# How to track AI visibility: a weekly measurement method you can run yourself > A do-it-yourself weekly method: build a question panel, run repeats per engine, log mentions and citations, and read trends without overreacting. Source: https://www.fame.so/post/track-ai-visibility · Fame Team · November 3, 2026 ## Key takeaways - Use a fixed panel of 20 to 50 buyer questions, mostly without your brand name. - Run each question several times per engine, because citations churn heavily from day to day. - Log mentions and citations as separate metrics, and note answers with no citations. - Do not change a page after one miss. Ahrefs suggests reassessing after two to three weeks. To track AI visibility yourself, fix a panel of 20 to 50 buyer questions, run each several times a week in ChatGPT, Perplexity, Claude and Gemini, and log mentions, citations and competitors in a spreadsheet. Report each engine separately and read multi-week trends. Single checks mislead, because AI answers and their cited sources change constantly. ## Why not just check once a month? Because one answer is a sample, not a result. [Ahrefs asked five AI tools the same questions daily for 20 days](https://ahrefs.com/blog/how-stable-are-ai-citations/), tracking about 37,500 citations, and found: EngineCitations gone by day twoLasted all 20 daysAnswers with no citationsGoogle AI Overviews50%6.6%0.9%Google AI Mode58%6.8%0.5%ChatGPT59%4.0%0.7%Gemini81%0.2%64%Copilot83%0.5%53% Ahrefs notes the Gemini and Copilot churn is inflated by empty answers. Excluding those, about 58% disappeared. The study also found about 75% of citations that dropped out came back within the period. Ahrefs sells an AI tracking product and the data covers one month, so treat the exact figures as a guide. The point holds anyway: a monthly screenshot can show either a win or a loss by chance. Engines also disagree with each other. [Analyze AI's vendor study](https://www.tryanalyze.ai/blog/state-of-ai-search-in-b2b-2026) found the brand missing from all three engines 47.2% of the time on generic prompts, and present in all three only 14.6% of the time. So report each engine separately. ## How do you build the question panel? Write the questions a buyer asks, using their words. Pull from sales calls, support tickets, reviews and Search Console queries. Then sort them into these groups: GroupExampleWhat it tells youCategory discoveryBest SOC 2 automation tools for a 100-person startupWhether new buyers find youProblemHow do we prepare for a SOC 2 audit in 90 days?Whether you are a source for the problemComparisonAcme vs a named rival for startupsHow you are framed against rivalsFact-checkWhat does Acme cost?Whether the facts about you are right Keep most of the panel unbranded. Analyze found that naming the brand in a prompt adds about 60 percentage points to mention rates, and category alone moves rates from 9% to 70%. A panel full of branded prompts will flatter you, so report prompted and unprompted results separately. Publish the panel internally, give it a version number, and avoid editing questions mid-quarter. If you must change one, mark it as new so the trend line stays honest. Ahrefs also found brand-question citations were most stable, and "best X" questions were among the least, so results by group will behave very differently. ## How do you run it each week? Set a fixed time, ideally the same day each week, and follow this routine: - Open a fresh chat with no memory or custom instructions, and use the same plan each time. Record which one. - Run each question three times per engine. The number is a practical compromise, not a standard. More repeats give steadier numbers. - Save the full answer text and every cited URL. - Fill in one spreadsheet row per run. - Calculate the summary numbers, then write three sentences on what changed. Use these spreadsheet columns: date, engine, model or plan, question ID, question group, run number, any citations at all (yes or no), brand named (yes or no), own site cited (yes or no), cited URL, competitors named, cited third-party domains, factual errors, and notes. Fifty questions across four engines at three runs each is 600 answers, which is too many by hand. A one-person team should begin with 20 questions and perhaps run the biggest two engines at three repeats, then expand. A smaller panel that runs every week beats a large one that gets skipped. ## Which numbers should you report? Keep the scorecard short and calculate it per engine. - Mention rate: answers naming your brand, divided by total answers. Fame's [share of answers