AI search measurement glossary

Terms used across this site, defined precisely enough to be argued with. Several of them are used loosely elsewhere in the category, which is how a measurement conversation goes wrong before it starts.

Consideration set: The group of companies an AI assistant is willing to name in answer to a given question. Membership is probabilistic: a company appears in some proportion of runs rather than occupying a fixed place.; This is why presence is reported as a share of runs with a sample count. Two companies at eleven of twenty and nine of twenty are not meaningfully apart.
Referrer stripping: A page instructing the browser not to send the referring page when a link is clicked, usually through rel="noreferrer" or a referrer policy. The visit happens; the destination never learns where it came from.; Affected visits land in direct traffic. No tool recovers the connection afterwards, because the information was never transmitted.
Direct traffic: Sessions arriving with no identifiable source. A mixed bucket containing people who typed your address, clicks from desktop email clients, visits from apps that pass no referrer, and AI referrals whose referring page was removed.
AI crawler: An automated client fetching your pages on behalf of an AI system. Most operators run more than one, separating training corpora, the search index behind answers, and fetches triggered by a person asking about a page in the moment.; Blocking them is three separate decisions rather than one, and a single wildcard rule usually produces an outcome nobody wanted.
AI Overviews: Google's generated summaries appearing above ordinary search results, which cite the pages they draw on.; Search Console counts these impressions inside your ordinary totals and does not label them, so no filter isolates them.
Crawl-to-click: The relationship between how often AI crawlers fetch a page and how many human visits that page receives from AI assistants. The two are related far more weakly than intuition suggests.; A heavily crawled page can send nobody. Treat crawl activity as attention rather than as outcome.
First-party data: Data collected by your own systems about your own site: analytics, server logs, CDN records, search console. Distinct from data a third party collects by polling AI models.
Share of answers: The proportion of repeated runs of the same question in which a company was named, reported with the number of runs behind it.; The honest alternative to a position, which implies a stability that does not survive asking again.
Key event: Google Analytics 4's term for what was previously called a conversion. An event marked as one of the outcomes that matter to your business.; Older guides referencing a conversions metric are describing the same thing under its previous name.
Session source: The origin of the session in which an event occurred, as distinct from first user source, which records where that person originally came from days or weeks earlier.; Building a channel on the wrong one produces plausible numbers answering a different question, usually a more flattering one.
Attribution window: The period during which a touchpoint can receive credit for a later conversion. Changing it changes the numbers without anything changing in the business.
Measured and estimated: The two confidence labels Veribix puts on every referral figure. Measured means counted from data that identified its source; estimated means inferred, with the reason stated.; Estimated is not hedging. It is a statement that the number was arrived at differently, and any surface showing one has to say so.

How these terms are applied in practice is set out in how Veribix measures.