How to measure AI search visibility

Why questions about your company name aren’t enough: Whether an AI system knows your name and whether it mentions you when someone asks a general question are two different things. In a study of 112 startups from the Product Hunt leaderboard, ChatGPT recognized 99.4% of the products when asked about them by name, and Perplexity 94.3%. Only 3.32% and 8.29% of the products, respectively, came up for general questions such as “What are the best AI tools launched this year?” The study queried both providers through their programming interfaces (APIs), not through their apps. Our guide to generative engine optimization (GEO) explains how these systems choose their sources.

Source: Sharma (preprint) (external site), Jan 1, 2026 · 112 startups, 2,240 queries

Why use several wordings: People ask for the same thing in very different ways. In an experiment by SparkToro, people wrote 142 prompts about the same purchase, and barely two of them were alike. Even so, the four most frequent brands each appeared in 55–77% of the 994 answers. So a fixed panel can’t cover every wording, but it does show whether you’re among the providers that get named. Even small edits to the wording change the sources of Google’s AI Overviews, as a study from the SIGIR 2026 conference (external site) shows.

Source: SparkToro (external site), Jan 27, 2026 · vendor study · 142 human-written prompts, 994 answers

How to do it:

1. Collect questions from inquiries, emails and first meetings. Add question-style queries from the Google Search Console performance report (external site) and search terms from Bing Webmaster Tools (step 5).
2. Assign each question to a group: company (“What does [company] offer?”), provider (“Who offers [service] in [place]?”) or expert question (“How much does [service] cost?”).
3. Phrase your most important questions in two or three different ways. A survey of GEO research (external site) recommends three to five paraphrases per information need for studies.
4. Give each question an ID, such as Q01, and a target page on your website. After that, never change the wording; a new wording gets a new ID.
5. Start with 10 to 20 questions; in our view, that’s enough to begin with. With every additional question, the share of answers that mention you becomes more reliable: SparkToro considers it a reasonable metric across dozens to hundreds of prompts, each run several times.

Example: Our own panel includes this expert question, phrased in German: “How can I measure my company’s AI visibility?” Its target page is the German version of this how-to.

Common mistake: Your panel only contains questions that include your company name. It then shows whether AI systems know you, but not whether they mention you to potential buyers.

Done when your list has 10 to 20 questions, each with an ID, a group, fixed wording and a target page, and all three groups are represented.

Why each system separately: Profound tracked the citations of 883,000 pages in seven AI systems over 12 months. Gains or losses in one system said nothing about the others. Profound itself sells a tool for this kind of measurement.

Source: Profound (external site), Sep 30, 2026 · vendor study · 883,000 pages, seven AI systems

Why no personal history: Being signed in changes the answers. According to OpenAI, ChatGPT draws on past conversations as long as memory (external site) is turned on; a temporary chat doesn’t use memory. Since February 2025, ChatGPT search has also worked without an account (external site). Gemini (external site) can personalize answers for signed-in users based on memory of their past chats. Google says Search results depend on the time, place, device and recent history of the person searching (Search Console Help on the Performance report (external site)).

Why the access method matters: Some tools query AI systems through their APIs. With ChatGPT, the answers there differ from those in the app. In one study, OpenAI’s API with web search and Google Search together found only about a quarter of the pages that ChatGPT cited in the browser. They found about half of the websites behind those pages.

Source: Jiang, Zhang, Xing, Chang, Xiao (preprint) (external site), Oct 11, 2026 · 80 questions, 20 runs each

How to do it:

1. Choose two to four systems your customers use; if in doubt, ask new contacts (step 6). Candidates include ChatGPT, Google’s AI Overviews and AI Mode, Microsoft Copilot, Perplexity, Gemini and Claude.
2. Ask while signed out, in a private browser window or a temporary chat. Use a new chat for each question, with no follow-up questions.
3. Copy and paste the exact wording. Ask from a German location, without a VPN that places you elsewhere. Query all systems on the same day, in the same week every month.
4. For each answer, note the system, the access method (app, browser or API), the model shown, whether the system searched the web, the answer text and all sources shown. A screenshot helps if questions come up later.

