Methodology
How to Measure Whether AI Assistants Cite Your Site
No tool reports this end to end. Three separate measurements exist, each counting something different, and the gap between them is where most of the confusion lives. Here is what each one actually measures, what it misses, and the numbers from our own 21-site portfolio.
There is no analytics product that tells you how often an AI assistant used your page to answer someone. What exists is three partial measurements: Bing Webmaster's AI Performance report, which is the only first-party citation count any major platform publishes; an analytics referrer segment, which counts the minority of citations that produce a click; and server logs, which show crawlers fetching pages and tell you nothing about whether the content was used. Used together they bracket the truth. Used alone, each one is misleading in a specific and predictable direction.
We publish this because we had to work it out ourselves, and because the gap between the measurements turned out to be the most important number in our own operation. Over an identical 28-day window our sites received 67 clicks from Google and were cited 14,394 times by Microsoft Copilot and partner models. Those two numbers describe the same 21 websites in the same month. If we had been watching only one dashboard, we would have concluded something completely wrong about what we had built.
Measurement 1: Bing Webmaster AI Performance (citations)
This is the closest thing to a direct citation count that exists. In Bing Webmaster Tools, a verified site gets an AI Performance report covering citations by Microsoft Copilot and its partner models. It gives you a daily time series, a total, and two tables that matter: the grounding queries that pulled your content into an answer, and your pages ranked by citation count.
Three practical notes from using it weekly:
- Verify the root domain, not each subdomain. The report aggregates across subdomains, so one read covers an entire portfolio. Ours returns 179 distinct cited pages across 21 subdomains in a single view.
- There is no API. A sweep of candidate Bing Webmaster API endpoint names in May 2026 returned 404 for every one. The report is interface-only, which means a human or a browser-driving agent has to read it. Budget for that; it is the reason most teams never look.
- The grounding-query table is the actionable half. Citation totals are a scoreboard. The queries tell you which questions your content is being pulled into, and the citation-share column tells you how much of that answer you own. That is a content brief, not a vanity metric.
What it misses: everything outside Microsoft's ecosystem. It says nothing about ChatGPT, Claude, Gemini or Perplexity. Treating it as total AI visibility is the mistake it invites.
Measurement 2: Analytics referrers (click-through)
Every major assistant passes a referrer when a user clicks a link in an answer. Segmenting analytics by those hosts gives you assistant-referred sessions. The practical host list:
chatgpt.com · chat.openai.com · perplexity.ai · claude.ai · gemini.google.com · copilot.microsoft.com · you.com · phind.com
Two caveats decide whether your number means anything. First, some assistant traffic arrives with no referrer and lands in Direct, so this is a floor, not a total. Second, GA4's built-in “AI Assistant” channel group and your own referrer segment will not match — ours disagree by a few percent because they use different host lists. Neither is wrong. But if you switch between them mid-quarter, your trend line is measuring your methodology. Pick one, write it down, keep it.
What it misses: the large majority of citations. A citation that answers the question in the chat produces no click and leaves no trace here at all.
Measurement 3: Server logs (crawling, not citing)
Your logs show assistant crawlers fetching pages. This is genuinely useful for one question — can they even reach my content? — and useless for the question people actually ask it. A fetch is not a citation. A page can be crawled daily and never used in a single answer.
Use logs to verify access and nothing more: that your robots rules permit the crawlers you want, that important pages return 200 to them, and that JavaScript-dependent content is not invisible in the raw HTML.
The Three Together
| Measurement | Counts | Blind to | Our 28-day figure |
|---|---|---|---|
| Bing AI Performance | Citations by Copilot + partner models | Everything non-Microsoft | 14,394 citations, 179 pages |
| Analytics referrers | Sessions from assistant links | Citations that produce no click | 126 sessions |
| Server logs | Crawler fetches | Whether content was used at all | Access verification only |
| For contrast: Search Console | Google clicks and impressions | All assistant surfaces | 67 clicks / 38,115 impressions |
The ratio between rows one and two is the finding. Roughly a hundred citations for every assistant-referred session means the overwhelming majority of AI visibility never touches your analytics. Any organisation judging its AI presence by session counts is reading the smallest of the available numbers.
