# How Do You Measure AI Visibility Over Time? > Track AI visibility by asking ChatGPT, Claude, Gemini and Perplexity the same buyer questions weekly, several times each, and count named and cited answers. Published: 2026-10-22 · Updated: 2026-10-23 Source: https://www.itgrows.ai/business/blog/how-to-measure-ai-visibility Publisher: ItGrows (https://www.itgrows.ai/business) — AI visibility for businesses --- Put the same fixed set of buyer questions to ChatGPT, Claude, Gemini and Perplexity every week, ask each question several times, and count how many answers name your business and how many cite your pages. Compare each week only with weeks measured on exactly the same basis, and judge the trend on a four-week average rather than on any single report. That is the whole method. The rest of this page covers the details that decide whether your trend line means anything or just records noise. ## What exactly should you count? Two numbers, kept apart. **Named** means the answer text mentions your business by name. That is what a buyer actually reads. **Cited** means one of your URLs appears among the sources the assistant used. When assistants search the web, they return sources in a structured form. OpenAI's web search tool includes inline citations by default, each as a `url_citation` annotation with the URL, title and location of the source ([OpenAI API documentation](https://developers.openai.com/api/docs/guides/tools-web-search)). Anthropic's web search tool also returns citations for sources drawn from live search results ([Claude Platform Docs](https://platform.claude.com/docs/en/agents-and-tools/tool-use/web-search-tool)). So citations can be counted by a script, not only by eye. The two move independently. You can be named without being cited, because the model knows you from training data or from someone else's page. You can be cited without being named, because your article supplied a general fact. Track both so you notice movement in either. A first citation is worth recording as its own event: in our weekly measurement of 20 October 2026, Perplexity cited an article we had published on a client's site, a German company-formation advisory. Also record who gets cited when you do not. For an orthopaedic practice in Bonn on 20 October 2026, the directory jameda.de was cited in 12 answers and competing practices' own websites were the most cited sources. For questions about getting into AI answers, on 21 October 2026 the most cited sources were the blogs of GEO agencies and tool vendors. That list of cited domains is your map of where the assistants get their material, and it changes week to week too. ## Which questions should you track? A fixed set of around 20, written the way your buyers phrase them, not the way you describe yourself. The research behind the term GEO took the same approach: Aggarwal et al. built a fixed benchmark of queries, GEO-bench, to measure content visibility in generative engine responses systematically ([Princeton University, KDD 2024](https://collaborate.princeton.edu/en/publications/geo-generative-engine-optimization/)). A stable question set is what makes week 12 comparable with week 1. Mix the question types deliberately, and know what each can show you: - **'Who should I hire' and 'how do I choose' questions.** These are where businesses get named. In a 23 September 2026 test of 20 German business questions, firms were named only on this type. For the Bonn practice, 'who to choose' questions named practices; questions about costs, risks and treatments named none. - **Informational questions.** These rarely name anyone, but they are where your articles can be cited. In the same September test the model answered 10 of 20 questions without searching at all and cited nobody; where it did search, it cited primary sources such as laws, ministries, the OECD and large accounting firms on 9 of 10. Once the set is chosen, freeze it. If you want to add questions later, run them as a separate series so the original trend stays clean. ## Why ask each question more than once? Because answers vary. Ask the same question three times and an assistant will often name different businesses. In the September test, two runs of the same question named two different sets of firms. A single ask per question turns your report into a coin toss. Our rule: a business counts as named for a question if any of three asks names it. Report two figures side by side: questions named out of your total (say 4 of 20), and named answers out of all answers (the Bonn practice was named in 12 of 80 answers across the four assistants together). The first tells you how much of the buyer's journey you cover; the second shows how consistently you appear. ## Why measure four assistants separately? They find and use pages differently, and a combined total hides that. On 19 October 2026 we asked assistants for the exact titles of 18 articles, each older than three weeks, on three client sites. ChatGPT found 1 of 18; Claude found 7 of 18. Google had indexed 29 of 30 articles of one of those sites. Being in Google's index did not mean an assistant could find the page. Those 12 of 80 answers for the Bonn practice split as Claude 5 of 20, Gemini 4 of 20, ChatGPT 2 of 20 and Perplexity 1 of 20. Google gives you one official signal: sites appearing in AI Overviews and AI Mode are included in Search Console's Performance report alongside the rest of Search ([Google Search Central](https://developers.google.com/search/docs/appearance/ai-features)). Useful, but it is reported together with ordinary search and covers Google only. It tells you nothing about ChatGPT, Claude or Perplexity, so you still need your own question set. Run the measurement through the APIs with web search switched on, and write down the model and tool version each week. Anthropic