absorb a growing share of research behaviour and return less traffic and less referrer information than a blue link did. Each of these changes made reach easier to buy and outcomes harder to see.

The practical consequence is uncomfortable. Media buying skill has deflated in value while measurement skill has inflated, and most marketing organisations are still structured for the previous ratio. Agencies sell campaign management. Job descriptions ask for platform certifications. Reporting cycles end with a dashboard screenshot rather than a decision. The work that now separates a profitable account from an unprofitable one — event architecture, value modelling, consent handling, holdout testing, warehouse joins between ad platforms and CRM — sits awkwardly between marketing, analytics, and engineering, and frequently belongs to nobody.

The rest of this analysis works through what conversion intelligence is made of, what the current evidence says about each component, where the numbers are contested, and what a team can build in the order that produces the earliest return.

Conversion intelligence defined without the marketing gloss

Conversion intelligence is the ability to answer three questions about marketing spend with enough accuracy to act on the answers: what outcome actually occurred, what that outcome is worth, and what share of it was caused by the spend. A business that can answer all three is running conversion intelligence. A business that can answer none is running on platform-reported numbers and hoping.

The three questions decompose into distinct technical problems, which is why so many teams solve one and assume they have solved the set.

Occurrence is a data collection problem. Somebody submitted a form, bought a product, booked a call, activated a trial, or walked into a shop. Capturing that reliably now requires server-side collection, deduplication across client and server events, identity resolution across sessions and devices, and a defensible treatment of the users who never consented to measurement at all. Client-side pixels alone miss a structural share of real conversions because of browser storage limits, ad blockers, and consent refusal.

Worth is a modelling and business problem, and it is the one most often skipped. A conversion count treats every lead and every order as identical. Almost no business works that way. Gross margin varies by product. Return rates vary by category and by acquisition channel. Lead-to-close rates vary by source, geography, and offer. Repeat purchase behaviour varies enormously between first orders that look identical in the checkout. Assigning a defensible value to each conversion, and passing that value back into bidding, converts a volume machine into a profit machine.

Causation is an experimental problem, and no amount of tracking solves it. Attribution answers the question of which touchpoints preceded a conversion. It cannot answer whether the conversion would have happened anyway. That question requires a counterfactual — a group that did not receive the advertising — which means geo holdouts, in-platform lift tests, or a model calibrated by such tests. The gap between what platforms report and what experiments measure is not a rounding error, and evidence gathered across 2025 and 2026 puts platform-reported ROAS somewhere between 20% and 60% above measured incremental lift depending on channel and design.

Three properties separate conversion intelligence from ordinary analytics.

It is decision-shaped. The output is a change to a bid strategy, a budget allocation, a value assignment, or a campaign that gets switched off, not a report that circulates and then dies. A team that produces a monthly attribution deck nobody acts on has analytics, not intelligence.

It is adversarially aware. Every platform reports its own contribution using its own attribution window, its own modelling, and its own definition of a view-through. Those reports are inputs from interested parties. Conversion intelligence treats them the way a CFO treats a supplier’s own quality report: useful, directional, and audited against something independent.

It is fed back into the machine. This is the part that separates 2026 from 2019. Measurement no longer just informs humans. It trains the bidding models. When a business sends corrected conversion values, offline outcomes, lead quality scores, or lifetime value estimates back into Google Ads and Meta, it changes what those systems buy on its behalf. Better measurement now converts directly into better media buying without a human touching a bid. That feedback loop is the mechanism by which conversion intelligence compounds, and it is why the gap between well-instrumented and poorly-instrumented advertisers keeps widening rather than stabilising.

Two things conversion intelligence is not. It is not a dashboard, though it usually produces one. And it is not the same as conversion rate work on a landing page, though the two are related. Improving a checkout flow raises the conversion rate. Knowing which traffic is worth sending to that checkout, and what each completed order is worth, is a different discipline that governs where the money goes in the first place.

Reach turned into a commodity the moment the auction learned to price itself

An auction that prices inventory well removes the arbitrage that used to reward skilled buyers. That is the mechanism behind the commoditisation of reach, and it deserves a precise explanation rather than a slogan.

In a manual auction, the buyer supplies the intelligence. A human decides that women aged 25 to 34 in three cities who visited a product page in the last fourteen days are worth €4 a click while everyone else is worth €1.20. If that human is right and competitors are not, the account wins cheap inventory. The platform captures less than the impression is worth, and the buyer keeps the difference. Every genuine media buying edge from the last decade was some version of this: a segment, placement, or moment that the market had underpriced.

Machine bidding removes the underpricing. When Google’s or Meta’s models estimate conversion probability per impression using their own behavioural data and then bid to a target, the price of each impression converges toward its expected value. The arbitrage does not move to a different segment; it disappears, because the system that runs the auction now knows more about the user than any advertiser buying in it. Meta described its Andromeda retrieval system as delivering a 14% improvement in ads quality when it reported third quarter 2025 results, and extended its Adaptive Ranking Model to offsite conversions in the first quarter of 2026. Each of those upgrades makes the platform better at finding the person likely to convert — and better at charging for the privilege.

The pricing evidence follows directly. Meta’s average price per ad rose 12% year over year in the first quarter of 2026, having risen 10% in the same quarter of 2025. Regional detail sharpens the point: price per ad rose 19% in Europe, 14% in the United States and Canada, and only 5% in Asia-Pacific, where impression supply grew fastest at 23%. Where supply grows faster than demand, price growth softens. Where demand concentrates, price growth accelerates. That is a commodity market behaving normally, not a marketing platform with a targeting secret to sell.

Meta itself attributes the price increase to advertising demand, which it links to improvements in its own targeting and measurement tools. The circularity there is worth naming: better platform measurement raises advertiser willingness to pay, which raises prices, which returns the advertiser to the same cost per outcome they started with. Any capability the platform gives to everyone gets competed away in the auction. Only capabilities that a business holds privately survive contact with the bidding system.

Three forces are compressing the reach advantage further.

Automation converged across vendors. Performance Max, AI Max, Advantage+, Smart+ and their equivalents are structurally similar products. When every buyer runs broad targeting with automated placement, the differences between accounts shrink to feed quality, creative, budget, and the conversion signal being fed in.

Budget migration concentrated demand. eMarketer’s 2026 forecast has Meta’s family of apps producing $100.86 billion in net US digital ad revenue, a scale that reflects budget moving into social from other channels. Money chasing the same auctions raises clearing prices for everyone in them.

Creative production stopped being scarce. Meta reported that the number of advertisers using its AI creative tools doubled to eight million in the first quarter of 2026. Generative production removes the volume constraint that used to protect well-resourced brands. The constraint that remains is knowing which creative produced profitable outcomes, which is a measurement question rather than a production question.

None of this means reach is worthless. Distribution still matters enormously, and a business with no reach has no business. The claim is narrower and harder: reach is now a purchasable input at a market-clearing price, available to any competitor with a credit card, and therefore cannot be the basis of a defensible position. The buyer who wins is the one who knows something about outcomes that the auction does not.

Meta and Alphabet’s own quarterly numbers show where the pressure sits

Platform financial disclosures are the most reliable public data on advertising cost trends, because the incentives run against exaggeration and the numbers are audited. Both large platforms reported first quarter 2026 results on 29 April 2026, and the detail is more useful than any agency benchmark study.

Meta reported total revenue of $56.31 billion, up 33% year over year, with advertising revenue of $55.02 billion, also up 33%, or 29% on a constant currency basis. The composition matters more than the headline: ad impressions delivered rose 19% while average price per ad rose 12%. Family daily active people reached 3.56 billion in March 2026, up 4%. Impression growth of 19% against user growth of 4% means the increase came mostly from showing more ads to the same people, which Meta attributes partly to ad frequency on its products.

Alphabet reported total revenue near $109.9 billion for the quarter. Search and other revenue grew 19% to roughly $60 billion, with paid clicks up 13% and average cost per click up 5%. YouTube advertising contributed $9.88 billion, up around 11%. Google Cloud grew 63% to $20.03 billion. Two-thirds of the market commentary focused on Cloud, but the advertising detail is where the operational lesson sits.

Google Network revenue fell 4% to $6.97 billion — the only Google advertising line that declined. Network is the business of placing Google-brokered ads on third-party publisher sites. It had already declined 2% in the first quarter of 2025 and continued sliding through the year. Every line where Google owns the surface grew; the line that depends on the open web shrank. Philipp Schindler, Google’s chief business officer, told analysts that queries are at an all-time high and that AI Overviews and AI Mode continue to drive search usage growth. More queries answered inside Google’s own interface, with fewer clicks leaving it, compresses the value of inventory that lives outside it.

For anyone planning a 2026 budget against 2025 assumptions, the arithmetic is unforgiving. If auction prices rise 12% to 14% in a market and click-through rate stays flat, cost per click moves in the same direction. Holding cost per acquisition flat then requires the conversion rate or average order value to absorb the difference. A business that improved neither is paying more for the same customer and will see it in gross margin rather than in the ads dashboard, which will keep reporting a respectable ROAS throughout.

One further disclosure from the Alphabet call is worth holding onto, because it reframes the reach conversation entirely. Management noted that ads have historically appeared against roughly 20% of search queries. The remaining 80% of queries carry no advertising because commercial intent was unclear or the format did not fit. Google’s stated ambition with Gemini-powered ad formats is to read intent well enough to monetise more of that inventory. If it succeeds, the supply of commercially relevant reach expands again — which is another reason to expect the reach advantage to keep deflating rather than recover.

The capital expenditure numbers put a floor under the trend. Meta guided 2026 capital expenditure to $125 billion to $145 billion, raised from a prior range. Alphabet lifted 2026 capital expenditure guidance to $180 billion to $190 billion. Companies spending that much on inference infrastructure will keep pushing automation into every layer of the ad stack, and they will keep pricing the resulting improvement into the auction. The advertiser’s share of that improvement is whatever they can protect with private information about their own outcomes.

The signal collapse that made reach look cheaper than it is

Reporting bias is the reason so many advertisers believe reach is still working. When measurement degrades, the numbers do not become obviously wrong — they become quietly generous in some places and quietly conservative in others, and the mixture is what makes it dangerous.

The collapse happened in layers, each with a different mechanism.

Safari began restricting third-party cookies through Intelligent Tracking Prevention in 2017 and later capped the lifetime of first-party cookies set through JavaScript. Firefox followed with Enhanced Tracking Protection. Brave blocks by default. Estimates of the share of global web traffic that is already cookieless without any Chrome change cluster around 17% to 20%, and the figure runs materially higher in markets where Apple device share is high. Any conversion path that crosses those browsers is either lost or reconstructed by modelling.

Apple’s App Tracking Transparency, live since April 2021, required apps to ask permission before accessing the device identifier used for cross-app attribution. Opt-in rates never recovered to a level that supported deterministic mobile attribution. App marketers moved to aggregated postbacks with delayed, coarse, and partially null conversion data.

Consent requirements removed a further slice. Under the ePrivacy Directive and GDPR as enforced across the EEA, storing or reading identifiers for advertising needs consent, which means a share of European traffic is legally unmeasurable at the individual level. Published consent rates vary enormously by sector, banner design, and country, and vendor-published figures in the 30% to 50% acceptance range should be treated as directional rather than settled. What is not in dispute is the direction: a large minority to a majority of European sessions cannot be tied to an individual conversion record.

Ad blockers and network-level filtering remove another portion, weighted toward technical audiences — which distorts B2B software measurement more than consumer retail.

The individually reasonable responses to these gaps produced a collective problem. Platforms filled the holes with modelling. Google models conversions for users who declined consent, provided an account clears activation thresholds. Meta models conversions where events are missing and reports them alongside observed ones. Both report the combined figure as a conversion count in the interface, usually without separating modelled from observed unless an advertiser goes looking. The modelling is competent and the alternative — reporting only what survived the browser — would be worse. But an advertiser bidding toward a number that is 30% or 40% model output is bidding toward the platform’s belief about its business rather than its business.

Bias then enters from two directions at once. Platform reporting overstates contribution where it claims credit for demand it did not create, most visibly in branded search and retargeting. Platform reporting understates volume where events never arrive, most visibly in long consideration cycles, offline conversions, and phone-based sales. Some accounts suffer both simultaneously in different campaigns, which is why blended figures look plausible while the underlying allocation is wrong.