Example: For our own measurement, we plan to use APIs to query Claude with web search, Perplexity and Google’s AI Mode, as well as ChatGPT, Gemini and Google’s AI Overviews. We’ll also check ChatGPT and Microsoft Copilot by hand in their regular interfaces as a spot check, signed out and from a German location. Our methodology has the details.

Common mistake: You measure in your own account, where you’ve already asked about your company many times. The answers may then be based on your earlier chats.

Done when the header of your tracking sheet lists the access method, sign-in status, language, location and measurement day for each system.

Why: AI answers vary even when nothing has changed. In this study, about two-thirds of the cited sources changed from one day to the next on average, and questions asked in quick succession varied almost as much. The study covered ChatGPT, Gemini, Google’s AI Mode and Perplexity, with questions from four industries, from January to March 2026. Our GEO guide summarizes other measurements of this fluctuation.

Source: Schulte, Bleeker, Kaufmann (preprint) (external site), Apr 8, 2026 · four AI systems, queried from servers in Switzerland

The table, from the same study, shows how precise the mention rate for a single question is after a given number of runs. With three runs, a question whose answers mention you half the time on average can therefore come out anywhere between about 13% and 87%. If you’re always or never mentioned, the authors say fewer runs are enough.

Source: Schulte, Bleeker, Kaufmann (preprint) (external site), Apr 8, 2026 · Appendix J, Table 16, mean of 1,216 brand series

How to do it:

1. Ask each question at least three times per system. The runs can follow each other directly: according to the study, they vary almost as much as runs on different days anyway.
2. With three runs, analyze groups of questions, such as all provider questions. Draw conclusions about a single question only once you have seven runs or more.
3. If you measure daily, for example with a tool, combine two to four weeks. The authors recommend this to get stable values for individual brands.
4. With fewer than 20 answers, report the count, such as “3 of 9 answers,” instead of a percentage.

Example: Our worked example under “From our work” shows how precise a rate across many answers is.

Common mistake: Each question runs once, and the result triggers an action for that one question.

Done when every question has at least three timestamped runs per system in your tracking sheet. Your report only draws conclusions about single questions with seven or more runs.

What to watch for when counting:

• All answers in the denominator: In the study with German-language prompts, ChatGPT showed at least one source in only 42.2% of runs. For questions about definitions, it often didn’t search at all (Schulte et al. (external site)). According to the survey (external site), answers without a search or without sources are “outcomes, not data to be discarded.” Leave them out, and your citation rate comes out too high.
• Domain first, then page: The specific page that gets cited changes even more often than the website (SISTRIX (external site)). Repeated runs in ChatGPT often retrieved different pages from the same websites (Jiang et al. (external site)). So count the citation for your domain, and note the page as well.
• Position as a range: In SparkToro’s research, the same brand list in the same order came up less than once in 1,000 runs. For SparkToro, a “ranking position in AI” is therefore nonsense. Report the range or the median across all answers with a list.
• No single score: Microsoft says the new metrics in Bing Webmaster Tools “do not turn AI visibility into a single ranking or score” (Microsoft Bing, Jun 16, 2026 (external site)). If a tool shows you an AI visibility score from 0 to 100, ask for the numerator and denominator.
• Define your core facts in advance: According to Bitkom, 42% of those who use AI for search say they’ve already received false or made-up information from an AI. Before your first measurement, decide which facts must be right, such as your name, location, services and website.

Source: Bitkom (in German) (external site), Nov 20, 2025 · 1,156 respondents aged 16+

Our how-to on getting recommended by ChatGPT shows how to sort cited websites by type and what to do with the result.

Common mistake: Only linked citations get counted, and mentions without a link drop out. Missing links are common: according to Bitkom, only just over a third of respondents (36%) feel that AI backs up its answers with enough links.

Done when your analysis gives the numerator and denominator for every metric and includes answers without sources.

