Practical guide · Measurement

How to measure AI visibility: tools and method

By Franck Gordi16 September 2026Updated: September 202611 min read

Measuring AI visibility means repeatedly querying generative engines with a fixed prompt set, then counting how often your brand is named, cited as a source, and outranked by competitors in the answers they return.

Most teams now know they need to be visible in ChatGPT, Perplexity or Gemini. Very few know where they currently stand. Without a starting number, every content decision is a hunch, and no budget argument survives three months.

1. AI visibility measurement, in one answer

There is no Search Console equivalent for generative engines. No model vendor currently publishes an interface that would tell a site how many times it was used in answers. Measurement is therefore necessarily indirect: you reproduce a user's behaviour, observe the answers, and count.

That constraint carries a methodological consequence worth accepting up front. You are not measuring your real visibility across all users, you are measuring the visibility observed on a sample of questions you chose yourself. The quality of your measurement therefore depends entirely on the quality of that sample, far more than on the tool running it.

2. The five metrics that matter

Monitoring platforms publicly describe indicators that are very close to one another. They can be reduced to five, each answering a different decision question.

The five AI visibility metrics

What each indicator measures, and the decision it supports

Mention rate: Share of tracked prompts whose answer names your brand. It answers one question: are you in the consideration set at all.
Citation rate: Share of tracked prompts whose answer links to your domain. A high mention rate with a low citation rate means the engine talks about you but sends the click elsewhere.
Average position: Where your brand sits in the answers that include it. Being named first and being named sixth are not worth the same thing.
Share of voice: Your share of mentions against your competitors on the same prompt set. It is the only metric that places your score inside a market.
Sentiment and accuracy: How the answer describes you, and whether it gets things wrong. A mention carrying a wrong price, scope or positioning costs more than no mention at all.

None of these metrics is standardised across vendors. Two tools can report two different scores for the same brand on the same day, because they query neither the same prompts nor the same engines. Compare trends inside one tool, never absolute values between tools.

The mention and citation split is the most useful

This is the cross-reading that produces the most actionable decision. Being named without being linked means the engine knows your brand but prefers to back its answer with third-party pages: comparison sites, directories, forums, press. The work is then on the page, not on awareness. Conversely, being linked without being named signals a useful page on a topic where your brand is not yet associated with the category.

This reading matches what we cover in our article on how LLMs choose their sources.

3. Building your prompt set

This step decides the value of the whole measurement, and it is the one most often rushed. A prompt set is not a keyword list: it is questions phrased the way a human asks an assistant, in full sentences.

Prompt family	Example phrasing	What it reveals
Category	"What is the best tool to do X?"	Whether you are in the default consideration set at all.
Comparison	"Your brand or Competitor, which should I pick for Y?"	How the engine positions you against a named competitor.
Brand	"Is Your brand any good, is it reliable?"	What the engine says about you, and which sources it leans on.
Problem	"How do I solve Z in this situation?"	Whether your content surfaces before the category is even named.
Objection	"What are the alternatives to Your brand?"	Who the engine offers when a user wants to leave.

Go wide: a useful set starts around thirty to fifty prompts. Below that, model variability drowns the signal.
Freeze it: a rephrased prompt breaks the series. Add prompts, do not edit existing ones.
Set language and market: the same question in English and in French does not return the same answer, and the cited sources differ sharply between markets.
Name your competitors: without them you measure your presence, not your place.

4. The monitoring method, in five steps

Step 1: establish the baseline

Run the full set across every engine you want to track, and store the raw answers, not just the scores. That first series is your zero point: without it, no later number is interpretable.

Step 2: record the cited sources

For every answer, note the domains cited. This is the most concrete output of the exercise: it tells you exactly which third-party pages carry authority on your topic, and therefore where to earn a presence. We covered that mechanism in our analysis of the most cited domains in AI engines.

Step 3: repeat at a fixed interval

A weekly rhythm suits most sites. What matters is regularity: same day, same prompts, same engines, same market. Any protocol change must be dated and logged, otherwise a break in the series will read as a drop in performance.

Step 4: cross-check with logs and traffic

Your server logs show visits from AI vendors' crawlers, and therefore whether your pages are even reachable. That is the technical prerequisite covered in our robots.txt and AI crawlers guide. Your analytics show outbound clicks from answers. Neither measures citation, but a gap between the three quickly reveals where the blockage sits.

Step 5: decide, then measure again

A measurement that triggers no action is a cost. Every reading should produce one short decision: a prompt you are absent from that deserves a page, a third-party source to work on, a factual error to correct. Then let at least four readings pass before judging the effect.