The damage is not confined to reporting. Broken conversion signal degrades the bidding itself, and that is the more expensive consequence. Smart Bidding and Advantage+ learn from the events they receive. Feed a system 60% of your conversions, systematically missing the ones that came from Safari users on mobile who bought three days later, and the model learns that those users are worth less. It will bid down on the pattern that produces them. The account then reports improving cost per acquisition while the business acquires fewer of its better customers — a failure mode that looks like success in every dashboard the team checks.

The industry’s own assessment reflects the situation. The IAB and BWG Strategy Global State of Data report for 2026 found roughly three-quarters of marketers saying their measurement systems lack the speed, accuracy, or trust they need. That is not a complaint about dashboards. It is an admission that the primary control system for the largest discretionary line in most marketing budgets is not trusted by the people operating it.

Privacy Sandbox died and the measurement problem stayed

For six years the industry organised its planning around a premise that turned out to be false. Google announced in 2019 that Chrome would remove third-party cookies and replace them with privacy-preserving APIs. Vendors built products against that roadmap. Conference agendas filled with cookieless readiness sessions. Deadlines slipped from 2022 to 2024 to 2025.

On 22 April 2025 Google confirmed it would not launch the standalone prompt it had proposed the previous July and would not deprecate third-party cookies in Chrome, leaving cookie controls where they already sat in Chrome’s privacy settings. The UK Competition and Markets Authority, which had extracted commitments from Google over the competitive effects of the change, noted the restated intention and began unwinding its oversight.

The second act arrived on 17 October 2025. Anthony Chavez, the Google vice-president responsible for Privacy Sandbox, published an update announcing the retirement of the Attribution Reporting API on Chrome and Android, IP Protection, On-Device Personalization, Private Aggregation including Shared Storage, the Protected Audience API on Chrome and Android, Protected App Signals, Related Website Sets, SelectURL, SDK Runtime, and Topics on Chrome and Android. The stated reasons were ecosystem feedback and low adoption. The same day, the CMA formally released Google from its Privacy Sandbox commitments. A small set of technologies survived — CHIPS for cookie partitioning, FedCM for federated sign-in, and Private State Tokens for fraud signals — none of which restores audience targeting or cross-site attribution.

Adoption data supports the low-uptake explanation. A longitudinal academic measurement of Privacy Sandbox deployment across the web found Protected Audience API usage falling from roughly 8% of measured websites in late June 2025 to close to zero by early August, before the retirement was announced, as early adopters stopped calling it. The replacement failed in the market before it failed in the roadmap. Chrome’s published schedule puts deprecation at milestone M144 in January 2026 and removal at M150 in July 2026.

Three consequences follow for anyone building measurement now.

First, there is no standard coming. The industry spent six years waiting for a shared replacement layer and will not get one. Whatever a business builds, it builds on its own first-party data, its own server infrastructure, and platform-specific conversion interfaces that each vendor controls and changes at will.

Second, the cookieless present arrived without the cookieless deadline. Chrome kept third-party cookies, but Safari, Firefox and Brave did not, and consent law applies regardless of what any browser permits. A business that postponed its measurement work because the Chrome deadline was cancelled has been losing signal on a fifth of its traffic the whole time and will keep losing it.

Third, the durable identity asset is authentication. Every product decision that raises the share of users who log in — accounts, loyalty programmes, saved carts, gated content, order tracking — produces measurement and compliance benefits that survive browser policy changes entirely. Logged-in users are the only audience an advertiser genuinely owns, and the only conversion records that reconcile cleanly across sessions and devices.

Consent became a first-class input to bidding

Consent used to be a legal chore handled by whoever owned the website. It is now a performance variable, and the mechanism is specific enough to be worth understanding line by line.

Google Consent Mode v2 requires a site to pass four signals — analytics storage, ad storage, ad user data, and ad personalisation — reflecting what the user agreed to. Google made v2 mandatory for advertisers serving the EEA and UK in March 2024, tied to Digital Markets Act obligations. Enforcement moved from documentation to automation on 21 July 2025: sites that fail to signal consent properly lose conversion tracking, remarketing, and demographic reporting for EEA and UK traffic. A misconfigured consent banner is now a direct cause of campaign underperformance in Europe, not a compliance risk sitting in a separate register.

The implementation choice matters more than most teams realise. Basic Consent Mode blocks Google tags entirely until consent is granted, which means a user who declines produces no data of any kind and no basis for modelling. Advanced Consent Mode loads tags in a restricted state that sends cookieless pings carrying the consent state, which gives Google enough aggregate information to model the conversions it cannot observe. Reported recovery from that modelling varies widely by account, and the figures circulating in vendor material — anywhere from 10% to 70% of lost conversions — should be read as a range shaped by traffic volume and consent rate rather than a promise.

Modelling also has activation thresholds, and accounts below them get nothing. Published guidance points to a requirement of roughly 700 ad clicks over seven days per country and domain pair, seven full days of collection, and a consent rate high enough for the model to calibrate. Small and mid-sized advertisers, and businesses split across many country domains, frequently sit under the threshold in several markets at once. They carry the full cost of consent refusal with none of the modelled recovery, which is one reason European mid-market accounts often report cost per acquisition well above what comparable US accounts report for the same product.

Consent rate therefore becomes a lever with direct revenue consequences. Banner design, wording, timing, the presence or absence of a reject button on the first layer, and the legal basis claimed for each purpose all move the acceptance rate, and the movement shows up in conversion volume, audience list size, and bidding quality. Raising acceptance from 35% to 55% does more for reported performance than most bid strategy changes, and it does so without buying a single additional impression. The consent management platform is measurement infrastructure that happens to have a legal function, and it should be owned accordingly.

There is a compliance boundary that responsible teams do not cross here. Interfaces engineered to make refusal difficult — pre-ticked boxes, hidden reject options, colour contrast designed to steer, cookie walls without a genuine alternative — attract enforcement across European data protection authorities, and the European Data Protection Board has taken a restrictive view of models where the only alternative to consenting is payment. The defensible route to a higher consent rate is a clearer explanation of what the user gets, fewer purposes claimed, and a banner that loads quickly and does not obstruct the page, not a dark pattern.

One technical dependency is easy to miss. For publishers and advertisers relying on the IAB Transparency and Consent Framework, TCF v2.3 replaced v2.2, with Google accepting the newer version from October 2025 and a migration deadline of 28 February 2026 for publishers and consent management vendors. Sites running an unmigrated platform can pass consent signals that the receiving systems no longer accept, producing a quiet failure that looks like a traffic quality problem.

The three layers of a working conversion data pipeline

The architecture that survives current conditions has three layers. Teams commonly build one, see partial improvement, and conclude the approach does not work. All three are needed because each solves a different loss.

Layer one is server-side collection. A server container — Google’s server-side Tag Manager is the common implementation, though the pattern is vendor-neutral — receives events from the browser at a first-party subdomain and forwards them onward. The gains are concrete: cookies set through HTTP response headers from the same domain are not subject to the JavaScript cookie lifetime caps that browsers impose, requests to a first-party endpoint are less frequently blocked, and page performance improves because third-party scripts move off the client. Consent state travels with each event, and the container decides which destinations may receive it. What server-side collection does not do is recover users who refused consent; that data is not missing because of technical friction, it is absent by instruction, and pretending otherwise is how compliance failures start.

Layer two is calibrated conversion interfaces. Each ad platform now offers a server-to-server route for conversion data: Meta’s Conversions API, Google’s enhanced conversions, the TikTok Events API, LinkedIn’s Conversions API and equivalents. These accept hashed identifiers — email, phone, name, address components — that let the platform match a conversion to a click or impression without a third-party cookie. Two implementation details decide whether the layer works. Deduplication requires a consistent event ID shared between browser and server so a single purchase is not counted twice. Match quality depends on how many identifier fields are supplied and how clean they are; an account passing hashed email only will match a materially lower share of conversions than one passing email, phone and postal fields. Meta exposes an event match quality score for exactly this reason, and treating that score as a monitored metric rather than a diagnostic curiosity is what separates functional setups from nominal ones.

Layer three is warehouse-first activation. The first two layers move events from a website to an ad platform. The third layer moves outcomes from the business systems where they actually live. A lead becomes an opportunity in the CRM three days later, closes at €4,200 six weeks after that, and churns or renews eleven months on. None of that exists in a browser. The pattern that works is a warehouse — BigQuery, Snowflake, Databricks or equivalent — holding joined data from the site, the CRM, the order system, and the ad platforms, with scheduled jobs pushing derived values back out through offline conversion imports, Customer Match audiences, and value adjustments. Google moved Customer Match uploads onto its current API path with a cutover on 1 April 2026, which broke pipelines still running against the legacy route; anyone maintaining these jobs learns quickly that they are production systems requiring monitoring, not scripts someone wrote once.

The sequencing question comes up constantly, and the answer is not the order above. The layer that returns value fastest is usually layer two, because conversion interfaces recover the largest volume of missing conversions per hour of engineering effort. A team with a functioning consent implementation and no server-side infrastructure should build Meta’s Conversions API and Google’s enhanced conversions first, then add server-side collection to consolidate and improve match quality, then build the warehouse joins that make value-based bidding possible. Teams that begin with the warehouse spend six months on data modelling before a single campaign behaves differently, and frequently lose executive patience before the payoff arrives.

Conversion values beat conversion counts once a bidder starts learning

The single change that produces the largest performance shift in most accounts is not a new campaign type. It is replacing a conversion count with a conversion value that reflects business reality.

Consider a retailer selling two product lines through one Performance Max campaign. Line A sells at €40 with a 65% gross margin, a 4% return rate, and a 38% repeat purchase rate within twelve months. Line B sells at €120 with an 18% gross margin, a 31% return rate, and almost no repeat purchase. A campaign counting conversions treats a sale of either as one conversion and bids equally for both. A campaign passing revenue as the value treats line B as three times more attractive. A campaign passing margin net of returns treats line A as roughly 30% more attractive than line B, and passing expected twelve-month contribution widens the gap to more than double. The three configurations produce three different businesses out of the same media budget, and only the third one reflects what the finance function cares about.

The value calculation is straightforward arithmetic that most teams never carry out. Take the order value, subtract cost of goods, subtract expected returns for that category and channel, subtract payment and fulfilment costs, then add a discounted estimate of repeat contribution derived from historical cohorts rather than from hope. The result is what the conversion is worth. Sending that number instead of the invoice total changes what the bidder buys within days.

Lead generation businesses face the same problem in a harsher form, because the gap between a form submission and revenue is wider and the distribution is more skewed. A B2B software company with a €900 average cost per closed deal might discover that leads from one campaign close at 11% and leads from another at 1.4%, while both report identical cost per lead. Bidding to cost per lead in that situation actively funds the worse campaign, because the worse campaign generates leads more cheaply — that is precisely why the leads are worse. The fix is to score leads on observable characteristics at submission, pass a value that reflects expected closed revenue, and import the actual outcome weeks later so the model learns the pattern rather than the proxy.

Platforms have built controls that assume this work has been done. Performance Max now carries customer lifecycle goals, including retention goals that allow higher bidding for lapsed customers, and reports customer acquisition cost in a dedicated column. Google previewed Product Value Adjustments at Google Marketing Live in May 2026, a pilot allowing percentage multipliers on conversion values for specific items inside Smart Bidding — a control that only means anything to an advertiser who knows which items deserve a multiplier. Every one of these features consumes private business intelligence as an input. An advertiser without it gets the automation without the advantage.

Three implementation cautions apply.

Value volatility breaks learning. A bidder receiving values that swing wildly for reasons unrelated to the user — currency conversion, a pricing test, a data pipeline error that occasionally sends cents as euros — will learn noise. Sanity checks on outbound values matter as much as the values themselves.

Delayed values need a bridge. If true value is only known ninety days after the click, a bidder receiving only the ninety-day figure learns too slowly to be useful. The workable pattern sends a predicted value immediately and corrects it later through value adjustments, so the model has something to learn from now and something accurate to learn from later.

Value-based bidding punishes bad data faster than count-based bidding does. A conversion count error inflates volume. A value error moves the entire spend distribution toward whichever segment happens to be mismeasured. Teams should verify the pipeline against finance figures before switching the bid strategy, not after.

