How Do You Track AI Visibility Over Time? A Monthly Routine That Separates Trends From Noise

You track AI visibility over time by holding the method constant: the same list of buyer questions, put to the same engines, on a regular schedule, with the same fields recorded each time. Anything that changes between runs is then a change in the answers, not in how you asked.

A monthly cadence suits most businesses. Weekly checks mostly capture noise, and quarterly checks are too slow to connect a result to the change that caused it. Before you build your own log, it helps to know where you stand: the free scan in the AEO Grader tool gives you a baseline score in about a minute. If you need the definitions first, our guide to measuring AEO performance covers which metrics matter, and the AI share of voice glossary entry defines the comparison metric. This post is about the routine that keeps those numbers comparable from one month to the next.

Two forces make one-off checks unreliable. The first is answer variance. Language models do not return the same text every time. Anthropic's glossary notes that even with temperature set to 0, results will not be fully deterministic, and identical inputs may produce different outputs across API calls. Consumer chat products add further variation through search results that change by the hour. Google adds its own: its AI features documentation says AI Overviews are only shown when its systems judge them additive to classic Search, and as such often do not trigger. A brand that appears in one run and vanishes in the next may not have lost anything.

The second force is lag. Engines do not see your fixes the moment you ship them. Google's guide to asking for a recrawl says crawling can take anywhere from a few days to a few weeks. OpenAI's crawler documentation says it can take about 24 hours from a robots.txt update for its systems to adjust, and Perplexity's crawler page says changes may take up to 24 hours to be reflected. Content changes travel further: a new comparison article on another site has to be published, crawled and picked up before it shapes an answer. A check run the day after a fix measures the old site.

The question set is the instrument, so build it carefully and then stop editing it. Aim for fifteen to twenty-five questions spread across the stages of a purchase. Write them the way a buyer would type them, not the way your marketing team would.

• Who are you: what does your company do, and who is it for?
• Can I trust you: is your company reputable, and what do reviews say?
• How do you compare: how does your company compare to a named rival?
• Alternatives: what are the best alternatives to your company?
• Unbranded category: what are the best options in your category for a specific type of buyer?

The unbranded questions matter most, because a mention there is earned rather than prompted by your own name. Keep the wording in a shared document with a version number. If you must add a question later, add it as new and keep the old ones running, so the history stays comparable.

Fix the conditions too. Use the same engines each month, such as ChatGPT, Claude, Perplexity, Gemini and Google AI Mode. Use a logged-out or clean session where the product allows it, note the country you asked from, and ask each question two or three times in fresh chats, because one run cannot tell you whether a mention is stable.

A spreadsheet with one row per question, per engine, per run is enough. Record the fields below every time, even when nothing seems to have changed, because a run of unchanged months is itself evidence. Paste the full answer text into a linked document rather than summarizing it; six months from now you will want to see exactly what was said, not what someone thought was important at the time.

Your own prompts show what engines say. Two free reports show what actually happened at scale, for the surfaces they cover.

Google announced Search Generative AI performance reports in Search Console on June 3, 2026. They show impressions within AI Overviews and AI Mode by page, country, device and date, with daily, weekly and monthly granularity, and the announcement notes that as of August 31, 2026 they had rolled out to all websites worldwide. Export the monthly view and paste it into the same sheet as your prompt log.

Microsoft introduced AI Performance in Bing Webmaster Tools on February 10, 2026. It reports total citations in AI-generated answers across Copilot and Bing's AI summaries, the grounding queries that led to them, and a timeline of citation activity. In June 2026 Microsoft added Intents, Topics, Citation Share and Compare; Compare overlays a previous time period on the current reporting view. Microsoft describes Citation Share as an observational metric rather than a ranking system, which is the right way to read every number in this post.

Neither report covers ChatGPT, Claude or Perplexity. That gap is why the prompt log stays at the center of the routine.

Once you have three months of data, the temptation is to react to every movement. Resist it with a few simple rules.

Count shares, not single answers. If you ask each question three times, a brand named in two of three runs is in a different position from one named in three of three, but a drop from three to two in a single month is well within normal variation.

Wait for persistence. Treat a change as real when it holds for two consecutive months, or when it shows up across several engines at once. A single engine moving once is a note in the margin, not a meeting.

Line changes up with your log. If visibility rose a month after you corrected a directory listing, and the engine now cites that directory, you have a plausible cause. If it rose with no change on your side, look at the sources column: a new third-party article is the usual explanation.

Watch accuracy separately. Mention rates can rise while the facts get worse. A wrong price repeated by two engines deserves action the same week, whatever the trend line says.

Put this on the calendar for the same week each month. The first run takes longer while you set up the sheet; after that, an hour is realistic for a twenty-question set. If you are making site changes at the same time, our AEO audit checklist shows what to fix and in what order, and local businesses should add the location questions from our guide on how local businesses show up in AI search.

If you would rather not run the prompts by hand, AEO Grader's monthly tracking automates the part the free reports cannot cover. It costs $9 a month and can be cancelled any time; the annual plan is $96 billed once a year, which works out to about $8 a month.

Each month it produces a fresh four-engine report, asking ChatGPT, Claude, Perplexity and Gemini the same panel of buyer questions it asks for every brand, built from the brand, website, category and competitor list on your account. Reports run on your own monthly anniversary rather than a fixed calendar day. A month only counts when all four engines answer; if one fails, the run is retried rather than sent, so you never get a fake drop caused by a missing engine. If a month's scan cannot be completed, for example because your site was unreachable, you get an email saying so instead of silence.

The monthly email is built by fixed rules from this month's numbers and last month's, with no language model writing it. The subject line leads with the most important change, in this order: an engine that stopped naming you, a competitor that overtook you, an engine that started naming you, then whether overall visibility rose, fell or held. The body shows your AI visibility out of 100 and its change, your share of voice and its change, a per-engine table of what moved, a list of competitors now named more often than you across the panel, and the single highest-impact action from your report. Every engine's full answer and cited sources are in the linked report.

You can try the free scan in the AEO Grader tool first; it is an estimate from one grounded Gemini and Google Search check, not a four-engine run. Nothing in tracking guarantees a mention; it tells you, every month, what changed.