Measuring AI visibility:
what can and cannot be measured today

Can you measure how visible a brand is in AI search?

You can measure part of it from first-party reports and estimate the rest. Google Search Console counts impressions in AI Overviews and AI Mode. Bing Webmaster Tools counts citations in Microsoft Copilot and AI summaries in Bing. ChatGPT adds utm_source=chatgpt.com to its referral links. The documentation I reviewed for ChatGPT, Perplexity and Claude describes no owner-facing report. For those, prompt sampling is the remaining method. It gives an estimate that depends on which prompts someone chose to test.

Every factual statement below is drawn from the public documentation listed under Sources.

"How visible are we in AI search?" has become easier to answer in 2026. Google and Bing have both started giving site owners reports dedicated to AI answers. The picture is still far from complete. This piece separates what can be measured from first-party data from what can only be estimated. It also marks what cannot be measured at all today.

What the operators now report

Google Search Console

Google announced Search generative AI performance reports on 3 June 2026. It says they reached all websites worldwide as of 31 August 2026. The help page for the Search report describes what is in them.

• The metric is impressions. That means how many times links to your site were shown to a user in a generative AI feature on Google Search.
• The features covered are AI Overviews and AI Mode. Experiments in Search Labs are excluded.
• Data can be grouped by page, country, device and date. It can be filtered between text-based and multimodal searches.
• In the chart, two links to the same site inside one AI feature count as a single impression, because the chart is aggregated by property.

The help page documents no click, query or position data for this report. Google's announcement says it is considering more metrics over time. Separately, Google's AI features guidance says traffic from these features is included in the main Performance report under the Web search type. That means clicks from AI features are counted there but not broken out.

Bing Webmaster Tools

Microsoft introduced an AI Performance report as a public preview on 10 February 2026. According to that announcement it covers Microsoft Copilot, AI-generated summaries in Bing and selected partner integrations. It shows four things.

• Total citations, meaning how often your pages were displayed as sources in AI-generated answers.
• The average number of unique pages cited per day.
• Citation counts for individual URLs, with a trend over time.
• A sample of grounding queries. These are the key phrases the AI used when retrieving the content it cited.

Microsoft adds a caution that is easy to skip past. The counts show how often a page is cited. They do not show its importance, ranking or placement within an answer.

Referrals and server logs

OpenAI's publisher FAQ says ChatGPT adds the parameter utm_source=chatgpt.com to referral URLs. So visits arriving from ChatGPT search can be identified in an ordinary analytics tool.

Server logs give a second view. OpenAI, Perplexity and Anthropic each document the user agents their systems use. OpenAI and Perplexity publish IP ranges for verifying them. A request from a user-triggered fetcher such as ChatGPT-User or Perplexity-User tells you that a page was fetched in connection with a user's action. It does not tell you the page was cited. It does not tell you what the question was. My guide to AI crawlers explains which agent does which job.

What can only be estimated

The documentation I reviewed for ChatGPT, Perplexity and Claude does not describe an owner-facing report like Google's or Bing's. For those products, the remaining method is to observe from outside. The same goes for any question about a brand being mentioned as opposed to a page being linked. You put a set of prompts to each AI system, record the answers and count mentions and citations. This method is useful. It also has three structural limits that anyone reading such numbers should understand.

Answers vary between runs

Language model output is not fixed. Anthropic's documentation says that even with temperature set to 0, results are not fully deterministic. Identical inputs can produce different outputs. A single run of a prompt is therefore one sample. It is not the answer.

Answers vary between people

OpenAI's help page on web search says ChatGPT typically rewrites a prompt into search queries. It may use a user's approximate location. Saved memories can shape the rewritten query when memory is enabled. A measurement tool asking the same question from a clean account does not see what a long-standing user in another city sees.

Google makes a related point about its own products. AI Overviews and AI Mode may use different models and techniques. So the links they show will vary. AI Overviews often do not trigger at all.

Nobody outside knows the real prompts

None of the sources listed below gives site owners the volume of prompts on a topic. Google's report gives impressions without queries. Bing gives a sample of grounding phrases. These are the system's retrieval queries. They are not the words a person typed. Any "share of voice" figure is therefore a share of the prompt set that someone chose to test. Change the prompts and the number changes.

What cannot be measured today

• The full population of questions people ask. No operator publishes it in the documents I reviewed.
• What a specific person was shown. Personalised and location-dependent answers are private to the conversation.
• Brand mentions without a link, from first-party data. Google's metric counts links to your site and Bing's counts citations. Neither report is described as counting unlinked mentions of a name.
• Clicks from Google's AI features on their own. They sit inside the Web totals.
• Influence with no visit. Someone who reads an answer and acts on it later leaves no referral.

A measurement approach I think is defensible

This section is opinion. I would order the evidence by how close it is to the source.

1. Start with the first-party reports in Search Console and Bing Webmaster Tools. They are limited but they are the operators' own counts.
2. Add referral data, including the ChatGPT parameter. Treat it as a floor. It counts only the visits that arrive with the parameter attached.
3. Use server logs to confirm the right agents can reach the site and to see which pages are being fetched.
4. Use prompt sampling for the questions the first three cannot answer. Read it as a trend. Fix the prompt set. Repeat each prompt. Compare like with like over time. Report ranges instead of single figures.

The mistake I would most want to avoid is treating a sampled visibility score as if it were a census. For a worked example of how much a small prompt set can swing, see my verdict on OtterlyAI's GEO experiments. For background on why the underlying behaviour is so hard to pin down, see how AI answer engines pick and cite sources. Terms are defined in the glossary.

Sources

1. Google Search Central Blog, Introducing Search Generative AI performance reports in Search Console (3 June 2026): developers.google.com/search/blog/2026/06/gen-ai-performance-reports
2. Google Search Console Help, Generative AI performance report (Search): support.google.com/webmasters/answer/16984139
3. Google Search Central, AI features and your website: developers.google.com/search/docs/appearance/ai-features
4. Microsoft Bing Webmaster Blog, Introducing AI Performance in Bing Webmaster Tools public preview (10 February 2026): blogs.bing.com/webmaster/2026/2/Introducing-AI-Performance-in-Bing-Webmaster-Tools-Public-Preview
5. OpenAI Help Center, Publishers and developers FAQ: help.openai.com/en/articles/12627856-publishers-and-developers-faq
6. OpenAI Help Center, Searching the web with ChatGPT: help.openai.com/en/articles/9237897-searching-the-web-with-chatgpt
7. OpenAI, Overview of OpenAI crawlers: developers.openai.com/api/docs/bots
8. Perplexity, Perplexity crawlers: docs.perplexity.ai/docs/resources/perplexity-crawlers
9. Anthropic, Does Anthropic crawl data from the web, and how can site owners block the crawler?: support.claude.com/en/articles/8896518
10. Anthropic, Glossary (temperature): platform.claude.com/docs/en/about-claude/glossary