Platform ROAS and incremental ROAS are two different products

Attribution answers which touchpoints preceded a conversion. Incrementality answers whether the conversion would have happened without the advertising. These are separate questions, they produce different numbers, and the difference is where most wasted budget hides.

The evidence on the size of that difference has firmed considerably. Analyses published across 2025 and 2026 put platform-reported return on ad spend somewhere between 20% and 60% above measured incremental lift, depending on channel and test design. Case studies published by the incrementality testing firm Haus, covering brands including Bombas, True Classic and Liquid Death, describe overstatement in the range of 1.5 to three times, widest on branded search and retargeting — the two places where the customer was most likely to convert anyway. Branded search in particular has produced incremental ROAS figures below 1.0 in published tests, meaning the campaign returns less than it costs once organic conversions that would have occurred regardless are removed from the credit.

Adoption has moved from specialist practice to mainstream expectation. A TransUnion survey conducted in July 2025 found 52% of US brand and agency marketers running incrementality tests. In retail media specifically, the Association of National Advertisers has reported that 71% of advertisers now rank incrementality as their most important key performance indicator. Incrementality has stopped being an advanced technique and become the standard by which a channel proves it earned its budget.

Measurement approaches and the question each one actually answers

Each row answers a narrower question than the row above it costs more to answer. A working measurement system runs several of them simultaneously and uses the expensive answers to correct the cheap ones.

Three test designs cover almost every practical case. Geo holdouts withhold or pause spend in a matched set of regions and compare outcomes against a synthetic control built from the remaining regions — Meta’s open-source GeoLift package and Google’s CausalImpact both implement this, and Google previewed Meridian GeoX at Google Marketing Live in May 2026 specifically to run publisher-agnostic geo experiments and convert results into model priors. In-platform conversion lift tests use ghost bids to hold out a randomised user group and are quick to run, though their scope stops at conversions the platform itself measures. Public service announcement holdouts serve a neutral creative to the control group, which controls for exposure but costs real impressions.

One methodological caution deserves attention because it undermines the cleanest-looking design. Randomised tests inside ad platforms are vulnerable to divergent delivery: the algorithm assigns different creatives or arms to systematically different audiences, so the comparison is no longer between identical groups. Research on this effect has been accumulating, and it means an in-platform lift result can carry a bias in the same direction as the platform’s commercial interest. Geo designs are noisier but harder to corrupt, which is why teams with real money at stake usually run both and treat agreement between them as the standard of proof.

Marketing mix modelling returned as a governance tool, not a report

Marketing mix modelling spent a decade looking like a relic. It was built for television, ran on aggregate weekly data, cost six figures a year from a consultancy, arrived as a slide deck two months after the period it described, and could not say anything useful about a search campaign. Digital attribution replaced it because attribution was faster, cheaper, and granular.

Signal loss reversed the comparison. Attribution depends on observing individual journeys, which is exactly what privacy changes removed. Modelling depends on aggregate spend and outcome series, which survived intact. A method that never needed user-level identifiers turned out to be the one method unaffected by their disappearance.

Google’s Meridian accelerated the return by removing the cost barrier. Released as an open-source Bayesian model in February 2025 alongside a partner programme of certified measurement partners, it addresses the specific weaknesses that made older models useless for performance media: it models reach and frequency rather than spend alone, it accepts results from incrementality experiments as priors that calibrate the model instead of letting regression guess from correlation, and it can use Google query volume as a control variable to separate advertising effect from underlying demand shifts. Google added a no-code Scenario Planner in February 2026 for teams without Python capability. Meta’s Robyn occupies similar ground with a different statistical approach.

The requirements are where enthusiasm meets reality. A useful model needs two to three years of clean channel-level spend and outcome data, weekly granularity at minimum, someone who can specify priors and read diagnostics without misinterpreting them, and enough variation in the historical data to identify effects at all. A business that has run the same budget split across the same three channels for two years has no variation for the model to learn from, and will get a model that confidently reproduces its own assumptions. Teams also discover that the Python package is published as google-meridian rather than meridian, that a GPU makes the difference between hours and days of fitting time, and that the hard part is neither installation nor interpretation but the data preparation that precedes both.

Adoption figures suggest wide use and shallow application. A survey by EMARKETER and Snap in July 2024 found 53.5% of US marketers using marketing mix modelling. Research Google commissioned from Harvard Business Review Analytic Services, fielded to 547 marketing professionals familiar with modelling between late September and early October 2025, was framed explicitly around an actionability gap — the distance between having a model and changing a budget because of it. That framing matches what practitioners report: most organisations that run a model do not reallocate money on the strength of it.

The right role for modelling is governance rather than fine-tuning. It answers portfolio questions — whether the total marketing budget is roughly right, how much of current sales is baseline demand that would exist with no advertising, whether one channel is saturated, what diminishing returns look like at higher spend. It cannot answer campaign questions, because aggregate data cannot resolve them. Any vendor claiming that a model will tell a team which ad set to pause is selling something the method cannot deliver.

The productive arrangement is a three-part loop with clear responsibilities. Modelling sets the portfolio view and flags channels whose estimated contribution has wide uncertainty. Experiments test those specific channels and return causal estimates. Those estimates become priors that update the model. Attribution then handles day-to-day decisions inside channels the other two methods have already validated. Each method covers the others’ blind spots, and the loop is what turns three disagreeing dashboards into one defensible allocation.

Geo experiments, holdouts, and the cost of never testing

A geo experiment is the cheapest way to buy certainty about a channel, and most organisations refuse to buy it because the price is visible while the benefit is not.

The design is simple in outline. Split comparable geographic markets into treatment and control. Change spend in treatment — increase it, cut it, or stop it entirely — while holding control constant. Compare the outcome difference against what the pre-period relationship predicted. The difference is causal lift, because nothing else changed between the groups except the advertising.

The execution details determine whether the result means anything.

Market matching comes first. Treatment and control regions need similar baseline sales, similar seasonality, similar channel mix, and enough historical correlation that a synthetic control built from the control markets predicts the treatment markets accurately during a pre-period where nothing was changed. A test whose pre-period fit is poor cannot produce a trustworthy post-period estimate, and this is the diagnostic most teams skip.

Power comes second, and it is the most common reason tests fail. A test can only detect an effect larger than its minimum detectable effect, which depends on outcome variance, the number of markets, the size of the spend change, and the test duration. Small businesses in small countries frequently cannot run a valid geo test on a national scale at all — there are not enough independent regions with enough volume. Running an underpowered test and reading its noisy output as a finding is worse than not testing, because it produces a confident wrong answer. Four to eight weeks is the usual window: long enough to capture delayed conversions, short enough that market conditions do not drift.

Contamination is the third issue. Spillover across regional boundaries, national television running simultaneously, retargeting audiences built before the test that continue serving in control markets, and shipping patterns that route online orders across regional lines all blur the split. Every one of these needs to be handled in the design or acknowledged in the interpretation.

The cost is real and worth stating plainly. A holdout means withholding advertising from customers who would have converted, which means foregone revenue for the duration. That is the price of the information. The relevant comparison is not against zero but against the alternative — continuing to spend against platform-reported numbers that published evidence suggests overstate contribution by 20% to 60%. A business spending €200,000 a month on paid media that runs no experiments is making a €2.4 million annual allocation decision on the basis of reports written by the parties receiving the money.

Published results illustrate why the information is worth buying. Test outcomes documented by incrementality vendors include campaigns that produced no measurable lift at all — where pausing spend in treatment markets changed nothing, and the budget was redirected to a channel that did produce lift. Those results are only visible through experimentation. No attribution model, however sophisticated, will report a zero for a campaign that has conversions flowing through it, because the conversions genuinely happened; they just did not happen because of the ads.

The practical cadence for a mid-sized advertiser is one properly powered test per quarter on the channel where uncertainty is highest and spend is largest, with an annual budget line set aside for it. Treating experimentation as a recurring cost of doing business, rather than a project that needs individual approval each time, is the organisational change that makes the method stick. Teams that require a business case for every test run roughly one per year, learn very little, and spend the intervening months arguing about dashboards.

AI search changed the shape of demand before it changed the budget

The clearest public evidence on how AI answers change search behaviour comes from Pew Research Center, which published findings in July 2025 based on the actual browsing activity of 900 US adults who agreed to share it. The dataset covered 68,879 Google searches during March 2025, of which 12,593 produced an AI-generated summary.

The results are specific enough to plan against. Around 18% of searches in the sample produced an AI summary, and 58% of participants encountered at least one. When a summary appeared, users clicked a link 8% of the time; when it did not, they clicked 15% of the time. Links inside the summary itself were clicked on 1% of visits. Sessions ended entirely after 26% of pages carrying a summary, against 16% without one. The typical summary ran 67 words and cited three or more sources in 88% of cases. Roughly two-thirds of all searches in the sample ended with no click on anything.

Independent measurements point the same way with different magnitudes. Ahrefs reported in May 2025 that the presence of AI Overviews was associated with organic clicks falling by about 34.5%. Seer Interactive published an analysis in September 2025 covering 3,119 informational queries across 42 organisations, 25.1 million organic impressions and 1.1 million paid impressions, describing continuous compression of click-through rates over fifteen months rather than a single step change. The estimates differ because the query mixes and methodologies differ; the direction does not.

Google’s own reporting is consistent with this without saying so directly. Search and other revenue grew 19% in the first quarter of 2026 while Network revenue — ads placed on third-party publisher sites — fell 4%. Management stated that queries were at an all-time high and credited AI Overviews and AI Mode with driving usage growth. More searches, answered inside Google, producing fewer exits to the open web. Total demand for information rose while the traffic that carries that demand to independent websites fell, which is a structural change in the shape of the funnel rather than a fluctuation in it.

For marketing strategy the consequence is precise and often misread. The visits that still arrive are a smaller, differently-composed set. Queries that a summary can resolve — definitions, comparisons of published specifications, opening hours, simple how-to — no longer produce visits. Queries that require a decision the model cannot make, a price only the seller knows, a stock check, a configuration, or a human conversation still produce visits. The traffic that remains is weighted toward the bottom of the funnel by subtraction rather than by any improvement in the site.

This is what makes reach-based reporting actively misleading in 2026. A site whose organic sessions fell 30% while revenue held flat has not declined; it has lost the sessions that never converted. A site whose sessions held flat while revenue fell 30% has a serious problem that a traffic report will completely conceal. Sessions and impressions have decoupled from outcomes to the point where any report built on them describes a different business than the one the finance team sees.

The measurement work this creates is unglamorous and specific. Traffic needs segmenting by whether an AI surface was involved, which requires custom channel grouping because default analytics configurations misclassify a large share of AI referrals. Conversion rate needs tracking per segment rather than in aggregate, because a blended figure will move for reasons of traffic composition rather than site performance. Content needs evaluating by conversion contribution rather than by sessions, which will change which pages a team considers worth maintaining. None of this is possible without conversion measurement that works at segment level, which returns to the same dependency: the businesses that can read the AI search transition correctly are the ones that already fixed their conversion data.

AI-referred visitors behave differently, and the evidence is messier than the headlines

The commercial value of AI-referred traffic has moved from an open question to a documented pattern, though the published numbers vary enough that anyone quoting a single figure is oversimplifying.

Adobe Analytics has the largest continuous dataset, drawn from more than a trillion visits across over 130 large North American retailers and tracked since October 2024. Its first quarter 2026 analysis reported that AI-referred traffic converted 42% better than non-AI traffic in March 2026 — a reversal from converting 38% worse a year earlier. Its May 2026 data put AI-referred retail visitors at 54% better than non-AI traffic, with retail traffic from AI sources up 138% year over year. Engagement metrics moved consistently: AI-referred shoppers spent 48% more time on product pages, viewed 13% more pages per visit, and showed a 12% higher engagement rate in the March retail data.

Shopify’s separately sourced commerce data corroborates the direction from a different sample. It reported in May 2026 that AI-referred sessions convert at nearly 50% higher rates than organic search on product detail pages, that AI outperformed organic search in 23 of 25 merchant categories by an average of 56% within those categories, and that AI-referred orders carried 14% higher average order values. AI referral sessions grew more than eightfold and AI-referred orders nearly thirteenfold between January 2025 and March 2026. More than half of AI-referred sessions began on a product detail page, against roughly 20% for organic search — the behavioural fact that explains the conversion difference better than any theory about intent.

