From Traffic to Recommendation

Why the measurement question arrived now

For most of the commercial web, the default marketing scoreboard rested on a single assumption: that discovery produces a visit, and that counting visits therefore counts discovery. Sessions, rankings, click-through rate, and cost per click all inherit that assumption. It was never perfectly true, since word of mouth, print, and offline recall always leaked, but it was true enough that a traffic report could stand in for a visibility report without much argument.

AI-mediated discovery narrows that assumption in a specific way. When a system answers a question directly, the answer can be delivered without a visit, and the visit is the thing the measurement system was built to see. This does not mean traffic is obsolete or that AI has replaced search. It means one instrument now reads a smaller share of the phenomenon it was hired to read, and the gap between the instrument and the phenomenon is widening rather than holding steady.

A Pew Research Center analysis of a browsing panel of 900 US adults found that among visits to Google search pages, users clicked a traditional result link on 8 percent of pages that carried an AI summary, compared with 15 percent of pages without one, and clicked a source cited inside the summary on 1 percent of such pages. That is one platform, one country, and March 2025 data, so it should not be generalized into a universal law. It is enough, though, to establish that the click is no longer a stable proxy.

What changed in the path, precisely

It helps to be exact about which part of the mechanism moved. Eligibility largely did not. Google documents that for a page to appear as a supporting link in AI Overviews or AI Mode it must be indexed and eligible to appear in Search with a snippet, and states there are no additional technical requirements beyond that. The same documentation discourages inventing AI-specific text files or chunking tricks. The retrieval surface, in other words, is substantially the familiar one.

What moved is delivery and settlement. Content is fetched by machines at one rate and returns human visitors at another, and those two rates are no longer close. Cloudflare publishes a crawl-to-refer ratio on Radar, computed as HTML requests from user agents associated with a given search or AI platform divided by HTML requests whose referer header names that platform. It is a blunt instrument by design, and it is publicly observable rather than inferred from any model's internals, which is exactly why it is useful.

The consequence is that a site can be read heavily, represented accurately, and cited fairly while its analytics show very little. Nothing in that sequence is a failure of the site. It is a settlement problem, and it is not fixed by working harder on the traffic number.

What the platforms report, and what they withhold

Reporting has improved, and it is worth knowing its exact shape rather than assuming it. In June 2026 Google added generative AI performance reports to Search Console, breaking out impressions from AI Overviews, AI Mode, and generative AI features in Discover. Google described this as isolating data that was already inside overall Search performance totals rather than as new data, which means aggregate figures do not change when the breakout appears.

The launch scope matters as much as the launch. The generative AI reporting covers impressions with page, country, device, and date dimensions. Click, click-through rate, and query breakdowns for those surfaces were not part of the initial release. So the report answers whether a URL surfaced, and does not answer what that surfacing was worth. Treating an impressions chart as a demand chart is the most available mistake in this area right now.

Referral-side measurement has a different limit. Google Analytics classifies traffic using rule-based channel groups that key off the referrer and campaign parameters, and site owners can define custom channels with their own matching rules. That works when an assistant passes a referrer. When an assistant does not, or when a person copies a name out of an answer and types it into a browser, the arrival is indistinguishable from direct. This is a structural property of referrer-based attribution, not a configuration error, and no channel regex resolves it.

• Search Console generative AI reports: impressions available, click and query detail for those surfaces not available at launch
• Analytics channel grouping: accurate only for arrivals that carry a referrer
• Assistant interfaces: no publisher-facing report of how often a brand was named without a link

Three layers worth counting

AIOFacts proposes separating the scoreboard into three layers, because collapsing them is what produces confused reporting. The layers are presence, accuracy, and consequence. They fail independently, they are measured by different methods, and a strong result in one says little about the others.

Presence asks whether a system names the entity when someone asks a question the entity should own. It is measured by sampling answers, not by reading logs. Accuracy asks whether what the system says is correct: the description, the category, the location, the pricing model, the claims about what the business does. Accuracy can be poor while presence is excellent, and that combination is more damaging than absence, because it scales a wrong statement. Consequence asks what happens afterwards, and it is where conventional analytics still earns its keep.