5. The tools, four families

Available solutions fall into four categories that do not measure the same thing. Conflating them is the most common mistake.

Family	What it measures	Main limit
Manual logging in a spreadsheet	Mentions, rank and sources on a small prompt set, reading the answers directly.	Slow, hard to sustain, and sensitive to your own session history.
Dedicated monitoring platforms	Automated prompt runs across several engines, scores, share of voice, sentiment, alerts.	Proprietary definitions that do not compare across vendors, recurring cost.
Server log analysis	Visits from AI vendors' crawlers, pages fetched and frequency.	Measures access, not citation. A crawler reading a page guarantees nothing.
Analytics and referrers	Sessions coming from a click inside a generative answer, and their landing pages.	Only sees the fraction of exposure that produces a click, so it structurally understates visibility.

How to choose within the second family

Is the score definition written down: a vendor that does not publish the formula behind its indicator is selling you an opinion, not a measurement.
Can you read the raw answers: without the full text you can neither verify a score nor spot a factual error about your brand.
Are market and language configurable: tracking calibrated on the English-speaking market does not describe your position elsewhere.
Is the data exportable: your history is the only asset this measurement produces. It has to be able to leave the tool.

6. What these numbers do not tell you

Three limits deserve to be stated plainly before you put a dashboard in front of a leadership team.

It is not market share: your score covers your prompts, not the questions your customers actually ask, whose volume nobody publishes.
Personalisation escapes measurement: user memory, history and location change the answers. An automated reading works on a neutral session, which belongs to nobody.
A model update can move everything: a version change at a vendor sometimes produces a break in the series with no relation to your work. Date those events in your tracking.

These caveats do not cancel the value of measuring. They simply bound what you can conclude from it. A stable trend across twelve readings remains the best signal available today, and it is far better than no measurement at all.

7. The action plan

Write thirty prompts this week: ten category, ten comparison with named competitors, five brand, five problem.
Run a first manual reading on two engines, in a logged-out session, and store the complete answers.
List the cited domains in those answers and rank them by frequency. That is your off-site roadmap.
Check technical access before investing in content: a site AI crawlers cannot read will never be cited, whatever the editorial budget.
Book a weekly slot and hold it for three months before concluding anything.

If you would rather start from the other end, our guide on how to be visible on ChatGPT covers the levers themselves, once measurement is in place.

FAQ

By repeatedly querying generative engines with a fixed prompt set, then counting how many answers name your brand, how many link to your domain, where you rank inside the answer and how much space competitors take. A single measurement means nothing: only repetition at a fixed interval on a stable prompt set produces a readable trend.

A mention is being named in the body of the answer. A citation is being used as a source, with a link to your domain. The two often come apart: an engine can recommend a brand while relying on third-party pages to back it up. A high mention rate with a low citation rate means your brand is known but your pages are not being picked as sources.

Because none of these metrics is standardised. Each platform defines its own prompt set, its own tracked engines, its own sampling frequency and its own way of counting a partial mention. Two tools can therefore report two different scores for the same brand on the same day without either being wrong. Track the trend inside one tool, do not compare absolute values between tools.

Yes, manually. Ask twenty to thirty questions representative of your market to each engine, record in a spreadsheet whether your brand appears, at what rank and which sources are cited, then repeat monthly under the same conditions. It is slow and sensitive to session bias, but it is enough to establish an honest baseline before paying for a tool.

Weekly is the right compromise for most sites. More often and you mostly measure the noise caused by model non-determinism. Less often and you miss the effect of a model update or a publication. Regularity matters more than frequency: a fixed interval is what makes the gaps interpretable.

They measure something adjacent but different: visits from AI vendors' crawlers to your pages. That is an access signal, not a citation signal. Knowing a crawler read a page does not tell you whether it was used in an answer. The two measurements complement each other, they do not replace each other.

Only partly. Outbound clicks from an AI answer arrive with an identifiable referrer, which lets you isolate that source in your analytics. But most exposure produces no click at all, because the answer is often enough for the user. Referral traffic therefore measures the bottom of the funnel, not visibility itself.

Sources

AmICited, product page "AI brand visibility tracking". Published definition of a visibility score and of the gap between share of voice and share of citations. Read the page.

Siftly, product page "AI Brand Monitoring Software". Published breakdown of a generative answer into several dimensions, including mention rate, position and sentiment. Read the page.

Peekaboo, product home page. Public example of a dashboard combining visibility, sentiment, average position and source domains. Read the page.