Three caveats belong with these numbers, and leaving them out is how the figures get misused.

Adobe’s comparison is against non-AI traffic in aggregate, which blends paid search, email, affiliate and organic. It is not a head-to-head against Google organic, so describing AI traffic as converting better than Google is shorthand rather than a precise claim. Both datasets are vendor-published and not independently audited, which is why the agreement between two independent samples matters more than either figure alone.

Sector variation is large enough to invert the conclusion. Adobe’s May 2026 travel data showed AI referrals to US travel sites up 194% year over year and up 2,215% since October 2024, with AI-referred travellers 21% more engaged, spending 70% longer per visit and bouncing 41% less — while still converting 28% less than non-AI traffic. The gap had narrowed by roughly 70% since October 2024, but it had not closed. A retailer and a travel operator reading the same aggregate headline about AI traffic quality would both be wrong about their own situation.

The per-engine conversion figures circulating widely — ChatGPT at 15.9%, Perplexity at 10.5%, Claude at 5%, Gemini at 3% against Google organic at 1.76% — come from aggregations of case studies rather than from a controlled cross-industry sample, and at least one of them traces to a single B2B client engagement. They are useful as an indication that engines differ. They are not benchmarks, and a business planning against them is planning against someone else’s account.

Attribution is the practical obstacle to using any of this. AI assistant referrals frequently arrive without a referrer or with one that default analytics configurations file as direct traffic, and published estimates of the misattribution share run high enough that most teams have been undercounting the channel materially. Google Analytics reportedly added a native AI assistant channel grouping in May 2026, which helps prospectively but does nothing for historical data. Until a team builds a custom channel group with pattern matching against the known assistant domains, its AI traffic sits inside direct, its conversion rate for direct looks unaccountably strong, and the channel that may be its highest-converting source remains invisible in every report it produces.

Attribution inside interfaces that hide the referrer

Assistant interfaces break attribution in a way that is different from cookie loss, and the difference matters because the fixes are different too.

Cookie restrictions broke the link between a click and a later conversion. Assistant interfaces break something earlier: the record that an interaction happened at all. A user asks a model to compare three options, receives an answer that mentions a brand, and arrives at the site an hour later by typing the name into a browser bar. There was an exposure, it was influential, and no technical artefact of it exists anywhere in the advertiser’s data. Nothing in the analytics stack will ever attribute that visit correctly, because the information required to do so was never transmitted.

Four partial substitutes are in practical use, and none of them is a replacement.

Server log analysis reveals which AI crawlers visit which pages and how often. Requests from the retrieval agents operated by the major model providers are identifiable in logs, and correlating crawl frequency against later citation appearance gives a rough read on which content is being ingested. This measures supply rather than demand — it says nothing about whether an answer citing the page led anywhere.

Citation monitoring runs a fixed set of commercially relevant prompts against the main assistants on a schedule and records whether the brand appears, in what position, and with what framing. The output is a visibility series comparable to rank tracking. Its weakness is that model outputs vary between runs and between users, so a single check is noise and only a consistent series carries information.

Self-reported attribution asks the customer. A single optional field at checkout or on a form — where did you first hear about us — produces data that no tracking system can generate, and it captures word of mouth, podcasts and assistant recommendations that are otherwise invisible. Response bias is real, recency bias is severe, and free-text answers need categorising. It is still the only method that observes the channels the technical stack cannot see, and the businesses that run it consistently find that its picture of demand differs sharply from their platform reports.

Branded demand tracking treats the volume of brand-name searches and direct traffic as a proxy for upper-funnel influence. When assistant answers mention a brand more often, branded search volume tends to rise. Movements are slow, contaminated by every other marketing activity, and useful only over months.

The structural conclusion follows from the failure of all four to be sufficient. When exposure cannot be observed, causation must be inferred from experiments rather than reconstructed from records. This is the same conclusion the incrementality section reached from a different direction, and the convergence is not coincidental. A measurement regime built on observation degrades as observation degrades. A measurement regime built on controlled comparison does not care whether the exposure was recorded, because it compares outcomes between groups rather than tracing paths between events.

Last-click attribution is not merely inaccurate under these conditions; it is structurally incapable of describing what happened. A model that assigns 100% of credit to the final touchpoint, in an environment where the influential touchpoint is systematically invisible and the final touchpoint is a brand-name search the customer performed because of it, will report that branded search generates the business. Every euro reallocated on that basis moves money from what creates demand to what harvests it, and the reports will look better each time.

Agentic commerce puts a machine between the shopper and the checkout

Three overlapping protocol efforts have emerged to let software agents browse, build carts and pay on a person’s behalf, and they sit at different layers rather than competing for the same one.

The Universal Commerce Protocol was announced on 11 January 2026 at the National Retail Federation conference in New York, developed by Google with Shopify and co-developers including Walmart, Target, Etsy and Wayfair, with more than twenty companies endorsing it at launch. It covers the full commerce journey — discovery, cart, checkout and post-purchase — and underpins commerce experiences in Google’s AI surfaces. Shopify opened UCP agent registration to any developer through self-serve access from 17 June 2026.

The Agentic Commerce Protocol, built by OpenAI with Stripe, handles the narrower problem of checkout execution inside a chat interface: how an agent collects a payment selection, passes a scoped token to the merchant, and lets the merchant charge it through a compliant payment provider while remaining the merchant of record. PayPal joined as a payment provider in October 2025, Stripe shipped a supporting suite in December 2025, and the specification is maintained publicly.

The Agent Payments Protocol, initiated by Google, addresses authorisation and trust — proving that a human actually authorised a given transaction. Google donated it to the FIDO Alliance on 28 April 2026 alongside a version 0.2 release, with contributions from around sixty organisations, and donated an accompanying verifiable intent system developed with Mastercard. Visa shipped a protocol-agnostic on-ramp in April 2026 accepting agents across multiple standards simultaneously. Anthropic’s Model Context Protocol, now governed by the Linux Foundation, provides the general connective layer between models and external systems including commerce APIs.

Early consumer adoption has been thinner than the announcement volume suggests. Reports indicate OpenAI wound down its in-chat Instant Checkout surface in March 2026 after a limited period with few merchants shipping against it, while the underlying protocol continued. The infrastructure is being built faster than the demand for it has been demonstrated, which is a reason for measured investment rather than either dismissal or urgency.

The measurement implications are more immediate than the revenue implications, and they are severe.

An agent-mediated purchase has no session in the conventional sense. There is no landing page, no scroll depth, no on-site search, no cart abandonment sequence, and no surface on which to run an experiment. The conversion happens inside an interface the merchant does not control and cannot instrument. Everything a merchant has built over fifteen years to influence conversion on its own site becomes unavailable at exactly the moment the purchase decision is made.

What replaces it is the product feed and the structured data behind it. When an agent compares options, it reads titles, attributes, specifications, availability, pricing, shipping terms and return policies. Feed completeness and accuracy become conversion levers in the literal sense — the agent’s recommendation is the conversion event, and it is determined by data quality rather than by persuasion. Google’s product discovery signals announced at Google Marketing Live in May 2026 point in the same direction: additional product attributes exist to make items legible to conversational surfaces.

The practical instrumentation task for now is narrow and worth doing early. Identify agent traffic in server logs and separate it from human traffic so it does not corrupt conversion rate calculations. Track orders that originate from agent-mediated channels as a distinct segment with their own margin and return profile. Audit the feed against what an agent would need to make a recommendation, including the fields most merchants leave empty. None of this requires a bet on which protocol wins, which is the point — the feed and the log analysis pay off regardless of the outcome.

Retail media grew because it sits closest to the transaction

Retail media became the fastest-growing category in advertising for one reason that gets buried under discussion of inventory and formats: the seller can observe the purchase. When a shopper sees a sponsored product on a retailer’s site and buys it there, the retailer holds both halves of the record. No cookie, consent flag or attribution model is required.

The growth figures are large under every methodology, though the definitions vary enough that cross-source comparison is unreliable. EMARKETER put US retail media spending at $58.79 billion in 2025 rising to roughly $69.33 billion in 2026, growth close to 18%. WARC and WPP Media estimated worldwide retail media at $174.9 billion in 2025, up 13.7%, and forecast $196.7 billion in 2026, up 12.4%, projecting the category at around 16% of all global advertising spend and overtaking linear and connected television combined during the year. IAB and PwC reported US retail media network revenue rising 23% in 2024 to $53.7 billion. EMARKETER’s longer view has US commerce media reaching $142.07 billion and 23.9% of all US digital advertising by 2030. Definitional differences between forecasters are large enough that any figure needs its source attached, and comparisons between reports from different houses tell you more about their taxonomies than about the market.

Concentration is the part most allocation discussions skip. Roughly 277 retail media networks were operating worldwide as of November 2025, yet EMARKETER projects Amazon and Walmart capturing around 89% of incremental US retail media spending in 2026 — approximately $9.42 billion of $10.53 billion in net-new investment. Amazon generated about $68 billion in advertising revenue in 2025, roughly 8% of its US gross merchandise value, while Walmart’s global advertising revenue reached $6.4 billion. For most brands the practical question is not how to allocate across 277 networks but how much goes to two of them and how much stays in channels the brand controls.

The closed-loop measurement that drives the category is also its central problem, and the market has noticed. A retailer selling advertising and reporting the resulting sales occupies the same position as a platform reporting its own contribution — the seller measures the outcome. Sponsored product placements sit directly next to purchase intent, which is exactly where incremental lift is lowest, because the shopper searching for a product category on a retailer’s site was frequently going to buy something in that category regardless. That is why the Association of National Advertisers found 71% of retail media advertisers ranking incrementality as their most important metric. The channel that solved the measurement problem created a new one: near-perfect observation of conversions that would have happened anyway.

Off-site retail media compounds the difficulty. Networks now activate retailer purchase data against inventory on social platforms and connected television, where the closed loop breaks and attribution reverts to modelling — but the reporting frequently retains the confident presentation of on-site results. An advertiser reading a single dashboard covering on-site and off-site activity is reading two measurement regimes with very different reliability presented in the same font.

For a brand building conversion intelligence, retail media requires the same treatment as any other channel with an interested measurer. Request incremental lift studies rather than attributed sales. Run holdouts where the network’s cooperation permits it, and treat refusal as information. Reconcile network-reported sales against internal shipment or sell-through data, which will frequently disagree. And separate the trade budget conversation from the media budget conversation, because retail media negotiations are routinely bundled with commercial terms in a way that makes the media performance question secondary to the shelf-space question.

The conversion definition problem most teams never fix

Before any of the sophisticated work matters, most accounts have a definitional problem that quietly invalidates everything downstream. It is unglamorous, it takes a day to audit, and skipping it is the most common reason measurement projects fail to produce results.

The pattern repeats across accounts of every size. Over three years, twelve conversion actions accumulated: purchase, add to cart, newsletter signup, contact form, phone click, PDF download, video view, two duplicate purchase tags from a platform migration, a chat widget open, an outbound link click, and something a former agency created called engaged visit. Nine are marked as primary. The bidding system therefore treats a PDF download as equivalent to a €400 order, and allocates budget toward whichever action is cheapest to generate. Cost per conversion looks excellent. Revenue does not move.

The corrections are specific.

One primary conversion per bidding goal, with everything else recorded as secondary for observation only. Machine bidding cannot serve two masters, and a mixed primary set produces a weighted average of goals that nobody chose.

Deduplicate ruthlessly. Duplicate tags from migrations and plugin conflicts are extremely common, and a duplicated purchase event halves apparent cost per acquisition while doubling apparent revenue. Verifying total platform-reported conversions against finance figures for the same period catches this in an afternoon.

Choose counting rules deliberately. Counting every conversion suits ecommerce, where each order is separate revenue. Counting one per click suits lead generation, where the same person submitting three forms is one opportunity, not three. Getting this backwards inflates lead volume and teaches the bidder to buy people who fill in forms repeatedly.

Set attribution windows to match the actual sales cycle, then hold them still. A thirty-day click window on a product with a ninety-day consideration period discards real conversions. Changing the window mid-quarter makes period comparison meaningless, and someone always changes it without telling anyone.