Source: Google Search Console Help (external site), Aug 31, 2026 · generative AI performance report

Source: Microsoft Bing (external site), Feb 10, 2026 · AI Performance, expanded Jun 16 and Oct 13, 2026

What an impression means: Google generally counts an impression as soon as a link is on the current results page, even if nobody scrolls down to it. All links in an AI Overview share a single position (Search Console Help (external site)). So an impression doesn’t mean that anyone has read your page.

What Bing’s grounding queries mean: Grounding queries are the search phrases the AI system used to retrieve content. You won’t see the question a person actually asked. Google’s AI Mode also breaks a question down into subtopics and searches for each one. Our article on query fan-out explains more. Since October 2026, Bing’s Topics Portfolio (beta) has also shown queries where other websites are cited and yours isn’t. According to Microsoft, smaller websites may not see this view at all (Microsoft Bing, Oct 13, 2026 (external site)).

What’s missing: As of November 2026, we found no comparable report for website owners from the providers of ChatGPT, Perplexity, Claude or the Gemini app. For these systems, your question panel remains the most important source. Our article on AI Overviews and AI Mode explains how Google chooses the sources for its AI answers.

How to do it:

1. Add your website to Google Search Console and Bing Webmaster Tools if you haven’t already.
2. Make sure your site is included in generative AI features in Search Console; otherwise the report stays empty. The first step of our ChatGPT how-to has more on this.
3. Export both reports on your measurement day and file them with your tracking sheet.
4. Add queries that keep showing up in Bing to your list of candidates for new questions.

Common mistake: Impressions from AI Overviews are read as visits.

Done when both reports are set up and you file their figures every month on your measurement day.

• Web analytics: Google Analytics has its own “AI Assistant” channel for visits from services such as ChatGPT, Gemini, DeepSeek, Copilot or Grok. According to Google (external site), this channel explicitly excludes visits from Google’s AI Overviews and AI Mode.
• Server logs: When an AI assistant fetches your page to answer a user’s question, the request shows up in your logs. The user agents for this are ChatGPT-User at OpenAI (external site), Claude-User at Anthropic (external site) and Perplexity-User at Perplexity (external site). Such a fetch shows that a system opened your page for an answer. It doesn’t show whether the system cited it. Our how-to How to track AI traffic beyond GA4 shows step by step how to analyze visits and fetches in your logs.
• Authenticity: A user agent can be faked; Google, for example, explains how to verify requests that claim to come from Google (Google for Developers (external site)). OpenAI and Perplexity publish the IP addresses they fetch from, so check the fetches against those lists. Our article on AI crawlers covers which crawlers exist and what they’re for.
• Asking new contacts: Ask them how they found you. That way, you also hear about recommendations in AI answers that nobody clicked.

Why rising numbers call for caution: AI services are growing themselves. On one large website, total referrals from ChatGPT grew 5.7-fold, but pages on the same site that hadn’t been changed still grew 3.5-fold. The authors call the effect of their changes “suggestive, not conclusive.” So always compare the pages you’ve worked on with pages you haven’t changed.

Source: Watanabe, Nakayashiki (preprint) (external site), Jun 3, 2026 · one website, web analytics and server logs

Example: Our contact form asks an optional question: “How did you hear about us?” The choices include ChatGPT, Perplexity, Microsoft Copilot, Gemini, Claude and “Google AI Overviews or AI Mode.”

Common mistake: Crawler fetches get counted as visits, or a rise in referrals is credited to your own work without a comparison with unchanged pages.

Done when you see visits from AI assistants separately, can tell user-initiated fetches apart from other crawlers and ask new contacts how they found you.

Why: A new model can swap the sources overnight. After ChatGPT switched from GPT-5 mini to GPT-5.5 on May 23, 2026, the domains cited in German-language answers shifted by 47% within 48 hours, according to SISTRIX. In the days before, the shift was 1–2%. SISTRIX itself calls this a correlation, and it sells a tool for this kind of measurement.