calculator](https://www.fame.so/tools/share-of-answers-calculator) can do this step. - Citation rate: answers citing your domain, divided by answers citing anyone. Excluding empty answers matters for Gemini and Copilot, where Ahrefs saw a majority with no citations. - Competitor share: mention rate for your top three rivals, using the same formula. - Accuracy: share of fact-check answers free of errors on price, product or positioning. - Stickiness: for each question, how many of the last seven days or runs cited you. Ahrefs recommends this sort of measure over single checks. Mentions and citations are different and should stay apart. In the Ahrefs data, the cited site's brand went unnamed in 54% of ChatGPT cases and 84% of AI Mode cases. Ahrefs also found mentions more stable, and suggests normalizing brand name variants before counting losses. Decide up front whether "Acme", "Acme Inc" and "acme.com" all count. ## How should you interpret changes? Use simple decision rules so that you do not chase noise: - One missing week: note it and do nothing. Ahrefs advises reassessing only if a citation stays absent for two to three weeks. - A new citation: note it, but do not celebrate until it repeats. Citations that lasted 8 to 14 days almost always returned, while about 35% of ChatGPT one-day citations never did. - Absent for three weeks: check Google ranking, recent page changes, and whether the page title still matches the question. - A competitor rising across a question group: open the sources cited for them. That usually tells you which page or list to look at next. - A factual error: fix your own page first, then the third-party page that likely caused it. A word of caution on small numbers. With three runs per question, a question moves in 33-point steps. Combine results by group over several weeks before claiming a trend, and avoid quoting precise percentages from a small panel to your CEO. Also bear in mind [Ahrefs' warning](https://ahrefs.com/blog/query-fan-out-chatgpt/) that a model update could shift retrieval patterns, so note model changes in your log. ## How do you connect visibility to pipeline? Add three checks to the weekly routine: - Referral visits in GA4. Look at session source for AI domains such as chatgpt.com and perplexity.ai. Expect low numbers. Fame's own site logged 41 AI assistant visits in 28 days against 42,966 Google impressions at its October 30, 2026 snapshot. - Search Console. [Google counts AI feature traffic](https://developers.google.com/search/docs/appearance/ai-features) inside the Web search type, so watch overall trends and question-style queries there. - A required free-text field on your demo form that asks how the buyer heard about you. Many buyers read an AI answer without clicking anything, so this catches some of what analytics miss. Referrals, in other words, understate influence. Self-reported data has its own flaws, but when both move together you have a stronger signal. ## When should you buy a tool or hand this off? Start with a spreadsheet and the [AI visibility checker](https://www.fame.so/tools/ai-visibility-checker). Buy a tool when manual runs are the reason the work slips. Before you buy, ask how many repeats it runs per question, how it accesses each engine, whether it reports mentions and citations separately, and whether you can export raw answers. Tracking alone changes nothing, so plan the fix as well. [The 90-day plan](https://www.fame.so/post/90-day-ai-search-plan) shows how to pair measurement with page work. [Fame Search](https://www.fame.so/search), Fame's paid service, includes weekly tracking across four engines plus site work for $2,000 a month. On its own site, 38-76% of tracked answers cited fame.so from June 1 to October 30, 2026. Those are the only results Fame has, and they do not forecast yours. ## FAQ **How many prompts do I need to track?** Start with 20 to 50. No standard number exists, but this is enough to cover several buyer jobs while remaining possible to review by hand. Add more only if you will act on them. **Should I track logged-in or logged-out answers?** Pick one and stay consistent, and write it down. Personalization and models differ by account and plan, so a changed setup can look like a changed result. **What is share of answers?** In this method, it is the percentage of your tracked answers that name your brand. Calculate it per engine and per question group rather than as one blended figure. **Do I need a paid tool?** Not to start. A spreadsheet works for a small panel. A paid tool becomes useful when the panel is large enough that manual runs stop happening every week.