Reconcile analytics and advertising definitions. Google Analytics key events and Google Ads conversions are counted, attributed and windowed differently, and the two will never match exactly. Teams that expect them to match spend weeks chasing a discrepancy that is a property of the systems, while teams that ignore the comparison entirely miss the discrepancies that do indicate breakage. The useful practice is a documented tolerance and an alert when it is exceeded.

Filter invalid submissions. Bot form fills, competitor testing and spam inflate lead counts and corrupt bidding. Every filtered submission removed from the conversion feed makes the model slightly more accurate about what a real lead looks like.

The organisational fix matters more than any of the individual corrections. Conversion definitions need an owner, a written document, and a change process, because they drift. A developer adds a tag during a site release. An agency creates a conversion action to demonstrate progress. A platform introduces a default that switches on automatically. Six months later the account is bidding toward something nobody chose, and the only way anyone finds out is a quarterly audit that compares what is configured against what the business decided it wanted.

Benchmarks mislead when the denominator moves

Conversion rate benchmarks circulate constantly and are used to justify budgets, fire agencies and set targets. Most of the disagreement between published figures is not disagreement at all — it is different denominators being compared as though they were the same measure.

Published ecommerce conversion rate benchmarks and what each one measures

The spread between these figures reflects sample composition and measurement definition rather than a genuine disagreement about how ecommerce performs, which is why a store comparing itself against the highest number will conclude it is failing when it may be typical.

Three distinctions do most of the damage when they are ignored.

Sessions against visitors produces different numbers for the same store, because one person browsing three times counts once in a visitor denominator and three times in a session denominator. Shopify’s published industry table uses visitors to orders while most platform reports use sessions to orders, which makes them structurally incomparable. A team that switches analytics tools and sees its conversion rate move has usually changed the denominator, not the performance.

Vertical spread is wider than any improvement a team can make to its own site. Food and beverage stores commonly report rates between 4.9% and 6.2% while luxury and jewellery sits between 0.8% and 1.5% — a difference of six times or more inside the same platform, driven by price point and purchase frequency rather than by site quality. A jewellery retailer converting at 1.1% and a snack brand converting at 5.8% may be equally well run, and comparing either against a blended global average produces a conclusion with no information in it.

Traffic mix moves the number more than most site changes do. Mobile carries somewhere between three-quarters and four-fifths of sessions in most consumer categories while converting at roughly half the desktop rate, so a shift in device mix alters the blended conversion rate with nothing on the site having changed. Returning customers convert at multiples of first-time visitors. A retargeting-heavy month reports a strong conversion rate. A prospecting-heavy month reports a weak one. The same site, the same product, different arithmetic.

The practical response is to abandon the sitewide conversion rate as a management metric. It is a composite of traffic composition, device mix, customer mix, category mix and site performance, and it moves for reasons that have nothing to do with the thing anyone is trying to manage. Segmented conversion rates — by channel, device, customer type and category, each tracked against its own history rather than against an industry average — are the only version of the metric that supports a decision.

Funnel-stage measurement adds the diagnostic layer. A store with a healthy add-to-cart rate and weak checkout completion has a payment, shipping or trust problem. A store with weak add-to-cart and healthy checkout completion has a product page or traffic quality problem. The sitewide rate is identical in both cases and points to neither.

Ecommerce economics under double-digit auction inflation

Auction prices rising 12% to 14% in a year does not damage an ecommerce business in proportion. It damages it in proportion to how thin the contribution margin already was, which is why identical percentage increases produce a mild irritation for one retailer and an existential problem for another.

The arithmetic is worth working through explicitly. A store with a €70 average order value and a 45% contribution margin after cost of goods, payment fees, fulfilment and returns has €31.50 of contribution per order to spend on acquisition and profit. At a €25 cost per acquisition it keeps €6.50 per new customer before overheads. A 14% increase in media cost moves acquisition to €28.50 and leaves €3.00. The media cost rose 14% and first-order profit fell by 54%. The same increase applied to a store with a 65% contribution margin reduces first-order profit by roughly 12%. Neither store changed anything it did; the sensitivity was set by margin structure before the auction moved.

Three responses exist and only one of them is media buying.

Raise the value of a conversion. Average order value responds to bundling, tiered shipping thresholds, subscription options and merchandising, and every euro added to contribution per order absorbs auction inflation directly. This is the fastest lever available and it sits with merchandising rather than with the media team, which is why it is frequently the last one anyone tries.

Raise the conversion rate on traffic already being paid for. A store converting at 1.8% that reaches 2.2% has cut its real cost per acquisition by 18% without touching a bid. The relevant work is unglamorous — page speed, express payment options, clearer shipping and returns information, product page completeness, reducing form fields — and its return is measurable against a control group if the team is willing to run the test properly.

Reallocate against measured contribution rather than reported ROAS. This is where conversion intelligence pays for itself. A store that discovers its branded search campaign and its lower-funnel retargeting produce little incremental lift can move that budget into prospecting or into retention channels and hold total contribution flat while reported ROAS falls. The reported number getting worse while the business gets better is the normal signature of a correct reallocation, and organisations that cannot tolerate it will not make one.

Blended measurement provides the discipline that campaign-level reporting cannot. Total media spend divided into total revenue — the blended ratio some businesses label MER — is immune to attribution disputes because it contains no attribution. It cannot say which channel worked, but it can say whether the whole portfolio is working, and it moves in the same direction as the bank account. Teams running incrementality tests alongside a blended target have a working control system: the blended figure catches drift, the tests explain it.

Payback period deserves a place next to the ratio. A business acquiring customers at negative first-order margin on the strength of a projected twelve-month payback is making a financing decision, not a marketing one, and it needs cohort data showing that historical customers actually repeated at the assumed rate. Cohorts acquired during a discount-heavy period repeat at materially lower rates than cohorts acquired at full price, so a lifetime value assumption drawn from one and applied to the other overstates payback systematically. Auction inflation makes this failure mode more expensive every quarter, because the acquisition cost being financed keeps rising while the repeat behaviour being counted on does not improve.

B2B pipelines need conversion intelligence more than ecommerce does

B2B marketing has the worst version of every measurement problem discussed so far, and the smallest tolerance for getting it wrong.

The structural difficulties compound. Sales cycles run months, so the outcome that matters is separated from the click by a period longer than any attribution window. Conversion volumes are low, so campaigns frequently sit below the threshold at which machine bidding can learn anything, and statistical significance on any test takes quarters rather than weeks. Buying committees mean multiple people from one account research independently, generating several apparently separate conversions that represent one opportunity. Deal values vary by an order of magnitude, so counting leads equally is not an approximation but a category error. And technical audiences block tracking and refuse consent at higher rates than consumer audiences, which skews the observable sample toward less technical buyers.

The failure mode this produces is specific and widespread: a marketing function that reports cost per lead against a target, hits the target, and does not generate pipeline. Cost per lead falls when lead quality falls, because low-intent people are cheaper to reach and more willing to exchange an email address for a whitepaper. Bidding to cost per lead is therefore a mechanism for buying worse leads more cheaply, dressed as improvement.

The corrections are all versions of moving the measured outcome closer to revenue.

Import CRM outcomes back into the ad platforms. Offline conversion import exists for this purpose: capture the click identifier at form submission, store it against the CRM record, and upload the outcome when the record reaches a stage worth bidding toward — qualified opportunity, proposal sent, closed won. Bidding toward qualified opportunities rather than form fills changes which campaigns receive budget within a few weeks, and the change is usually large.

Assign stage values that reflect probability. A qualified opportunity worth €40,000 with a 25% historical close rate carries an expected value of €10,000. Passing that number rather than the deal size gives the bidder a signal that is both informative and available months earlier than the closed figure.

Use lead scoring at submission as a bridge for low-volume accounts. When closed-won volume is too low to train anything, an interim score based on firmographic and behavioural attributes observable at submission provides a proxy that correlates with eventual quality. It is inferior to real outcomes and vastly superior to a raw count.

Build the sales feedback loop as a process rather than a favour. Marketing needs disposition reasons on rejected leads, and it needs them consistently enough to aggregate. Most organisations attempt this, most abandon it within two quarters, and the ones that maintain it can identify which campaigns, geographies and offers produce leads sales will not work — which is the highest-return information a B2B marketing team can hold.

Platform capability is moving in this direction. Google’s 2026 announcements included bidding that accounts for the whole lead journey and call analysis using Gemini to identify strong leads and surface what drives sales. Both features consume outcome data the advertiser must supply. The platform is offering to act on pipeline intelligence, which means an organisation that cannot produce pipeline intelligence has been given a control it cannot use.

Two further points specific to B2B. Self-reported attribution carries more weight here than in consumer categories, because so much B2B influence happens in communities, podcasts, peer conversations and assistant answers that leave no technical trace — a single question on the demo form frequently reveals a channel that no analytics configuration attributes at all. And branded search deserves particular scepticism: in a market where a competitor’s sales team may be the reason someone searched a brand name, the incremental value of bidding on that name is an empirical question with a testable answer, and the answer in published tests is often lower than the reported ROAS suggests.

Local service businesses and the call that never gets scored

Local service businesses — trades, clinics, legal practices, repair, installation, home services — operate under conditions where the standard measurement playbook fails for reasons of volume and channel rather than sophistication.

Most conversions are phone calls, and most phone calls are never scored. A plumbing business receiving forty calls a week from paid search knows the call count from its call tracking. It usually does not know that eleven were outside the service area, six were price shoppers who booked nobody, four were existing customers calling about a previous job, two were suppliers, and seventeen produced work worth between €90 and €4,200. The bidding system receives forty conversions of equal value. The campaign that generates the most out-of-area price shoppers looks like the best performer, and receives more budget accordingly.

The fix is call outcome scoring, and it is a business process rather than a technology purchase. Someone tags each call — booked, not booked, out of area, wrong number, existing customer, quoted and lost — and the tags flow back as conversion events with values. Modern call systems and recent platform features assist with classification, including Gemini-based call analysis in Google Ads that identifies strong leads and time-wasters. The classification still needs a definition of what counts as a good call, which only the business can supply, and a consistent habit of applying it.

Job value variance makes value-based bidding unusually productive here. A boiler repair worth €140 and a full system installation worth €6,000 arrive through the same phone number from the same campaign. Passing the eventual job value, or a tiered estimate available at booking, redirects budget toward the queries that produce large jobs. For many local businesses this single change moves revenue more than any other available action, because the query patterns that produce large jobs are distinct and the platform can learn them once it knows which ones they are.

Low volume forces methodological adjustments that larger advertisers do not need. Weekly reporting is noise at forty conversions a week; monthly and quarterly comparison is the honest cadence. Geo experiments are usually impossible because there are not enough independent markets, which leaves before-and-after comparison with seasonal adjustment as the practical substitute — weaker evidence, honestly labelled. Value tiers of three or four bands hold up better than precise values, because precise values on small samples produce volatile bidding.

Two structural factors deserve attention. Google’s Local Services Ads operate on a separate lead-based model with its own dispute process, and the leads it produces need the same outcome scoring as everything else — the difference being that a business can contest leads it considers invalid, which only works if it is tracking them. And review volume and rating function as conversion rate variables in local categories at a magnitude that rivals anything on the website, because the comparison a prospective customer runs happens in a map interface or an assistant answer rather than on the business’s own pages.

The measurement discipline that pays off for a local operator is narrower than for an ecommerce brand but no less demanding: know which calls became work, know what the work was worth, feed both back, and resist the temptation to read weekly fluctuations as trends.

Subscription and app businesses live on the value curve, not the first order

For a subscription or app business the first transaction carries almost no information about whether the acquisition was profitable. A trial start, a first month at a promotional price, or an install tells you nothing about the twelve-month contribution that determines whether the media spend was worth making. Every measurement decision follows from that gap.

The consequence is that these businesses cannot use the conversion event their platforms most readily accept. A trial start is easy to instrument and easy to bid toward, and bidding toward it selects for people who start trials — a population that overlaps imperfectly with people who pay for a second year. Accounts that switch bidding from trial start to first payment usually see reported volume fall and revenue rise, which is the signature of a measurement correction rather than a performance decline.

Predicted lifetime value is the intended answer, and it is harder than vendors suggest. A usable prediction needs cohort history long enough to observe the behaviour being predicted, which means a business less than two years old is extrapolating. It needs segmentation, because the retention curve for customers acquired through branded search differs from those acquired through paid social, and applying a blended curve to both misprices each. And it needs recalibration, because product changes, pricing changes and competitive shifts move retention in ways that invalidate the model that was accurate last year.

