How to measure AI visibility without fooling yourself

To measure AI visibility, build a fixed panel of prompts based on real audience questions, run them on a schedule across ChatGPT, Gemini, Perplexity and Google AI Overviews, and record whether your brand is mentioned, cited or linked, and how it is described. Repeat runs, because answers vary, and report trends rather than single results.

Why is AI visibility hard to measure?

Three reasons, and it helps to say them out loud before anyone builds a dashboard.

• Answers are non-deterministic. Ask the same question twice and you may get different brands, different sources and different wording.
• Answers are personalised and contextual. Location, account history, conversation context and model version all shift results.
• There is no published "position". Unlike a search ranking, there is no official rank to look up, and most AI platforms provide limited analytics to brands.

So any single screenshot of an assistant naming your brand, or failing to, is an anecdote. Measurement means sampling.

What should you measure?

We track four layers at SNMRush, matching our name: Search, Neural answers and Mentions.

1. Presence: the share of prompts in which your brand is mentioned at all.
2. Citation: the share in which one of your pages is shown as a source or linked.
3. Accuracy: whether the description is correct (category, offer, location, no confusion with other brands).
4. Share of mentions: how often you appear relative to named competitors across the same panel.

Alongside those, keep the classic search metrics, because many AI answers are grounded in search results: rankings for the same questions, Search Console impressions, and referral traffic from AI tools in your analytics (sources such as chatgpt.com or perplexity.ai now appear in many referral reports).

How do you build a prompt panel?

The panel is the heart of the method, and its quality decides whether the numbers mean anything.

Most teams start by brainstorming prompts in a meeting. The result is a list of questions the marketing team would ask, which is rarely what customers ask. We build panels from evidence instead. SNMRush uses SOMIN, an AI audience-research platform, to surface the questions and tensions people voice about a category in real social conversation, and we phrase prompts in the audience's own language. The method is described in finding the questions your audience asks.

A balanced panel usually includes:

• Category questions: "What is the best way to…", "Which tools help with…"
• Problem questions: the tensions behind the search, phrased as people phrase them.
• Comparison questions: "X vs Y", "alternatives to X".
• Brand questions: "What is [brand]?", "Is [brand] good for…"
• Local variants if you operate in more than one market.

Fifty to a hundred and fifty prompts is a workable range for most brands. Freeze the panel for a quarter so trends are comparable.

How often should you run it?

Monthly is enough for most brands, with each prompt run several times per engine to smooth out variation. Record the date, engine, mode (for example, with or without web search), location and full response. Keep the raw answers; you will want to reread them when a number moves.

Definitions

• Prompt panel: a fixed list of questions used to sample AI answers over time.
• Presence rate: percentage of panel runs that mention your brand.
• Citation rate: percentage of runs that show your domain as a source.
• Share of mentions: your mentions as a proportion of all tracked brand mentions in the panel.

What are the common mistakes?

• Reporting a single run. One answer is a sample of one.
• Changing the panel every month. You lose comparability.
• Counting mentions without reading them. A mention that describes you wrongly is a problem, not a win.
• Claiming causation. If presence rises after you publish a page, that is encouraging, not proof. Models update, competitors change, and engines tweak retrieval.
• Ignoring the money. Visibility is a means. Tie it to enquiries and revenue where you can.

That last point deserves emphasis. Our sister app Trafix makes the case for paid media that vanity ROAS hides true profit; the AI visibility equivalent is a rising mention count that never turns into a conversation with a buyer. Track assisted conversions and enquiry sources alongside the visibility metrics.

How should results be reported?

A one-page monthly readout works best: the four headline rates with trend lines, the prompts that moved most, three examples of answers quoted verbatim, the sources AI engines cited most in your category, and a short list of actions. The SOMIN reporting page shows a similar philosophy applied to marketing reporting more widely: every number should trace back to the evidence behind it.

For a sense of how agencies make evidence-led reporting part of a client relationship, the Analytic Partners case study is worth a look.

What does a good baseline look like?

There is no universal benchmark, and you should be suspicious of anyone quoting one. A good baseline is simply your own first quarter of data, measured consistently. From there, the questions become practical: which prompts never mention us, which sources do engines cite instead, and what would it take to become the better source?

One more practical tip: log the model or mode each engine reports where it is visible. When a platform ships a new model, results can shift overnight, and you will want to separate that from the effect of your own work.

That is where measurement turns back into work: the answers point to answer-first content to write, entities to clarify and mentions to earn. Measure honestly, act on the gaps, and measure again.

Frequently asked questions

Can I track AI visibility in Google Search Console?

Search Console includes AI Overviews and AI Mode traffic within overall Search performance, but does not separate it out in detail. A prompt panel is needed to see mentions and descriptions.

How many prompts should a visibility panel have?

Fifty to a hundred and fifty is workable for most brands, each run several times per engine per month, and frozen for at least a quarter.

Are AI visibility tools accurate?

They are useful for sampling at scale, but all face the same variability. Check how they run prompts, how often, and whether you can read the raw answers.