Keeping the layers separate also disciplines the reporting language. Presence findings are sample estimates and should carry that caveat. Accuracy findings are checkable and can be stated flatly. Consequence findings are attributed, which is to say partially inferred, and should never be written as though causation were established.

Measuring presence without fooling yourself

Presence measurement has a methodological problem that is easy to underestimate: the same question asked twice can return different brands, in a different order, with different framing. A 2026 arXiv paper decomposing variance components in non-determinism for brand answers describes exactly this reproducibility issue and notes the common practice of resampling each prompt several times, often around five, before averaging. Single-shot querying cannot characterize presence, and a screenshot of one answer is an anecdote presented as a measurement.

Several other sources of variance sit on top of sampling noise. Providers update models, system prompts, retrieval configurations, and safety filters without publishing a changelog for publishers. Results can differ by geography, by account state, and by how recently the underlying index was refreshed. Platform behavior varies between assistants in ways that are not documented externally, so cross-platform comparisons describe different systems rather than one system measured twice.

The practical response is not to abandon the measurement. It is to make it boring and repeatable: a fixed prompt set written from real buyer questions, repeated sampling per prompt, a recorded date and platform version where one is exposed, and results reported as ranges rather than points. Movement is then interpretable. A single reading is not.

• Fix the prompt set before measuring, and change it deliberately rather than opportunistically
• Resample each prompt and report a range, not a single answer
• Log platform, date, and locale with every observation
• Never compare a sampled presence figure against a logged traffic figure as though they were the same kind of number

Reconciling the two scoreboards

The recommendation scoreboard does not replace the traffic ledger. It sits beside it, and the two are joined at outcomes rather than at sessions. Revenue, qualified enquiries, and signups remain countable. What becomes unreliable is the middle of the chain, where a session was previously assumed to carry the causal story from discovery to outcome.

There is some evidence that the smaller volume of AI-referred traffic behaves differently once it arrives. Adobe Digital Insights, measuring US retail sites in March 2026, reported that traffic arriving from AI sources showed an engagement rate 12 percent higher than non-AI traffic and viewed 13 percent more pages per visit, alongside very large year-over-year growth from a small base. This is vendor-published analysis of one sector in one country, so it should inform expectations rather than settle them. It does suggest that judging AI referrals purely on session count understates them.

Two supporting indicators are worth watching because they move when recommendation presence moves and they are measurable with tools already in place: branded search volume, and the share of direct arrivals that are new rather than returning. Neither proves anything on its own. Both are consistent with a discovery path that ends in a name being remembered rather than a link being clicked, and both are cheaper to track than any dedicated tooling.

What a defensible scoreboard contains

A reporting pack that survives scrutiny tends to look the same across organizations. It states what was sampled and when. It reports presence as a rate across repeated samples rather than a binary. It carries an accuracy register listing every incorrect statement found about the entity, with the source that appears to support the error, because that register is actionable in a way a score is not. It keeps generative impression data and referral data in separate sections, clearly labelled, because they count different events.

It also names its blind spots on the page rather than in a footnote. Answers delivered without a link, mentions inside voice responses, and assistant sessions inside closed applications are not observable from the publisher side today. A scoreboard that does not say so implies a completeness it does not have.

The failure mode to avoid is the composite index: a single AI visibility score built by weighting presence, accuracy, and referral data together. It looks decisive and it is unfalsifiable, because no external party can check the weights and the underlying components move for unrelated reasons. Report the components.

The limits of the evidence here

Most of what is documented concerns Google surfaces, because Google publishes publisher-facing reporting and documentation and most assistant operators do not. Findings about AI Overviews and AI Mode should not be read as findings about ChatGPT, Claude, Perplexity, Copilot, or Gemini in the assistant context, which expose no comparable reporting to the sites they read.

None of the sources cited here describe how any system internally selects or ranks what it cites, and this piece makes no claim about that. What is observable is external: whether a URL surfaced, whether a referrer arrived, whether an answer named an entity when sampled, and whether the statement it made was correct. That is a narrower evidence base than the confident dashboards currently on the market imply, and stating the narrowness is part of the measurement.