How to measure AI visibility — and what you can't measure

Why there is no “rank” in AI answers

A search ranking is a position in a list that is fairly stable and that search engines themselves report through tools like Search Console. AI-generated answers are different: they are written fresh for each request, influenced by the exact wording, conversation history, location and account state, and they change as platforms update their systems. No major AI platform currently offers publishers a report of how often they appear in answers.

So AI visibility has to be estimated by sampling, the same way pollsters estimate opinion.

Step 1: Build a prompt set that reflects real buyers

A prompt set is the list of questions you track. Good sets are:

• Realistic — phrased the way buyers actually ask, often as full questions.
• Balanced across intent — category (“what tools help with X?”), comparison (“X vs Y”), problem (“how do I fix…”), and brand (“is Prisol any good?”).
• Stable — kept constant over time so trends are comparable; add new prompts as a separate cohort.

Sources for prompts include search query data, sales call notes, support tickets and the “related questions” features in search results.

Step 2: Sample repeatedly, per platform

Run each prompt several times on each platform you care about — typically ChatGPT, Google Gemini and AI Overviews, Perplexity and Microsoft Copilot — on a fixed schedule. Record the date, platform, mode (for example, with or without web search), and location. Repetition matters: a single run tells you what can happen, not what usually happens.

Step 3: Record separate signals

Signal	Definition
Mention rate	Share of sampled answers that name the brand
Recommendation rate	Share that suggest the brand as an option to choose
Citation rate	Share that link one of the brand’s own pages as a source
Average position	Where the brand appears in the answer relative to others
Accuracy	Whether facts stated about the brand are correct and current
Sentiment	Whether the description is positive, neutral or negative
Cited third-party sources	Which other domains are used as sources for the topic

Keep these separate. Being mentioned but never recommended is a different problem from being cited but described inaccurately.

Step 4: Compare against competitors

A 30% mention rate means little in isolation. Track the same signals for the competitors buyers actually consider, and report share of answers relative to that set.

Step 5: Connect to outcomes where you can

• Referral traffic. Visits from AI products often appear in analytics as referrals from their domains (for example chatgpt.com or perplexity.ai). Segment them.
• Branded search. Increases in branded search impressions in Search Console can indicate growing awareness, though many factors contribute.
• Self-reported attribution. Adding “How did you hear about us?” with AI assistants as an option to forms is simple and surprisingly informative.

What you can’t measure (and should say so)

• Actual exposure. Platforms don’t publish how many people asked a question or saw your brand.
• Personalised answers. Logged-in users with history may see different answers from your samples.
• Small changes. Differences of a few percentage points between months are often within normal run-to-run variation.
• Causation. A rise in mentions after a piece of work is encouraging, not proof; platforms change too.

Every report we produce lists the prompts, platforms, run counts, dates and locations, and flags changes that fall within normal variation. If a vendor’s AI visibility dashboard doesn’t tell you its method, treat its numbers with caution.

Getting started

You can run a basic version of this with a spreadsheet: twenty prompts, four platforms, three runs each, once a month. If you would rather have it done systematically with competitor benchmarks, see our AI visibility service or start with an AI Visibility Audit.