How to measure your visibility in AI answers

Why you cannot measure this like rankings

A search ranking is stable enough to report as a number. An AI answer is not. The same question can produce a different list of brands, in a different order, each time it is asked.

SparkToro and Gumshoe.ai asked volunteers to run 12 brand-recommendation prompts through ChatGPT, Claude and Google’s AI features, 2,961 prompts in total, each repeated 60 to 100 times per platform. Reported by Search Engine Journal in January 2026, they found that ChatGPT and Google’s AI features returned the same brand list in under 1 percent of repeated runs, and the same list in the same order in under 0.1 percent. Across 994 varied headphone prompts, four brands still appeared in 55 to 77 percent of responses.

Read that carefully. It says exact lists and positions are unreliable, while the overall set of brands an assistant tends to name is more stable. The study is not peer reviewed, and one author runs an AI tracking company, which they disclosed. It is evidence for caution, not a final answer.

Build the prompt set

Choose 20 to 30 buyer questions across the buying journey, weighted toward the commercial ones. Write them as buyers would. Include head questions ("best accountant for an online store"), branch questions ("how much does an accountant for an online store cost", see query fan-out), comparison prompts and local prompts.

Keep the set stable. Add new questions deliberately, tagged with the month they joined, rather than constantly changing it, or the trend line means nothing.

Run it repeatedly, across assistants

• Test the main assistants your buyers use, and note whether you ran them logged out or signed in, since personalization changes answers.
• Run each prompt several times, and use at least 5 to 10 runs for prompts you care about. No one has published how many runs make a number reliable, so report your run count.
• Record whether you are named, whether you are linked, how you are described and which sources were shown.
• Note which competitors appear and how often.

Report rates, not anecdotes

A single answer is an anecdote. The useful number is the share of runs that name or cite you, per prompt and overall, tracked monthly. Worked example: 25 prompts times 8 runs is 200 answers per assistant. If you are named in 46 of them, your share of answers is 23 percent. Next month the same method gives a comparable number.

Separate a brand mention from a citation, because they mean different things: a mention is your name in the text, a citation is a link to your page. Also record sentiment and accuracy, because being named wrongly is a finding too. See brand, entity and mentions for what drives them.

Use the first-party data you already have

• Search Console: Google announced Search Generative AI performance reports on June 3, 2026. They show impressions of your URLs in AI Overviews and AI Mode by page, country, device and date, and the data also stays in the overall performance report. Google said it would roll the reports out to a subset of sites first, and a note on the post says they reached all sites by August 31, 2026. The metrics Google lists are impressions, pages, countries, devices and dates, so read them as exposure.
• Google Analytics 4: Google’s default channel group documents an "AI Assistant" channel for sessions from sources such as ChatGPT, Gemini, Copilot or Grok. It excludes AI Overviews and AI Mode, which count as Organic Search. Expect undercounting, because some clicks arrive without a referrer.
• Server logs: check which AI crawlers fetch your pages, using the user agents the providers document. See AI crawlers and robots.txt.
• Your own customers: add "ChatGPT, Perplexity or another AI assistant" as an option to the "How did you hear about us?" field, and ask on the first call.

Tie it to the business

Visibility only matters if it leads to enquiries. Put monthly share of answers next to AI referral sessions, form submissions that mention an assistant, and sales calls where the buyer says they found you in an answer. Connect enquiries to the pages assistants cite.

Many AI-influenced buyers will never click. They read the answer, then search your brand name or call. So also watch branded search impressions and direct traffic for movement, and treat those as supporting evidence, not proof.

The limits, and how to choose a tool

Assistants personalize and change, so visibility is a trend, not a fixed fact. Tools differ in how they sample. Google’s generative AI guide warns that third-party tools do not have access to Google’s internal systems, so be sceptical of any tool claiming to show them.

Before buying a tracker, ask how many times each prompt is run, which assistants and which versions it covers, whether it uses the public interface or an API, how it detects a mention, and whether you can export the raw answers. Fishkin’s advice in the SparkToro coverage was to ask providers to show their method. We publish ours so the numbers can be judged.