The pattern that works in practice is a short-horizon proxy validated against long-horizon outcomes. Rather than predicting twelve-month value at signup, identify the earliest observable behaviour that correlates strongly with twelve-month retention — activation of a specific feature within seven days, a second session within seventy-two hours, an invited colleague, a completed profile — and bid toward that. A seven-day signal that predicts annual retention with reasonable accuracy is more useful to a bidding system than a perfect annual figure that arrives too late to influence anything.

Mobile app measurement operates under tighter constraints than web. Apple’s framework returns aggregated, delayed and coarsened postbacks with limited conversion value encoding, which means fine-grained value passing is not available on iOS in the way it is on the web. The workable approach uses the available value encoding to represent a small number of value tiers rather than continuous values, accepts that the signal is delayed, and leans harder on incrementality testing and modelling because user-level attribution simply is not there to be recovered.

Cohort discipline is where most of these businesses lose money invisibly. A cohort acquired during a 50% first-year discount campaign retains at a materially lower rate than one acquired at list price, because a share of it was buying the discount rather than the product. Averaging both into one lifetime value figure and using it to justify higher acquisition bids on the discounted offer produces a spend increase justified by a number the offer itself degraded. Every lifetime value assumption needs to be segmented by the offer that produced the cohort, and needs re-verifying each time the offer changes.

Retention goals in Performance Max and the equivalent controls elsewhere let advertisers bid differently for lapsed customers, and reporting on customer acquisition cost now sits in the interface. These controls only produce different outcomes for a business that knows which customers lapsed, what they were worth, and which are worth winning back — the same private information dependency that runs through every part of this subject.

Publishers and lead resellers face the sharpest version of the problem

Two business models sit at the extreme end of the reach-to-conversion shift, and their experience is a preview of pressures that reach everyone else later.

Publishers monetise attention directly. When attention arrives through search and search stops sending it, the revenue line moves immediately. The Pew data quantifies the mechanism: clicks falling from 15% to 8% of searches when an AI summary appears, and only 1% of visits producing a click on a source cited inside the summary. Google’s own Network advertising revenue — the line that places ads on third-party sites — fell 4% in the first quarter of 2026 to $6.97 billion while every Google-owned surface grew. A business whose entire model is renting attention to advertisers is being disintermediated by an interface that answers the question without the visit.

The responses available are all versions of moving away from reach. Registration and authentication convert anonymous audience into addressable audience that survives browser and consent changes. Subscription revenue substitutes reader payment for advertiser payment. Content licensing to model providers monetises the ingestion directly rather than hoping for referral traffic. First-party data products let publishers sell audience access on their own measurement terms. Each of these reduces dependence on the click, and each requires knowing far more about individual reader value than a page-view business ever needed to know.

Lead resellers and marketplaces face a different squeeze. Their product is a lead, their margin is the difference between what they pay to generate it and what a buyer pays for it, and their traditional advantage was arbitrage — buying traffic more cheaply than competitors. That arbitrage has thinned for the same reason it thinned for everyone: automated auctions price inventory closer to its value.

What replaced it is quality accountability. Buyers who have built conversion intelligence now know which sources produce leads that close. A lead buyer running offline conversion imports and closed-loop reporting can rank its suppliers by closed revenue per lead rather than by cost per lead, and will pay a premium for the top of that ranking while cutting the bottom entirely. A reseller that cannot demonstrate downstream outcome quality is competing on price in a market where its buyers can now see quality, which is the worst possible position.

The corollary is an opportunity for resellers who invest in the measurement their buyers use. A supplier that can report closed-won rates by source, geography and lead attribute, and can price accordingly, sells a different product than one shipping undifferentiated form fills. The same logic applies to affiliates, who face buyers increasingly willing to run holdout tests on affiliate traffic and increasingly unimpressed by last-click coupon-code attribution.

Both models illustrate the general point in unusually clear form. When reach was scarce and measurement was crude, intermediaries captured value by controlling access to audiences. Now that reach is abundant and measurement is improving, value accrues to whoever can prove which outcomes were caused. That reallocation is still in progress, and the businesses on the losing side of it are mostly still reporting the metrics that describe the world as it was.

A build order for teams starting from a broken setup

Most organisations reading this already spend money on paid media and already have a measurement setup that partly works. The sequence below assumes that starting point and orders the work by return per unit of effort rather than by architectural elegance.

Weeks one and two: audit before building. List every conversion action in every ad platform and analytics property, with its counting rule, attribution window, primary or secondary status, and creation date. Compare total platform-reported conversions and revenue against finance figures for the same three-month period. Check the consent implementation with a tag debugging tool under both accept and reject. Identify which conversions are observed and which are modelled. This produces a written document that almost no organisation has, and the discrepancies it surfaces usually redirect the rest of the plan.

Weeks three and four: fix definitions and consent. Reduce primary conversions to one per bidding goal. Remove duplicate tags. Set counting rules and attribution windows to match the sales cycle and document the choice. Implement or repair advanced consent mode with all four signals passing correctly, and verify that modelled conversions appear where they should. For accounts serving the EEA and UK, this step is the difference between measurement working and measurement being switched off by the platform.

Weeks five to eight: build server-to-server conversion interfaces. Meta’s Conversions API and Google’s enhanced conversions first, because they recover the largest conversion volume per engineering hour. Use a shared event ID for deduplication. Pass as many hashed identifier fields as legally available, and monitor the match quality score as an ongoing metric rather than checking it once. Add other platforms in order of spend.

Months three and four: assign values. Calculate contribution per conversion — order value less cost of goods, returns, payment and fulfilment — and pass that instead of gross revenue. For lead businesses, build an interim lead score and pass expected value. Verify outbound values against finance before switching the bid strategy, then switch and hold the configuration still for at least four weeks so the effect is observable.

Months four to six: connect the outcome systems. Capture click identifiers at submission and store them against CRM records. Build scheduled offline conversion imports for the stage that matters. Establish the warehouse joins between site, CRM, order system and ad platforms if the data volume justifies it. Treat these pipelines as production systems with monitoring, because platform API paths change — Google’s Customer Match cutover on 1 April 2026 broke jobs that had run unattended for years.

Month six onward: start experimenting. Run one properly powered geo holdout or in-platform lift test per quarter, beginning with the channel carrying the highest spend and the widest uncertainty. Branded search and retargeting are the usual first candidates because the published evidence suggests the largest gap between reported and incremental performance sits there. Set the testing budget as a standing line rather than a per-test approval.

Month nine onward: add portfolio modelling if the scale justifies it. A business spending under roughly seven figures a year on media across three or four channels will get more from experiments than from a model. Above that, and with two to three years of clean channel history, a Bayesian model calibrated by experiment results answers budget-level questions that experiments alone cannot.

Two things to skip until later, despite vendor pressure. A customer data platform purchased before the conversion definitions are fixed will faithfully distribute bad definitions to more destinations. And multi-touch attribution software bought to resolve channel disputes will produce a fourth number that disagrees with the other three, because the underlying observability problem is not a software problem.

One measurement of the programme itself is worth instituting from the start. Track the share of reported conversions that are observed rather than modelled, and the share of revenue that reconciles to finance. Both should improve monthly. If neither moves after a quarter of work, the programme is producing activity rather than intelligence, and the diagnosis is usually that nobody with authority over the ad accounts is acting on what the measurement says.

Team structure, skills, and the hiring mistakes that stall progress

Conversion intelligence fails organisationally more often than technically. The work sits across marketing, analytics, engineering and finance, and in most companies it belongs to nobody with the authority to complete it.

The skills required do not match the skills most marketing teams hire for. Platform certifications, campaign management experience and creative judgement remain useful, but the work described in this analysis needs SQL, an understanding of event schemas and identity resolution, familiarity with API authentication and scheduled jobs, enough statistics to design a test and read a confidence interval, and enough finance literacy to calculate contribution margin. The profile closest to this in most organisations is an analytics engineer, and marketing departments rarely have one.

The common hiring mistake is recruiting a senior performance marketer to solve a data engineering problem. The appointment is logical — the symptom appears in the ad accounts — and it fails because the person hired can diagnose the problem precisely and cannot fix it. The second mistake is the reverse: hiring a data engineer with no marketing context, who builds a technically sound pipeline carrying conversion definitions nobody validated against how the business makes money.

Agency arrangements introduce a structural incentive problem worth naming plainly. An agency paid a percentage of media spend has no financial reason to demonstrate that a quarter of that spend is not incremental. An agency paid a fixed retainer to manage campaigns has no reason to spend three weeks on server-side infrastructure that produces no visible campaign activity. Neither of these is dishonesty; both are what the contract rewards. An organisation that wants incrementality evidence should either commission it from a party with no stake in the media budget or pay explicitly for the measurement work as a separate line rather than expecting it as a by-product of campaign management.

The division that works in practice keeps three responsibilities distinct. The business defines what counts as an outcome and what it is worth, because only it knows the margins and the sales process. Whoever owns the data infrastructure — internal or contracted — builds and maintains the pipelines that move those definitions into the platforms and the outcomes back out. The media team operates campaigns against the resulting signals. When the first responsibility is delegated to either of the others, the definitions drift toward whatever is easiest to measure.

Documentation is the unglamorous requirement that determines whether any of this survives staff turnover. A conversion definitions document, a data flow diagram showing what travels where under which consent state, a record of attribution window choices with the reasoning, and a log of experiments with their designs and results. Organisations without these rebuild the same understanding every time someone leaves, and rediscover the same duplicate tag every two years.

One realistic note on scale. A business spending €20,000 a month on media cannot justify an analytics engineer, a warehouse and a quarterly testing programme. The proportionate version at that scale is correct conversion definitions, working consent signalling, two conversion APIs, margin-based values, and one annual holdout test — perhaps twelve days of specialist work in the first year and two days a quarter after that. That is achievable, and it puts a small advertiser ahead of most of its larger competitors, because the larger competitors usually have more tooling and the same unexamined definitions.

Tooling costs, vendor claims, and where the money actually goes

The measurement stack has a vendor for every layer, and the pricing is opaque enough that many organisations discover the real cost only after committing. A realistic breakdown helps in deciding what to buy and what to build.

Server-side collection carries an infrastructure cost that scales with traffic. A container running on managed cloud hosting for a mid-traffic site typically lands in the low hundreds of euros a month, rising with request volume; managed hosting services that handle the configuration cost more but remove the maintenance burden. The larger cost is the initial implementation and the ongoing attention, because a server container that silently stops forwarding events looks exactly like a performance decline in every report.

Consent management platforms range from free tiers adequate for small sites to enterprise contracts in the tens of thousands annually where multi-brand, multi-jurisdiction configuration and audit logging are required. The relevant question when comparing them is not banner appearance but whether the platform correctly propagates consent state to a server container, supports the current framework version, and produces records that survive a regulator’s request.

Customer data platforms are where budgets disappear fastest, frequently at six figures annually for enterprise deployments. The honest assessment is that many organisations buying one need a warehouse and a scheduling tool rather than a platform, and would be better served spending a fraction of the licence on an analytics engineer. A platform distributing unvalidated conversion definitions to more destinations increases the speed at which bad data propagates.

Modelling is free as software and expensive as labour. Meridian and Robyn cost nothing to license. Fitting, validating and interpreting a model requires an analyst with Bayesian competence, whether employed or contracted, and the data preparation preceding the model usually consumes more time than the model itself. Certified partner engagements exist for organisations without internal capability and are priced accordingly.

Incrementality vendors charge platform fees plus, in some arrangements, a share of measured savings. The larger cost of experimentation is not the vendor but the foregone revenue in control markets, which is the actual price of the information and should appear in the budget explicitly rather than being discovered afterwards.

Four vendor claims deserve direct scepticism.

Precise recovery percentages. Any promise that a product recovers a specific share of lost conversions is describing an outcome that depends on traffic volume, consent rate, geography and match quality — variables the vendor does not control. Ranges are honest; single figures are marketing.

Complete attribution. Claims of full-journey visibility in an environment where a large share of exposure is structurally unobservable describe a modelled reconstruction. That has value when labelled correctly and is misleading when presented as observation.