alone documents several versioned iterations of its web search tool, such as web_search_20250305, web_search_20260209 and web_search_20260318. A tool change can move your numbers without anything changing on your site, and your notes should show that. ## How do you compare one week with the next? Only on the same basis: same questions, same number of asks, same four assistants, same settings. If any of these changes, the week-over-week difference measures your method, not your visibility. Then smooth it. Take the average of the last four weekly reports and watch that line. With three asks per question, one ask flipping can move a weekly figure, and reacting to a single good or bad week leads to rewriting pages that were fine. The four-week average shows whether a change holds. Annotate the timeline with what you did: articles published, robots.txt edits, sitemap submissions, indexing requests. Without those notes you cannot connect a rise to its cause. And check crawl access in the same report. On our own site, a robots.txt rule meant to hide a single redirect page (/b) was written as a prefix, and from 11 August to 21 October 2026 it also closed /blog, /business and every business article to all crawlers. Bing Webmaster Tools had listed the business page as 'discovered but not crawled' since 11 August 2026. We changed the rule to /b$ on 21 October 2026. On the same day we took a day-0 measurement: 20 English and 20 German questions, four assistants, three asks each. ItGrows was named in 0 of 160 answers and none of our pages was cited. That is our baseline, and the full write-up is in [We Checked Our Own AI Visibility and Were Named in 0 of 160 Answers](https://www.itgrows.ai/business/blog/we-measured-our-own-ai-visibility). A zero next to a blocked page means something different from a zero next to an open one. ## Can traffic or server logs replace this? Not as the main measure. In about two weeks of logs from three of our course websites, roughly 1.2 million requests in the two busiest logs, human visitors arriving from AI assistants totalled one from Copilot and zero from ChatGPT, Perplexity, Gemini, Claude, Meta AI and others. Assistants mostly answer in place. More on that in [Does AI Search Send Traffic to Websites?](https://www.itgrows.ai/business/blog/does-ai-search-send-traffic-to-websites) Logs are still a good secondary signal. Live retrieval agents such as ChatGPT-User, OAI-SearchBot and Perplexity-User fetch a page while an assistant is answering a real person. In the same fortnight an edtech platform received 1,918 such fetches on content pages; a client blog of a few dozen articles received 56 across 24 different articles. Counting these weekly shows which pages assistants actually pull into answers. For business outcomes, ask new leads how they found you. Clients report results such as FreightSizer, a B2B SaaS for freight dimensioning, getting 2 new B2B customers in one month, both from ChatGPT, and appslift, an app store optimisation service, getting 2 to 5 leads a month from AI assistants. Those figures come from the clients, not from a click report. ## What does it cost to measure? Less than most people expect, in API terms. With API prices of October 2026, one question asked once with web search costs about: | Assistant | Cost per ask with web search | Cost for 20 questions x 3 asks | | --- | --- | --- | | Claude | about $0.069 | about $4.14 | | ChatGPT | about $0.010 | about $0.60 | | Gemini | about $0.009 | about $0.54 | | Perplexity | about $0.0013 | about $0.08 | A full report of 20 questions x 4 assistants x 3 asks comes to about $5.40, so four weekly reports cost a little over $20 a month. The real cost is building and maintaining the script: matching name variants, extracting cited domains, storing results so weeks stay comparable, and keeping up when tool versions change. If you want a first reading before building anything, the free check at [itgrows.ai](https://www.itgrows.ai/) generates five buyer questions from your homepage and gets answers from ChatGPT and Perplexity in 25 to 40 seconds. For a manual version, see [How to Check Whether AI Assistants Mention Your Business](https://www.itgrows.ai/business/blog/how-to-check-if-chatgpt-recommends-your-business). ItGrows runs the weekly measurement described here, alongside publishing articles and fixing crawler access, for $499 a month (plus a one-time $499 setup, free with 3 months or a year). Nobody can guarantee that an assistant will name a business, and a measurement only shows you where you stand. ## Frequently Asked Questions **How often should I measure AI visibility?** Weekly is a sensible rhythm. Answers vary between asks, so a daily figure mostly records noise, while a monthly one gives you too few points to see a trend. Judge progress on the four-week average. **How many questions do I need to track?** Around 20 buyer questions, asked three times on each of four assistants, gives a workable report. Include both 'who should I hire' questions, where businesses get named, and informational questions, where your pages can be cited. Keep the set fixed once you start. **Is Google Search Console enough to track AI visibility?** It covers part of it. Google includes AI Overviews and AI Mode in the Performance report alongside the rest of Search, but that says nothing about ChatGPT, Claude or Perplexity, which you have to measure with your own questions. **How long before the numbers start to move?** There is no fixed timeline, and pages first have to be crawlable and findable by each assistant. Named and cited answers move independently, so track both to catch movement in either. See [How Long Does It Take to Start Appearing in ChatGPT Answers?](https://www.itgrows.ai/business/blog/how-long-does-it-take-to-appear-in-chatgpt-answers) for what to expect.