How to track AI visibility: what to measure every week

What should you count before you pick a tracker?

AI visibility tracking comes down to six numbers, and the time to define them is before you compare tools. Trackers name and compute these numbers in their own ways, so two dashboards can show different percentages for the same business. Know what each one means, and you can judge any tool or run the count yourself.

Here is how they separate, in an invented round of 80 answers: 16 name you, 6 link to your site, one of the 16 states an old price, and you and four fixed rivals are named 64 times in all. That gives a mention rate of 20%, a citation rate of 7.5%, a share of voice of 25% and a wrong-fact rate of about 6%. These figures are illustrative, not a target. We know of no benchmark we could verify for a good share, so compare yourself with your own earlier weeks.

Why is there no rank inside an AI answer?

On a Google results page, your page holds a place in a list, and Search Console can average that place over many searches. An assistant has no such list. It writes a new answer each time, with its own names in its own order, and the length changes too: a January 2026 SparkToro study saw answers with as few as 2 or 3 recommendations and, equally often, 10 or more. “Position 3” has no shared list behind it.

Google’s help page on how Search Console counts its data reflects this: “An AI Overview occupies a single position in search results, and all links in the AI Overview are assigned that same position.” So Search Console gives no order among the links inside one Overview. Microsoft, announcing its AI Performance report in February 2026, says page-level citation activity “reflects how often pages are cited, not page importance, ranking, or placement.”

So what does an AI rank tracker report? We read the help pages of several trackers on November 4, 2026. They use “position” or “rank” for at least three different things.

• Mention order. Where your name came in one answer, where 1 means first mentioned, usually averaged across runs.
• Link order. Where your URL came among the sources cited in an answer or an AI Overview.
• Rank by visibility. Your place among all the businesses in the results, sorted by how often each one appears.

Each of these summarizes the sample that tool collected. None is a place you hold inside the assistant. Before you act on one, ask how many runs stand behind it and whether it counts the answers that did not name you: one help page we read defines average position as where a brand tends to rank “when it appears”.

Google’s guide to generative AI features, last updated July 10, 2026, adds: “No third-party tool has access to our internal ranking or AI systems.” It also says to use such tools if they help your workflow.

How do AI Overview trackers differ from ChatGPT trackers?

An AI Overview sits on a Google results page, so it is checked much like any Google search. An AI Overview tracker typically runs a keyword from a chosen location, records whether an Overview appeared, and notes whether your page is among its links.

Tools that follow ChatGPT, Gemini, Claude and Perplexity work differently, whether they are sold as LLM trackers (LLM means large language model), ChatGPT rank trackers or prompt tracking. They send a full question, called a prompt, and read the written answer.

A tool that covers one kind does not automatically cover the other, so check which ones any tracker lists.

The last row matters most for prompts. In a second SparkToro test, volunteers wrote 142 prompts for one need: headphones for a family member who travels. Barely two looked alike, yet across the 994 answers four brands appeared 55% to 77% of the time. Your fixed list is a sample of what buyers want, not a copy of what they type.

How different can two answers to the same question be?

The clearest public test we could read at its source is that SparkToro study, published by its CEO, Rand Fishkin, with Patrick O’Donnell of Gumshoe.ai, a startup in AI tracking. Fishkin raises that tie himself as a possible conflict of interest. In November and December 2025, 600 volunteers ran 12 prompts through ChatGPT, Claude and Google’s AI (an AI Overview, or AI Mode when no Overview showed) at their usual settings: 2,961 runs in all, 60 to 100 per prompt on each tool.

• The same list came back less than 1 time in 100, whatever the question.
• The same list in the same order came back less than 1 time in 1,000.
• Presence was steadier than order. One hospital appeared in 69 of 71 ChatGPT answers about West Coast cancer care, and came first in only 25.
• The size of the category shapes the numbers. In a wide field, brand design agencies, the highest figures sat in the 30s and 40s percent, well below the top headphone brands.