Cookieless tracking that identifies individuals. Some products marketed under this heading rely on probabilistic device recognition. In the EEA and UK, techniques of that kind attract the same consent requirements as cookies, and a product that promises identification without consent is offering a compliance liability rather than a measurement solution.

Attribution software as a dispute resolver. Adding a fourth measurement system to three that already disagree produces a fourth number. The disagreement is caused by missing observation, and no vendor’s model creates data that was never collected.

The proportionate total for a mid-market advertiser doing this properly — consent platform, server-side hosting, one testing engagement a year, warehouse compute, and specialist time — sits well below what a single misallocated month of media spend costs at scale, which is the comparison that makes the business case. The expensive option is continuing to allocate seven figures against numbers written by the recipients.

Regulatory pressure is now part of the conversion architecture

European enforcement has moved from warnings to fines large enough to change platform product behaviour, and each change alters what advertisers can measure. Treating regulation as a legal matter separate from measurement no longer works.

The Digital Markets Act produced its first penalties in April 2025: €500 million against Apple for App Store anti-steering restrictions, and €200 million against Meta for its pay-or-consent advertising model, covering non-compliance from March 2024, when gatekeeper obligations became binding, to November 2024, when Meta introduced a revised model. The Commission’s objection was that a binary choice between full personalisation and paid ad-free access did not offer users a less personalised but otherwise equivalent service.

The resolution changed the advertising product. On 8 December 2025 the Commission accepted Meta’s undertaking to offer EU users a genuine third option, presented from January 2026: consent to full data sharing and fully personalised advertising, or share less personal data and receive advertising with more limited personalisation based largely on context, age, coarse location and gender. Every EU user who selects the reduced option becomes a person whose behaviour is not available for behavioural targeting, custom audiences or lookalike modelling, while remaining a person who can be reached. European campaign performance on Meta now depends partly on the take-up of that option, which advertisers cannot observe directly and can only infer from their own results. The European Data Protection Board’s separate view — that large platforms generally cannot obtain valid consent when the only alternative is payment — reinforces the direction under data protection law rather than competition law.

Enforcement against Google escalated in July 2026. On 23 July the Commission issued two non-compliance decisions totalling €890 million: €460 million for self-preferencing its own comparison, shopping, hotel, travel and sports services in Search rankings, and €430 million for restricting app developers on Google Play from steering users to cheaper channels outside the store. Google was given 60 days to end both infringements, with exposure to periodic penalty payments of up to 5% of worldwide turnover for continued non-compliance, and may appeal. Separately, specification decisions adopted on 16 July 2026 require Google to share anonymised search data with rival search engines and AI services from January 2027, and to open eleven Android system-level features to competing AI assistants by the next major Android release in 2027.

The forward-looking element matters more than the fine. The Commission stated it is examining how the principles prohibiting self-preferencing apply to AI Overviews and AI Mode — the generative summaries occupying the top of the results page. It reached no finding, and deferred the question. If that examination concludes that a proprietary model generating the dominant answer above all third-party content constitutes the same behaviour as promoting Google Hotels above rivals, the shape of the European results page changes again, and with it the traffic and conversion patterns every European advertiser is currently planning against.

Three planning consequences follow for anyone building measurement in or for European markets.

Signal availability is a policy variable, not a technical constant. It can shrink through enforcement and expand through settlement, on timescales shorter than a martech implementation. Architectures that depend on any single identifier source are fragile by construction.

Contextual and creative quality carry more weight where behavioural signal is reduced. As a larger share of EU impressions is delivered on context rather than on profile, the differences between advertisers shift toward message and offer, which are harder to buy and harder to copy.

First-party data becomes the stable input. Consented, authenticated, business-owned outcome data is the one asset unaffected by DMA remedies, browser policy shifts and platform product changes. Every regulatory development of the past two years has raised its relative value, and none has reduced it.

Privacy and data handling limits that shape what can be measured

Conversion intelligence involves collecting more outcome data and moving it between more systems, which raises the legal stakes of getting the handling wrong. The constraints are not obstacles to work around; they define what a defensible system looks like.

Under the GDPR, every processing operation needs a lawful basis, and analytics and advertising measurement are treated differently across supervisory authorities. Storing or accessing information on a device for advertising purposes requires consent under the ePrivacy Directive regardless of what basis is claimed for subsequent processing, which is why the consent layer sits upstream of everything else. Uploading customer records to a platform for audience matching or offline conversion import is a separate processing operation from collecting them, and needs a basis of its own plus disclosure in the privacy notice — a step routinely skipped when a marketing team discovers Customer Match.

Hashing is pseudonymisation, not anonymisation. Sending a hashed email address to an ad platform for matching does not remove the data from the scope of the GDPR, because the platform can and does link it to an identified person. That is the entire function. Teams that describe hashed uploads as anonymous in internal documentation are creating a record that will not survive scrutiny.

Data minimisation applies with force to warehouse projects. The instinct when building a measurement warehouse is to collect everything in case it becomes useful. The legal requirement is to collect what is necessary for a specified purpose and to define retention periods. A warehouse holding five years of granular behavioural records because nobody set a deletion policy is a liability that grows quietly.

Special category data creates hard limits in specific verticals. Health, finance, sexual orientation, religion and political opinion attract stricter treatment, and inference matters as much as collection — a conversion event named after a specific medical condition, a pregnancy product category, or a debt-relief enquiry can constitute health or financial vulnerability data once it leaves the business. Platform advertising policies also restrict personalisation in sensitive categories, so an event name or audience segment that reveals the underlying condition is both a compliance exposure and a policy violation. Generic event naming with the sensitive detail retained internally solves both.

Processor arrangements need to be current. Enforcement in the United States has begun to cite advertising technology contracts specifically — reporting on a California settlement in September 2025 identified deficient ad-tech contract terms among the findings — and stale data processing addenda that predate a company’s current stack are a common weakness. Any vendor receiving event data needs a current agreement covering purpose limitation, sub-processors and transfers.

The practical rule for what leaves the building is short. Send hashed identifiers, not raw ones. Send values and outcome types, not free-text notes or internal comments. Send event names that describe a commercial action rather than a condition or characteristic. Keep the granular, sensitive and joined data inside systems the business controls, and export only what a platform needs to match a conversion and price a bid. A measurement architecture that keeps the interesting data internal and sends the minimum outward is both more defensible legally and usually more accurate, because it forces the definitional work to happen before the export rather than after.

Creative became the last uncommoditised input

When placement, audience selection, bidding and budget pacing are all decided by the same class of model across every platform, the variables left to an advertiser are the offer, the conversion signal it feeds in, and the creative it supplies. Creative is the one of those three that competitors can see, and it is still the largest driver of performance variance between accounts buying the same inventory.

Production has stopped being the constraint. Meta reported that advertisers using its AI creative tools doubled to eight million in the first quarter of 2026. Generating fifty variants of an ad is now a matter of hours for anyone, which removes the advantage that belonged to brands with production budgets and replaces it with a different problem: when everyone can make more creative than they can evaluate, the bottleneck moves from production to measurement.

Evaluating creative properly is harder than most testing practice acknowledges. Running two ads in one ad set and reading the results is not an experiment, because the delivery system is choosing who sees which one. Divergent delivery means the better-performing ad may simply have been shown to a more responsive audience, and the effect runs in the same direction as the platform’s incentive to demonstrate that its allocation was correct. Research on this bias has accumulated to the point where in-platform creative comparisons should be read as directional at best.

The methods that produce trustworthy creative conclusions are the same ones that work elsewhere. Randomised holdouts where the platform assigns users rather than choosing between creatives. Separate campaigns with equal budgets and matched targeting, accepting the extra cost. Geo-split creative tests for larger changes, where a new campaign concept runs in one set of markets and the incumbent in another. Each is more expensive than reading the ad-level report, and each answers a question the ad-level report cannot.

Proxy metrics deserve careful handling. Hook rate, hold rate, click-through rate and thumb-stop measures are useful diagnostics for why a creative is failing, and unreliable as predictors of what it will earn. Reporting from practitioner analyses in 2026 has repeatedly described generatively produced creative winning on click metrics while underperforming on conversion, with the gap widening at higher average order values — plausible, consistent with what many teams observe, and not yet supported by a controlled published study. The safe reading is that engagement metrics and outcome metrics can move in opposite directions, and that any creative programme managed on engagement will eventually discover this expensively.

Two structural points matter more than any individual test. Distinctive brand assets — the colour, character, sound, phrase or format a category associates with one company — retain their function in an environment of infinite variation, and arguably gain it, because they are the only element that makes generated output recognisable as belonging to a specific brand. And the offer inside the creative usually moves conversion more than the execution around it: a genuinely better guarantee, price, delivery promise or bundle outperforms a better-looking version of a weaker proposition, which is why creative testing that never varies the offer plateaus quickly.

Brand investment under a conversion-led budget

There is a real objection to everything argued above, and it deserves to be stated at full strength rather than acknowledged and dismissed.

Every method described in this analysis measures effects that occur inside the measurement window. Geo holdouts run for four to eight weeks. In-platform lift tests run for two to four. Conversion values are assigned to actions observable within a sales cycle. A marketing activity whose main effect is to make a purchase more likely eighteen months from now — a category association, a memory structure, a reason to think of one brand first when a need arises — produces almost nothing inside those windows. A measurement regime that funds only what it can measure will systematically defund the activity whose returns arrive latest, and it will look correct doing so at every step.

The evidence for long-term effects is not speculative. Analysis of advertising results databases by researchers in the UK institute tradition has consistently found that campaigns generating large business effects allocate a majority of budget to brand-building rather than to short-term activation, with the frequently cited benchmark sitting near a 60:40 split for many categories. The mechanism is that activation harvests demand that exists while brand-building creates the demand to harvest later. A business that shifts entirely to activation reports improving performance for several quarters and then finds its activation getting more expensive, because the pool of people predisposed toward it is shrinking.

The interaction with conversion intelligence is uncomfortable in a specific way. Incrementality testing is at its most decisive on exactly the channels closest to purchase — branded search, retargeting, shopping ads — and its findings there are usually that reported returns were overstated. That is correct and useful. The risk is that the same organisation, having learned to demand causal evidence, applies the standard uniformly and cuts brand investment because no four-week test can defend it. The evidential standard is then being applied asymmetrically in practice: activation gets judged on measurable lift, brand gets judged on the same measurable lift, and the method that cannot detect brand effects is treated as having disproved them.

Three approaches keep the discipline without the distortion.

Longer-window portfolio experiments. A geo holdout on brand activity needs six to twelve months rather than six weeks, and it needs to accept regional noise as the price of measuring a slow effect. Few organisations run these. The ones that do get an answer nobody else has.

Intermediate outcome measurement. Branded search volume, direct traffic share, unaided awareness in tracking surveys, and share of category conversation move earlier than sales and correlate with the mechanism brand investment is supposed to produce. None is a substitute for a revenue effect, and all are better than nothing.

Modelling with carryover. This is where marketing mix modelling earns its place. Adstock and carryover parameters exist precisely to capture effects that persist beyond the period of spend, and a model fitted over two or three years can attribute a share of current baseline demand to prior brand activity in a way no experiment inside a quarter can. The estimate is uncertain and it is directionally informative, which is the honest description of most useful marketing evidence.

The defensible position is that conversion intelligence should govern the allocation of activation budget and inform, rather than determine, the split between activation and brand. The failure mode of the reach era was spending on attention nobody could connect to outcomes. The failure mode of the conversion era is spending only on outcomes visible inside a test window, and it is the more insidious of the two because the reporting improves as it happens.

Failure modes of conversion intelligence programmes

Programmes fail in recognisable patterns. Most of these are visible in retrospect and preventable in advance.

Measurement without authority. The most common pattern. An analyst or contractor builds accurate measurement, produces a finding that a large campaign is not incremental, and nobody with budget control acts on it because the campaign belongs to someone senior or a client relationship depends on it. The measurement was correct and the programme produced nothing. Deciding in advance who will act on an uncomfortable result, and what threshold triggers action, is more important than the technical quality of the test.

Goodhart’s law applied to conversion values. Once a team is measured on ROAS calculated from values it supplies, the values drift upward. Lifetime value assumptions get generous. Predicted repeat rates get revised in the favourable direction. Nobody commits fraud; a series of individually defensible assumptions accumulates in one direction. Value inputs need periodic verification against realised finance figures by someone who is not measured on the resulting ratio.

