How to Measure Your Brand’s Visibility in AI Answers

Ask ChatGPT for the best payroll software for a ten-person company and you get a short list. Ask again an hour later and the list changes. Ask Perplexity and you get a third answer with different sources underneath it. Anyone who has tried to report on “AI visibility” from a handful of screenshots has run into this. A single answer is one draw from a distribution, and it tells you very little on its own.

This guide sets out a method that holds up when a client or a CFO asks how you know. It is the same method our AI Visibility Monitor uses, but nothing here depends on our software. You can run it with a spreadsheet and some patience.

What an answer engine does before it answers

When ChatGPT search, Perplexity, Gemini, Claude or Grok answers a question with web access turned on, it usually runs one or more searches first, reads some of the results, and writes its reply from that material. Google describes a version of this for AI Overviews and AI Mode as “query fan-out”: several related searches across subtopics, run while the answer is being built. The general term for tying an answer to sources fetched at question time is grounding, and AsymiLink Meta’s AI glossary has a short, plain-language definition if you want one to share with a client.

Two practical consequences follow.

• You have to be retrievable. If the engine’s search crawler is blocked, or your main content only appears after JavaScript runs, you are out of the running before the model writes a word. Our guide to AI crawlers and robots.txt covers which bots matter.
• The sources shape the answer. Engines often cite review sites, directories, forums and comparison articles rather than vendor pages. If those pages name your competitors and not you, the answer will too.

Step 1: Write the prompts your buyers would actually type

Start with a prompt set, not a keyword list. People type full questions into assistants, often with context about themselves: “I run a small dental practice in Brisbane, which booking software works with Medicare?” That sentence would never show up as a keyword, but it is exactly what you need to test.

A useful set has a few properties:

• Mostly unbranded. Around four in five prompts should leave your brand out entirely: problem questions, “best X for Y”, how to choose, cost, alternatives. Branded prompts (“is Acme any good”, “Acme vs Globex”) are worth tracking, but they flatter you, because the engine has been handed your name.
• Spread across the funnel. Early questions about the problem, middle questions comparing options, late questions about pricing, reviews and who to hire. Visibility often looks very different at each stage.
• Grouped into a few topics. Four to eight topic labels are enough to see where you are strong and where you never come up.
• Local where it matters. If the business sells in particular cities or countries, include prompts that name them.

Fifty prompts is a reasonable start for one brand. Fewer than twenty and the numbers below get too noisy to act on.

Step 2: Run every prompt more than once

Because answers vary from run to run, ask each prompt on each engine several times and treat the results as a sample. Three runs per prompt per engine is a sensible floor. For each prompt, record how often the majority outcome repeated. If your brand appears in two of three runs on one engine and none of three on another, that prompt is unstable, and a single screenshot of either result would have misled you.

Use the engines’ own APIs with web search switched on rather than a logged-out browser session. API answers are not identical to what a signed-in consumer sees, since the consumer apps can add personalization, memory and their own settings, but they are repeatable, they return the cited URLs in a structured form, and you can run them on a schedule.

Step 3: Report four numbers, not one

For every answer, record whether each tracked brand is named, where it appears in the answer, and which URLs are cited. From that you get:

Mention rate and citation rate move independently, and the gap between them is useful. A brand that is often named but rarely cited is being described from other people’s pages. A brand that is cited but not named has useful content that is not tied clearly enough to the brand. For the position weighting we use 1.0 for first place, 0.775 for second, 0.55 for third, 0.325 for fourth and 0.1 after that. The exact weights matter less than keeping them fixed so that month-on-month changes mean something.

Step 4: Put error bars on it

Suppose your brand is named in 5 of 20 answers. That is a 25% mention rate, but with only twenty answers the 95% Wilson score interval runs from about 11% to 47%. A competitor at 35% on the same sample is not clearly ahead of you. At 50 mentions in 200 answers the rate is still 25%, but the interval narrows to roughly 20% to 31%, which is tight enough to compare.

The Wilson interval is the standard choice for a proportion from a small sample. Unlike the textbook “p ± 1.96 × standard error” formula, it behaves sensibly near 0% and 100%. Show the interval next to every rate in your reports, and do not announce a change unless the old and new intervals barely overlap.

Step 5: Read the sources, not just the answers

The most actionable part of the data is the list of cited domains. Sort them by how often they are cited in answers that mention a competitor but never mention you. These are your gaps: a review site where you have no profile, an industry directory you are missing from, a comparison article that predates your product, a forum thread where nobody has mentioned you.

Fixing those gaps is ordinary, honest marketing work: claim and complete the profiles, ask real customers for reviews, give journalists and comparison writers accurate information. It is slower than editing your own pages, but it is usually what moves mention rate on unbranded prompts.

Your own pages still matter for citation rate. The pages that get cited tend to:

• answer one question directly near the top, in text rather than in an image or a script-rendered widget;
• state specific facts (prices, limits, supported integrations, locations) that an engine can quote;
• carry structured data that matches the visible text;
• sit on a site the engines’ search crawlers are allowed to fetch.

Step 6: Break it down by engine and by topic

An overall score hides most of the story. Report each engine separately, because they search differently and cite different kinds of sources. Then cut the results by topic and funnel stage. It is common to find a brand that does well on decision-stage prompts, where buyers already know the names, and barely registers on awareness prompts, where the shortlist is formed.

Google AI Overviews and AI Mode are a special case. Google says clicks from them are counted in Search Console’s Performance report under the Web search type, and that there are no special requirements beyond normal SEO. So for Google you already have click data. What the method above adds is the answers themselves: who was named, and which pages were used.

How often to measure

Weekly is enough for most brands. Daily runs mostly measure noise unless you are watching a launch or a crisis. Keep the prompt set fixed between runs, and when you add prompts, report them separately for a cycle so the trend line is not distorted. Set an alert threshold in points (for example, a change of 10 points or more in mention rate) rather than reacting to every wobble.

Where AsymiLink SEO fits

Our AI Visibility Monitor automates this method across five engines: ChatGPT, Perplexity, Gemini, Claude and Grok, each called through its vendor’s API with web search enabled. It generates a starting prompt set from your site, runs repeat samples, reports the four metrics above with 95% intervals and per-prompt stability, lists the source domains where competitors are cited and you are not, and can run on a daily, weekly or monthly schedule with an alert threshold. The answers come straight from the engines; there is no third-party panel or estimated data in between.

If you would rather run it by hand first, the steps above are the whole method. Start with fifty prompts, three runs each, and one honest spreadsheet.