Two Things We Did Not Expect
Citations look like a stock, not a flow. Our most-cited property has published nothing since early July 2026 — it is on maintenance only — and it remains the largest single source of citations in the portfolio, holding 4,827 of our top-25 citation count. Rankings decay when you stop publishing. Citations, so far, have not. We are re-measuring at 60 and 90 days before believing it, but if it holds it changes what a content budget is buying.
The pages that get cited are not the pages built to rank. Our most-cited content answers a question a person actually asks out loud, puts a direct answer in the first line, and cites a primary source next to each claim. Pages engineered for search rankings do markedly worse on this surface. Evidence-shaped, not SEO-shaped.
A Minimum Viable Setup
- Verify your root domain in Bing Webmaster Tools and read the AI Performance report weekly at the 3-month window. Save the totals, the per-page table and the grounding-query table to a dated file — the interface shows you the present, and you will want the trend.
- Build one assistant-referrer segment in your analytics with the host list above. Write the definition down. Do not change it.
- Check your logs once for crawler access, then stop looking at them for this purpose.
- Report the citation count and the session count side by side, over identical windows. Separately they mislead; together they are the actual picture.
Frequently Asked Questions
Can you track when ChatGPT or Copilot cites your website?
Partially, and only through three separate measurements that no single tool combines. Bing Webmaster Tools has an AI Performance report that counts citations of your pages by Microsoft Copilot and its partner models, including a per-page and per-query breakdown — it is the only first-party citation count any major platform publishes. GA4 (or any analytics tool) can be segmented by referrer to count sessions arriving from assistant domains such as chatgpt.com and perplexity.ai, which measures click-through rather than citation. Server logs show the AI crawlers fetching your pages, which measures neither. Citations without a click are invisible to analytics entirely, which is why the Bing report matters out of proportion to Bing's search share.
Why don't AI citations show up in Google Analytics?
Because a citation is not a visit. When an assistant reads your page and summarises it in an answer, the user often never clicks — they got what they needed in the chat. Nothing reaches your analytics. GA4 only records the minority of cases where someone follows the link. In our own portfolio the gap between the two is roughly two orders of magnitude: 14,394 citations recorded by Microsoft against 126 assistant-referred sessions in GA4 over the same window. Judging AI visibility by analytics sessions alone therefore undercounts it dramatically.
Where is the AI Performance report in Bing Webmaster Tools?
Under the AI Performance section for a verified site, with 7-day, 30-day, 3-month and 6-month windows. It reports total citations, average cited pages per day, a daily time series, and two tables: one listing the grounding queries that pulled your content into an answer, and one listing your pages by citation count. There is no API for it — an endpoint sweep in May 2026 found no citation endpoint on the Bing Webmaster API — so it must be read from the interface. If you verify your root domain, the report aggregates across all subdomains, which makes one read cover an entire portfolio.
Which analytics referrers indicate AI-assistant traffic?
The practical list is chatgpt.com, chat.openai.com, perplexity.ai, claude.ai, gemini.google.com, copilot.microsoft.com, you.com and phind.com. Two caveats matter. Some assistant traffic arrives without a referrer at all and lands in Direct, so this is a floor rather than a total. And GA4's own 'AI Assistant' default channel group and a hand-built referrer segment will not agree exactly — ours differ by a few percent — because they use different host lists. Pick one definition, write it down, and keep it stable, or your trend line will measure your methodology instead of your traffic.
Do AI citations decay the way rankings do?
Our data suggests less than expected, though the observation is young. Our most-cited property has published nothing new since early July 2026 and remains the largest single source of citations in the portfolio, holding 4,827 of the top-25 citation count. That looks more like a stock than a flow — a ranking decays when you stop publishing, while a citation appears to keep being earned by the page that already answers the question. We are treating this as provisional and re-measuring at 60 and 90 days rather than building a strategy on one reading.
Method and Caveats
Figures come from our own 21-site portfolio over the 28 days ending 24 July 2026: Bing Webmaster AI Performance (citations), GA4 referrer segmentation (sessions) and Google Search Console (clicks and impressions). Bing states that its AI Performance data is a sample of overall activity and may be refined as more data is processed, so treat the totals as indicative of magnitude and trend rather than exact. The citation-durability observation rests on a few weeks of data on one property and is the least reliable claim here — we flag it as provisional deliberately. Our full operating numbers, including the experiments that failed, are published in the Thicket Report.