Fishkin began the work expecting tracking to be pointless. He ended by calling visibility percentage “across dozens to hundreds of prompts run multiple times” a reasonable metric, and by rejecting ranking position as a number to follow. He says the authors are not professional researchers, and he asks for far larger samples.

One reply says nothing, and a share across many replies says something. And a weekly sheet is a rough instrument: to know one prompt’s set of recommendations, Fishkin suggests asking “usually at least 60-100X”, while a weekly round asks each question once per assistant. So pool the answers across the list and across weeks, and read the result as a trend, not question by question.

Do API answers match what a person sees in the app?

Not necessarily. A tracker may ask through a company’s API, the paid channel built for software, or drive the app the way a person would. The companies’ own documentation shows how the two differ.

• ChatGPT. OpenAI’s help article on web search says ChatGPT “may use an approximate location based on your IP address” and may use relevant saved memories when rewriting a search query. Its API documentation lets the developer set the location instead.
• Gemini. Gemini Apps Help describes a personalized experience “using memory of your past chats”. Google’s API documentation for its search tool does not say how the two compare.
• Claude. Anthropic’s documentation says the web and mobile apps use a system prompt, which is a set of standing instructions, and that “These system prompt updates do not apply to the Claude API.”
• Perplexity. Its API FAQ answers the question directly: “The API uses the same search system as the UI with differences in configuration, so their outputs may differ.” It adds that the underlying AI model might differ too.

None of the four publishes how large the gap is. Surfer, which sells a tracker, reports that brand lists from APIs and from scraped apps overlapped only 15.5% to 23.8% (1,000 prompts, one sample each, August 2026). Gumshoe, which measures through APIs, argues that an earlier Surfer run used default API settings. Fishkin reports early data suggesting the gap may be less of a problem than he assumed, and still calls it an open issue. MWY Pro would use APIs, so we have a stake here too.

A manual run is a sample too: a private window lacks the history a real buyer has. So use one method per sheet, record it in a column, and never join app weeks and API weeks into one trend.

What can Search Console and Bing Webmaster Tools show you for free?

Google and Microsoft both report AI appearances to site owners. We checked each page below on November 4, 2026.

Google Search Console. Google’s help page says the Generative AI performance report reached all websites worldwide as of August 31, 2026. It covers AI Overviews and AI Mode and describes impressions: how many times links to your site were shown, by page, country, date and device. The page lists no click, position or query figures. Google says a click on a link in those features counts as a click, reported within the Web search type of the main Performance report, so AI clicks sit mixed with the rest. The report may not appear if your site has too few impressions.

Bing Webmaster Tools. Bing has an AI report. AI Performance entered public preview in February 2026 and covers Microsoft Copilot, AI-generated summaries in Bing and select partner integrations. It shows total citations, average cited pages per day, citations per URL, a timeline, and grounding queries: the key phrases the AI used when retrieving content, based on a sample. On June 16, 2026, Microsoft added four preview features, among them Citation Share: your percentage of all citations for one grounding query, with no competitors named.

Google Analytics. Its default channels now include AI Assistant, for “sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok”. AI Overviews and AI Mode are excluded and count under Organic Search. The setup and its gaps are in our guide to ChatGPT traffic in Google Analytics.

We found no public report of mentions or citations from OpenAI, Anthropic or Perplexity, and Google’s report lists Search features, not the Gemini app. For those four, asking and counting is the only way to see whether you are named.

What does a weekly tracking spreadsheet look like?

Freeze the question list. Start from the buyer questions described in our guide to AI visibility, give each one an ID, and stop editing. If you must change a question, start a new version and note the week. Keep 3 to 5 questions with your name in them on a separate tab, such as “What does [your business] charge for a new roof?” They show how you are described, and they stay out of the mention rate.

Give it to one named person. Anyone careful can run it, since the work is copying and marking. The owner reads the wrong-fact column each week, knowing the right hours and prices.

Keep the conditions the same. Same weekday, same account state, same location, web search on wherever the assistant offers a choice.

Save every answer in full, so you can recount when a definition changes.

