What are the limits of AI visibility measurement?

No, and the reason is mechanical rather than statistical: in the ordinary case the reader never clicks, so there is no event for any analytics system to record. Pew Research Center’s March 2025 browsing panel put clicks on a source inside an AI summary at about 1% of visits, and the magnitude of what a citation is worth as an impression is covered in the GEO versus SEO stage.² The field’s own 2026 survey rates the claim that citation scores predict clicks, conversions or revenue at very low confidence, its lowest grade.³

Three mechanisms break the chain, and knowing which is operating tells you which workaround is worth trying. The first is the missing click: a reader satisfied by the answer never becomes a session. The second is the missing referrer: chat clients frequently send none at all, so a visit that did happen arrives as direct traffic. The third is the indistinguishable agent: a Tow Center for Digital Journalism analysis published 30 October 2025 reports that to a website, an agentic browser’s agent “is indistinguishable from a person using a standard Chrome browser”, and that such browsers retrieved subscriber-only material the same vendors’ standard interfaces could not.⁴ That class of traffic is growing quickly: Cloudflare’s network-wide 2025 review reported user-triggered crawling growing more than fifteenfold across the year.⁵ The share of your AI-mediated touches that are invisible in analytics is therefore rising, not falling.

No workaround recovers the missing coefficient, and be suspicious of anything presented as one. State a modelling assumption as a sentence instead, with its inputs named and its unknowns marked unknown, and never render it as a chart. A chart is quoted as a measurement by the second meeting.

No, and this is the limit most often mistaken for noise in the data. A panel runs a clean-room session: fresh context, no signed-in account, a fixed locale, search left on, one run per prompt. A buyer holds none of that still. Anthropic documents that Claude “saves memory as a set of individual topics as you chat”, carries that into later conversations, and lets the user switch it off, so two people asking the same question are not querying the same system.⁷ Nothing you run reproduces their state.

The other half is non-determinism, and unlike personalisation it can at least be quantified. The March 2026 statistical treatment of AI visibility reports that “citation rankings are unstable across samples, not only among top-ranked domains but throughout the frequently cited domain set”, and that bootstrap intervals place many apparent differences between domains inside “the noise floor of the measurement process”.⁶ The 2026 critical survey records the same pattern from the audit side: commercial audits show “low source overlap, substantial run-to-run variability, and persistent fidelity gaps”.³

The working rule that follows is to stop treating one run as an observation. Repeat each prompt, report a rate with an interval rather than a point, and call a week-on-week move real only once it clears that interval; the survey’s own checklist asks studies for multiple paraphrases and repetitions for this reason.³ How many repeats that takes is arithmetic, worked out in the statistical-power article. What repetition never buys is the buyer’s own answer, which stays unobservable, so keep your session settings written beside the number rather than in an appendix.

Bounding a question means answering it with a stated range and a stated assumption instead of a point estimate, and it takes three sentences. Name the estimand: write down precisely what you wish you knew, as in “how many buyers formed an opinion of us inside an assistant last quarter”, because the vague version is what lets a proxy quietly stand in for it. Name the instrument: say what you actually have and what it measures, so that a 200-observation panel measures the share of your chosen questions that named you, on five engines, in a date range. Name the falsifier: say what observation would show the reading was wrong, which for a bound usually means naming the assumption whose failure breaks it.

Written out, a bounded claim reads like this: “On our 60-prompt panel, four of five engines named us in 28–36% of answers in July, at 200 observations per engine. We do not know how often those questions are asked, so this is not a market share. If our buyers ask questions materially different from our panel, this number does not describe them.” That paragraph is shorter than most executive summaries and it cannot be misquoted, which is the point.

The same discipline is what the research literature asks of itself. The 2026 critical survey’s recommendation is to build a retrievable page and then “measure retrieval, citation, and fidelity separately”, and its minimum-study checklist requires the product, mode, model, date, locale, account and whether search was enabled to be recorded on every reading, and null outcomes to be retained rather than dropped.³ None of this stops you measuring. It stops the measurement being asked to carry a claim it was never built for.