Measuring AI search visibility

AI search visibility can be measured in parts, and no single number covers all of it. As of October 2026, Google and Microsoft report on their own AI features, analytics tools catch some of the visits that AI answers send, server logs show which bots fetched which pages, and prompt tracking samples the answers directly. Each source sees a different slice and has its own blind spots, so a useful report says which source each number came from.

Reports from Google and Microsoft

Google's Search Console has a Generative AI performance report, which Google says reached all websites on August 31, 2026. It shows impressions from AI Overviews and AI Mode, broken out by page, country, device, and date, where an impression means a link to the site was shown in one of those features. The help page describes impressions only. Google separately says that traffic from its AI features counts toward the regular Performance report under the Web search type.

Microsoft added AI Performance to Bing Webmaster Tools as a public preview in February 2026. It counts citations of a site's pages across Microsoft Copilot, AI-generated summaries in Bing, and some partner integrations, with page-level counts, a trend over time, and a sample of the grounding queries the AI used when it retrieved the cited content. Microsoft notes that these counts don't indicate ranking or placement inside an answer. We aren't aware of a comparable publisher report from OpenAI, Anthropic, or Perplexity as of this writing.

Referral visits in analytics

When someone clicks a link in an AI answer, the visit can show up in analytics as a referral from a domain like chatgpt.com, perplexity.ai, copilot.microsoft.com, gemini.google.com, or claude.ai. Links in ChatGPT answers often carry a utm_source=chatgpt.com parameter, which helps when the referrer is missing. Referrers do go missing, because many apps don't pass one, and those visits land in direct traffic. AI referral counts are a floor, and the real number may be well above it.

Clicks also understate influence. The Pew Research Center tracked the Google searches of 900 US adults in March 2025 and found they clicked a regular result on 8 percent of visits to results pages with an AI summary, against 15 percent of visits without one, and clicked a link inside the summary on 1 percent. A page can shape an answer that few readers click through from, which is why impressions and citations are worth tracking alongside visits.

Bot activity in server logs

Server and CDN logs record visits from AI crawlers and user-triggered fetchers, identified by user agent. User agents are easy to fake, so verify them against the IP ranges the providers publish. OpenAI and Perplexity offer JSON lists for each bot, and Apple supports reverse DNS lookups for Applebot. A verified fetch by ChatGPT-User or Perplexity-User is one of the most direct signs available that someone's question led an assistant to a particular page.

Logs have two large gaps. Engines often answer from pages they indexed earlier, so a citation can happen with no visit at the moment of the question. Google's AI features rely on ordinary Googlebot crawling, so nothing in a log separates an AI Overview from a regular search result. Most crawlers also never run the analytics script on a page, which leaves logs and CDN data as the only place they appear. The crawler and agent access guide lists the user agents to look for.

Prompt tracking and its margin of error

Prompt tracking means sending a fixed set of questions to AI engines on a schedule and recording which brands and sources appear. It's the only way to sample answers directly, and its limits are worth stating up front. Answers vary from run to run, and a 2026 study by Ronald Sielinski estimated that a domain's citation share on OpenAI's search model needs 150 or more queries before its 95 percent confidence interval narrows to five percentage points. A three-point change measured on a few dozen questions can easily be noise.

Method matters as much as volume. What a provider's API returns can differ from what its consumer app shows, and the app personalizes answers by approximate location and, with memory turned on, by saved memories. A sound setup keeps the question set fixed, runs each question more than once, records the date, engine, and access method for every response, and reports shares with a range. That's the approach behind the monitoring described on our advisory page, and the reasons answers shift in the first place are covered in how AI search engines choose and cite sources.

What stays out of view

Some things can't be measured from outside today. None of the engine reports we've seen show how often an assistant mentions a brand without linking to it, how training data shapes what a model says, or how many people saw an answer citing a page inside ChatGPT, Claude, or Perplexity. Google and Microsoft report only on their own surfaces. Any tool that presents a single, complete AI visibility score is combining partial sources like these, and it should say how.