Direct answer: A citation shows that a supported AI surface displayed a link to your page; a referred session shows that a visit was recorded with a source; a qualified action shows that a visitor did something useful. Report these separately. None alone proves that the AI answer caused a sale.
AI visibility dashboards now make a familiar reporting mistake easy: a page was cited, so a report calls that “traffic.” The units are different. Bing Webmaster Tools' AI Performance preview reports citations displayed as sources in supported AI answers and page-level citation counts. Bing explicitly says the counts do not indicate position, ranking, authority, or the role of a page in a particular answer. Its grounding queries are a sample, not a complete list of user searches.
Google's Search Generative AI report uses another unit: impressions when a site URL appears in AI features on Search or Discover, with page, country, device, and date views. Google says this is included in overall performance data. A Google AI impression and a Bing citation should not be added into a single “AI mentions” total. The platforms count different events on different surfaces.
Keep a three-stage ledger
| Stage | Record | What it does not prove |
|---|---|---|
| Appearance | AI-feature impression or displayed citation, by platform and page | That someone clicked or read the cited page |
| Arrival | Sessions and landing pages with a relevant source or medium | That every AI-influenced visitor was recognized |
| Action | Qualified inquiry, booking, purchase, or other agreed outcome | That AI exposure alone caused the action |
For arrivals, use session-level acquisition data rather than a first-user acquisition report. Google Analytics explains that its Traffic acquisition report is scoped to sessions, whereas User acquisition is scoped to new users. Referral information can also be missing: GA4's direct-traffic documentation lists lost tracking parameters, redirects, offline documents, and blockers among the reasons traffic may lack a clear source. Therefore, a small AI referral count is a lower-confidence observation of recognized visits, not a full count of everyone influenced by an answer.
Two research papers show why the first stage is a limited proxy. Liu, Zhang, and Liang's 2023 human audit found that citations in four then-current generative search engines did not always support the sentence beside them. Strauss and colleagues' 2026 analysis compared retrieved URLs with displayed citations in roughly 14,000 LMArena interaction logs from March–April 2025 and found an attribution gap in several tested model configurations. The first paper asks whether a citation supports a claim; the second asks whether a retrieved source was cited. Both concern specific historical systems and samples; the second also relies on provider-disclosed logs that may omit or overcount pages used. Neither supplies a current conversion rate for your site. Together they argue for auditing answers and measuring downstream visits separately.
Start with a monthly page-level sheet: platform, visibility metric and definition, recognized sessions to that landing page, and qualified actions from those sessions. Add a few manually reviewed answer examples, recording the date and whether the link actually supports the claim. If the sample is tiny, report the counts plainly and avoid percentage swings. Compare the pages that appear with the pages that receive useful visits; that gap is a question to investigate, not evidence that AI search is working or failing.
The existing campaign attribution guide addresses causal claims about leads. This article defines the narrower AI-search measurement chain. ARCHV can design the landing and inquiry path for a scoped website; analytics instrumentation and attribution analysis should be agreed separately. A website estimate can start with the pages and actions the business actually values.