Source: SISTRIX, Jun 1, 2026

How to do it:

1. Keep a change log with the date, system, event and source. Google reports major Search updates in the Google Search Status Dashboard (external site), such as the core update from May 21 to June 2, 2026. Add model changes as soon as a provider announces them.
2. Calculate the monthly value as the mean of the run values. The range runs from the lowest to the highest run value.
3. Only treat a change as a trend if the monthly value falls outside the previous month’s range and stays there the following month. Our methodology applies the same rule.
4. If a model change falls between two measurements, report the values separately and restart the trend rule.
5. When you change something yourself, note the date and the pages affected. Then compare questions whose target page you changed with questions where nothing changed.

Worked example: Your three runs in February produce mention rates of 13%, 23% and 17%. The monthly value is 18%, and the range runs from 13% to 23%. A March value of 20% is within the range, so it’s a fluctuation. If March comes in at 30% and April also stays above 23%, treat that as a trend.

Common mistake: A good month gets celebrated, and a bad one sets off hasty changes.

Done when your change log is running and every monthly report says whether a change is a fluctuation or a trend.

Questions to ask about a tracking tool:

1. Which systems does it query, and through which access method: app, browser or API?
2. Does it ask in German and from a German location?
3. Does it use your own questions, or does it generate questions itself? Can you see every question?
4. How many runs does it do per question, and does it show the variation?
5. Can you see the answer text and all sources of every answer, and can you export the raw data?
6. Does it count answers without sources?
7. Does it flag model changes?
8. How is an overall score calculated, and what’s in the denominator?

Why be wary of promises: Google advises site owners to “be wary of third-party tools that promise ranking success” or claim to use internal Google metrics. According to Google, no third-party tool has access to its internal ranking or AI systems (Google Search Central, as of Jul 10, 2026 (external site)). Our how-to on choosing a GEO agency lists the questions to ask an agency about its measurement.

Example: We’ll run our own panel with a question list and one row per answer. The template is below, under “From our work.”

Common mistake: A tool gets bought for its score, without anyone knowing which questions and runs are behind it.

Done when you can answer all eight questions for your tool, or you measure by hand with the template.

• No view inside the system: You see answers, but not the reasons behind them. The survey describes the path from search to answer as subject to chance and only partly observable.
• A citation doesn’t prove a correct statement: The same paper notes that “a visible citation may be incorrect.” So for questions about your company, read what the answer actually says.
• Visibility has many levels: The research literature uses the term “visibility” for at least nine different quantities, from whether a search is triggered to clicks and revenue. A mention isn’t a click yet, let alone an order.
• Many studies come from vendors: Several of the analyses cited here come from companies that sell measurement tools, and others are preprints that haven’t been peer-reviewed. We say so each time; for example, one of the authors of the SparkToro study works for an AI tracking vendor.

Source: Martinez (preprint) (external site), Jul 15, 2026 · survey of 45 studies

What our own measurement can’t do is set out in our methodology under What this measurement isn’t.

Common mistake: A value from your panel is presented as a market share in ChatGPT.

Done when your report states the period, systems, number of answers and limits of the measurement for every figure.

• Countries differ: According to SISTRIX, 74% of the domains ChatGPT cited in Germany were new each week, compared with 60% in the UK and 42% in France. SISTRIX measured from December 2025 to April 2026. In another study, the same English questions produced an AI Overview less often from Germany than from the US: for 65% instead of 81% of the questions (Kirsten et al. (external site)).

  Source: SISTRIX (external site), May 1, 2026, updated May 12, 2026 · 82,619 prompts, six countries, 17 weeks

• Where the studies ran: The study with German-language prompts collected its data from servers in Switzerland. The authors themselves write that the location can affect which sources get chosen.

• Language of the sources: Note whether German- or English-language pages get cited. The section on Germany and the EU in our ChatGPT how-to explains why this matters.

• Server logs: When you analyze them, check IP addresses against the providers’ lists, then store only the result, such as “ChatGPT user-initiated fetch.” That way, your analysis needs very little data. Our own website sets no cookies, and we’ll count fetches from our server logs without IP addresses (our measurement principles).

Done when your tracking sheet states the language and location of every query and the language of the cited sources for every answer.