Treating modelled conversions as observed. Reporting a conversion count without knowing what share is model output leads to confident decisions built partly on a platform’s inference about the business. The share should be monitored and reported alongside the count.

Stopping tests when the result looks right. An experiment checked daily and stopped when it reaches significance produces a biased estimate. Duration and success criteria have to be fixed before the test starts, and honoured when the interim numbers are encouraging.

Reallocating on noise. Small samples produce large swings. A campaign that appears to have doubled its return over two weeks at low conversion volume has usually not changed at all. Teams that reallocate weekly at low volume spend their time chasing variance, and the churn itself damages performance because bidding models restart learning with every change.

Over-fitting to a single test result. A geo holdout run during a promotional period, in a seasonal peak, or while a competitor was running an unusual campaign measures that situation. Treating one result as a permanent property of a channel leads to a cut that was correct in March and wrong by September. Findings need re-testing at intervals.

Pursuing completeness instead of usefulness. Some teams spend two years attempting to reconstruct full journey visibility that current conditions make impossible, and make no decisions differently in the meantime. The purpose is better allocation, not a complete record. A measurement system that resolves 70% of the question and changes behaviour beats one that promises 95% and ships next year.

Consolidating into one number two functions define differently. Marketing’s revenue figure and finance’s revenue figure diverge because of returns, cancellations, taxes, currency and timing. Presenting a single blended ratio to a board without a reconciliation invites a dispute that discredits the whole programme at the worst possible moment. Reconciling once, documenting the difference, and reporting both is dull and protects everything else.

Market reading for the next four quarters

Forecasting in this area has a poor record — the industry spent six years preparing for a cookie deprecation that never happened — so what follows is separated by confidence level rather than presented as a single outlook.

High confidence, because the mechanism is already visible. Auction prices continue rising. Both large platforms are increasing capital expenditure into ranking and retrieval systems, both attribute demand growth to the resulting performance improvements, and both price that improvement into the auction. Advertisers should plan 2027 budgets assuming another year of high single-digit to low double-digit cost increases per unit of reach, and assuming that the improvement they receive in exchange is available to every competitor simultaneously.

Automation continues absorbing execution. The direction announced at Google Marketing Live in May 2026 — a unified Gemini agent spanning Google Ads, Analytics, Merchant Center and the marketing platform, natural language campaign guidance, one-click adoption toggles — points toward campaign construction becoming a conversational task. The skill of building a campaign continues deflating in value. The skill of deciding what the campaign should be told to buy does not.

Measurement fragmentation persists rather than resolving. With no industry standard replacing third-party identifiers and each platform controlling its own conversion interface, the near future involves several imperfect measurement systems reconciled by whoever is willing to do the work. Vendors will keep selling unification; the underlying observability problem is not one a vendor can solve.

Moderate confidence, because the direction is clear and the pace is not. AI-mediated discovery keeps taking share of research behaviour, and the traffic it returns keeps converting differently from search traffic — better in retail, still worse in travel, varying enough by sector that aggregate figures mislead. Businesses that segment and measure it will price it correctly; those that leave it inside direct traffic will keep making decisions with their best-converting channel invisible.

Regulatory intervention continues reshaping signal availability in Europe. The Commission has explicitly deferred the question of whether generative summaries constitute self-preferencing under the same principles applied to Google’s vertical services, and has ordered search data sharing with rivals from January 2027. Both threads have the potential to alter European results pages and traffic patterns inside the planning horizon.

Incrementality moves from advanced practice to procurement requirement. With published adoption already past half of US brand and agency marketers and retail media buyers ranking it their top metric, the trajectory points toward advertisers demanding causal evidence as a condition of spend rather than as an occasional audit.

Low confidence, genuinely uncertain. Agentic commerce volume. The protocol infrastructure is being built rapidly by well-capitalised parties, and consumer adoption of in-chat purchasing has so far been thin enough that one flagship surface was wound down within months. Merchants should keep feeds and structured data in good order — which pays off regardless — and defer larger commitments until transaction volume rather than protocol announcements justifies them.

The one structural prediction worth making is about distribution of capability rather than technology. The gap between advertisers who have built conversion intelligence and those who have not will widen rather than narrow, because the feedback loop compounds. A business feeding accurate values into automated bidding gets better allocation, which produces better outcome data, which improves the next allocation. A business feeding conversion counts gets a system trained on a distorted objective, and every improvement the platforms ship makes that system better at pursuing the wrong target. The platforms will keep improving for both. Only one of them gets to keep the benefit.

Open questions the current evidence cannot settle

Several questions central to this subject remain genuinely unresolved, and treating them as settled is how expensive mistakes get made.

The true size of the incrementality gap by channel. Published figures cluster between 20% and 60% overstatement, with case studies describing 1.5 to three times on branded search and retargeting. Those figures come from vendors with an interest in the gap being large and from case studies selected for being interesting. There is no large-scale independent audit across advertisers, categories and geographies. The direction is well supported; the magnitude for any specific business is only knowable by testing it.

Whether AI referral conversion advantages persist as volume grows. The current advantage may reflect early adopter composition — people using assistants for shopping in 2026 skew toward deliberate, high-intent researchers. As usage broadens toward the general population, the mix may look more like ordinary search traffic. Adobe’s own travel data, where AI referrals still convert below other sources while the gap narrows, is consistent with either interpretation.

Whether modelled conversions are accurate enough to bid on. Platforms report modelled and observed conversions together and disclose little about model performance. No independent validation exists comparing modelled conversion counts against ground truth at account level. Advertisers are bidding against numbers whose error bars nobody outside the platforms can estimate.

Whether agent-mediated purchasing becomes material. Protocol development has outpaced demonstrated consumer demand. The infrastructure could support a large shift or remain a niche within a few categories. Anyone claiming to know which is forecasting, not analysing.

Where the AI Overviews self-preferencing question lands. The European Commission has stated it is examining whether the principles in its July 2026 decision extend to generative summaries, and has not decided. A finding either way changes European search economics materially, and no one outside the process can predict the outcome or the timing.

Whether creative generated at volume produces worse outcomes at higher price points. Practitioner reporting describes generated creative winning engagement metrics and losing conversions as order values rise. It is a plausible pattern with a plausible mechanism and no controlled published study behind it.

How much brand investment conversion-led measurement is quietly destroying. The mechanism described earlier is well established in advertising research. The magnitude of the effect in businesses that adopted strict incrementality discipline over the past three years is unknown, and by construction will not be visible for several more years. It is the question most likely to be answered painfully rather than analytically.

A final honest caveat about the argument in this analysis. The claim that conversion intelligence is the remaining durable advantage rests on the assumption that platforms cannot acquire the private business information that makes it work — margins, lead quality, lifetime value, sales outcomes. That assumption holds today because the information sits in systems advertisers control and is only shared deliberately. If platform integrations reach far enough into commerce and customer relationship systems that the information flows automatically, this advantage commoditises too, and the search moves on to whatever remains scarce after that. Every advantage in advertising has eventually been automated. The reasonable position is that this one lasts longer than the last one, not that it lasts forever.

Questions marketers ask about conversion intelligence

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

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

Meta Reports First Quarter 2026 Results Primary earnings release containing the first quarter 2026 figures for revenue, ad impressions delivered, average price per ad and daily active people.

Meta Platforms quarterly report on Form 10-Q, period ended 31 March 2026 Regulatory filing giving the revenue breakdown and management’s explanation of impression and pricing growth.

Alphabet first quarter 2026 earnings call Investor relations page for the call in which Google management discussed Search revenue growth, Network decline and Gemini deployment across the ads stack.

Alphabet earnings call, Q1 2026: Sundar Pichai’s remarks Company account of the quarter, including Search and Other advertising growth and the role of AI Mode and AI Overviews in query volume.

Alphabet reports first quarter 2026 earnings Business press summary with the segment figures for Search, YouTube advertising, Cloud and traffic acquisition costs.

Alphabet Q1 2026: Google Network ad revenue falls 4% as AI reshapes the web Trade analysis of the Network advertising decline against growth in Google-owned surfaces, with the timeline of publisher traffic complaints.

Meta Q1 2026: $56.3B revenue as AI tools double advertiser adoption Coverage of the quarter’s advertising product detail, including AI creative tool adoption and the extension of the Adaptive Ranking Model to offsite conversions.

Google users are less likely to click on links when an AI summary appears in the results Pew Research Center analysis of real browsing behaviour from 900 US adults, with click rates, session termination rates and summary characteristics.

Google’s AI Overviews are hurting clicks: Pew study Trade coverage summarising the Pew findings and their implications for publishers and search marketers.

Pew Research confirms Google AI Overviews is eroding web ecosystem Detailed reading of the Pew dataset, including query counts and the share of searches producing an AI summary.

Google pulls the plug on Topics, PAAPI and other major Privacy Sandbox APIs Trade reporting on the October 2025 retirement announcement and the UK regulator’s release of Google from its Privacy Sandbox commitments.

Google is retiring Privacy Sandbox Report quoting the full list of retired technologies from the announcement by the Google vice-president responsible for the programme.

Google’s changing approach to third-party cookies Timeline of the deprecation reversals from the original roadmap through the April 2025 confirmation, with the consequences for measurement.

The rise and fall of Google’s Privacy Sandbox Academic longitudinal measurement of Privacy Sandbox adoption across the web, documenting the collapse in Protected Audience API usage before retirement was announced.

Privacy Sandbox Reference overview of the initiative’s history, including the 2023 general availability milestones for Topics, Protected Audience and Attribution Reporting.

Meridian is now available to everyone Launch announcement for the open-source Bayesian marketing mix model, describing calibration against incrementality experiments and the certified partner programme.

Google’s Meridian makes MMM more accessible Analyst assessment of the release, including survey data on the share of US marketers already using marketing mix modelling.

Marketing mix modelling with Meridian Google’s own guidance on the model, covering query volume as a control variable and the research framing around the gap between modelling and action.

Google Marketing Live 2026: news and announcements Official collection of the May 2026 announcements, including AI Search ad formats, bidding and budgeting changes and retail commerce features.

Google Marketing Live 2026 collection Business-facing summary describing the unified Gemini agent spanning Ads, Analytics, Merchant Center and the marketing platform.

Steering the machine: our take on the agentic shift at Google Marketing Live 2026 Agency analysis of the announcements, including the Product Value Adjustments pilot and the argument that execution friction is disappearing from advertising.

New features and announcements Google Ads release documentation covering customer lifecycle goals, retention bidding and acquisition cost reporting in Performance Max.

Commission fines Google €890 million for breaches of the Digital Markets Act Official European Commission decision notice covering the two non-compliance findings, the individual fine amounts and the ranking obligations for gatekeepers.

EU fines Google €890 million and gives it 60 days to fix search Detailed trade reporting on the decision, the compliance deadline, periodic penalty exposure and the parallel specification decisions on search data and Android.

Meta commits to give EU users choice on personalised ads under the Digital Markets Act Commission statement on the undertaking to offer EU users a less personalised advertising option from January 2026.

Reviewing European antitrust activity in 2025 and what it all means for 2026 Policy analysis of the pay-or-consent finding and the wider enforcement pattern under the Digital Markets Act.

AI referrals to travel sites surge 194% as visitor quality improves Reporting on Adobe Analytics data covering AI referral growth, engagement metrics and the conversion gap in travel against the reversal in retail.

AI referrals drive higher ecommerce traffic and conversions Coverage of Adobe and Shopify data on AI-referred retail traffic growth, landing page composition and revenue per visit.

Commerce media ad spending forecast 2026 Analyst forecast for commerce media spending and its projected share of total US digital advertising.

US ad spending 2026 Forecast covering total and digital US advertising spending by channel and platform, including Meta’s projected net digital ad revenues.

Digital retail media advertising spending in the United States Statistical series for US retail media spending with year-over-year growth and forecasts through 2029.

GeoLift 101: your guide to geo-based incrementality testing Practitioner explanation of geo experiment design, synthetic control construction and the reasons market-level testing survives identifier loss.

Why are marketers turning to quasi geo-lift experiments Technical treatment of experiment designs, including the divergent delivery problem that undermines randomised tests run inside ad platforms.