The rates then take one formula each.

Mention rate:    =COUNTIFS(A:A,"W46",E:E,"Y")/COUNTIF(A:A,"W46")
Citation rate:   =COUNTIFS(A:A,"W46",H:H,"Y")/COUNTIF(A:A,"W46")
Wrong-fact rate: =COUNTIFS(A:A,"W46",J:J,"?*")/COUNTIFS(A:A,"W46",E:E,"Y")

For share of voice, add a Y or N column for each rival in your fixed set, and divide your count of Y by the total count of Y for you and the rivals.

Then keep a second tab with one row per week: the rates, Search Console’s AI impressions, Bing’s total citations, AI Assistant sessions from analytics, and anything you changed on your site that week, with the date.

How many weeks before a change means anything?

Ordinary margin-of-error arithmetic gives a first answer. Twenty questions across four assistants is 80 answers a round. Assume, as an illustration, a true share of 20%, or 16 in 80. At the usual 95% confidence level:

Run four rounds to set a baseline. Change one thing and note the date. Run four more, then compare the two averages, and treat a gap under about 5 in 80 as no news.

Treat this as a rule of thumb and a floor. The arithmetic assumes every answer is an independent draw, while answers to the same question tend to resemble one another, so the real noise is larger. SparkToro lists the number of runs needed as an open question. And none of the four assistants publishes how soon it reads a changed page again, so four rounds may still be too early.

• One assistant alone is noisier. Each gets 20 answers a round, so four rounds of one assistant carry the same noise as one round of all four. Read those numbers as hints.
• A shorter list takes longer. With 10 questions, you need twice as many weeks for the same precision.
• Wrong facts are the exception. A wrong price or wrong hours need no statistics. One is enough to act on.

When does a paid AI visibility tracker save more than it costs?

Judge it by hours and question count. What follows is our rule of thumb, not a measured benchmark.

The time for a round is questions, times assistants, times runs per question, times minutes per answer. Time your first two rounds to get your own minutes. At an illustrative one minute per answer, 20 questions across four assistants take 80 minutes. Three runs of each question make 240 answers: four hours, every week.

A spreadsheet is enough while you have one website, one country, about 20 questions and one run of each, and the round fits in about two hours a week. Paid AI visibility monitoring starts to make sense when any of these becomes true.

• The round passes about two hours a week, or keeps getting skipped. A sheet with gaps cannot show a trend.
• You want several runs per question, which manual work rarely sustains.
• You track several websites, countries or languages. Each one multiplies the round.

Then do the sum: the hours saved each month, times what that person’s hour costs you, against the subscription. If the saved hours cost less, keep the sheet.

Before you pay, ask how position and share of voice are defined. Our guide to AI visibility tools covers the other questions.

What the MWY checker does and does not track

The free AI visibility checker does not count any of these six numbers. It answers the question underneath them: what on your site stops an assistant from reading, knowing and quoting you? It reads your pages as raw HTML without running JavaScript, which is how outside tests have mostly found AI bots reading pages, skips anything your robots.txt closes to MWYGeoCheck, and scores 24 items in three groups: can AI read it, does AI know who you are, can AI quote it. You get the score and the 5 items that matter most in a few seconds, with no sign-up.

The score uses fixed rules, so the same site gets the same result: a steady column beside numbers that swing.

Pro, which opens soon, adds a weekly round. MWY writes 20 buyer questions from the content of your site, you can edit them, and every week each one goes to ChatGPT, Gemini, Claude and Perplexity through their APIs with web search on. The report shows which answers name your site, which other businesses they name, which pages they cite, and what changed since the week before. An answer through an API can differ from what one person sees in the app, so the report reads weeks, not single answers.

The checker does not track AI Overviews or AI Mode, does not read your Search Console, Bing Webmaster Tools or analytics, and does not report a rank inside an answer. MWY also does not promise that any assistant will name your business: nobody outside the AI companies controls that. What can be done is to remove what stops them from reading you, and to count honestly whether they do.