01Method
How to measure AI visibility credibly
Measuring AI visibility means sampling a fixed set of buyer prompts across each platform, repeatedly, and reporting the distribution rather than a single reading. This guide describes the method well enough for you to run it yourself, without hiring anyone.
Quick answer
Measuring AI visibility means sampling a fixed set of buyer prompts across each platform, repeatedly, and reporting the distribution rather than a single reading.
This guide describes the method in enough detail to run it yourself, without hiring anyone. The tedious parts are included.
Key takeaways
- Build the prompt set from real buyer language, not a keyword tool. Keyword tools give search syntax; assistants get natural questions.
- Freeze it. A prompt set that changes between cycles makes trend data meaningless.
- Agree the competitor list before sampling. Choosing it afterwards is how measurement becomes marketing.
- Three to five runs per prompt. One run records variance and calls it a finding.
- Never blend platforms. The two most similar shared only 24.1% of cited pages.
- A large share of answers cite nothing. Any vendor claiming complete coverage is overstating what is possible.
Short answer
Freeze a prompt set built from real buyer language. Run each prompt several times per platform on a fixed cadence. Code every response for presence, citation and accuracy against documented rules. Report per platform, never blended, always with sample size and date.
Step 1 — Build the prompt set from real language
Sources: sales call recordings, support tickets, lost-deal notes, existing search queries, and the questions prospects ask on first calls. Not a keyword tool. Keyword tools give you search syntax; assistants get natural questions.
Aim for 40 to 120 prompts across three stages, and keep the stages labelled because they behave differently:
- Problem awareness. "How do companies usually handle [problem]." You will rarely be named here and that is normal.
- Solution evaluation. "What should I look for in a [category] platform." This is where the criteria you are later judged against get set.
- Vendor comparison. "Who are the main vendors for [use case]." This is the step that decides whether you are in the process at all.
Then freeze it. A prompt set that changes between cycles makes trend data meaningless. Version it if you must change it, and note what changed.
Step 2 — Agree the competitor set first
Share of voice is meaningless without a named competitor list, and choosing it after seeing results is how measurement quietly becomes marketing. Five to eight companies you genuinely lose deals to. Write it down before you sample.
Step 3 — Sample with repetition
Run each prompt three to five times per platform, from consistent locale settings, in a fresh conversation each time. Non-determinism has to show up as a distribution. If you run once, you are recording variance and calling it a finding.
Two practical traps. Personalisation and chat memory contaminate results, so start clean each run. And your own location changes answers, so fix the locale and record it.
Step 4 — Code against written rules
For each response record: cited with a link, named without a link, or absent; which competitors appeared; whether the description of you was accurate; which of your URLs was used. Write the rules down so a different person reaches the same numbers.
The accuracy field is the one people skip and the one that most often produces the highest-value finding. Being described wrongly is a different problem from being absent, and it is usually cheaper to fix. See the three-state diagnosis.
Step 5 — Report per platform
Do not blend. In 2026 research the two most similar platforms shared only 24.1% of the pages they cited. A blended score averages away the finding, that you are strong on one surface and absent from another, which is the only actionable part.
| Measure | How to record it | The trap |
|---|---|---|
| Citation presence | Runs where your domain is cited ÷ total runs | Reading one cycle as a trend. Wait for three |
| Mention without citation | Runs where you are named but not linked | Counting it as absence. It is influence you cannot see |
| Share of voice | Your appearances against the named competitor set | Changing the competitor list mid-programme |
| Description accuracy | Coded against a canonical fact set you write first | Not having a canonical fact set to code against |
| Cited URL | Which of your pages was used | Assuming it will be the page you optimised |
Reading the result without fooling yourself
- Three periods before you call a trend. Two points is a line, not a direction.
- Accuracy drops beat presence drops for urgency. Being described wrongly costs more than being absent.
- Annotate platform changes. Retrieval shifts move numbers with nothing changing on your side. Note them rather than claiming or accepting credit. Our published methodology requires this.
- Keep prompt-level detail. Aggregates hide which specific queries you are losing, and those are where the commercial value sits.
Connecting it to pipeline, carefully
Visibility is not revenue, and the lag between them is long. In a B2B category with a nine-month cycle, work done this quarter shows up in closed-won three quarters later. Anyone attributing this month's pipeline to last month's citation rate is guessing.
Report the two separately, state the lag, and label anything connecting them Estimated rather than Verified. Then watch branded search alongside citation rate, because an unlinked mention often surfaces later as someone searching your name directly.
What this method cannot see
A large share of AI answers cite nothing. Those mentions are real, they influence buyers, and no tool including ours can measure them. Any vendor claiming complete coverage of AI visibility is overstating what is possible, and that is worth testing them on.
Related: our full client measurement methodology, ongoing tracking, and the two-week baseline if you would rather not run it yourself.
02Common questions
Questions that come up while running this
How many prompts do I need?
Forty to a hundred and twenty for a commercial programme. Ten is enough for a directional read you can do yourself in an afternoon. What matters more than the count is that the prompts come from real buyer language and that the set is frozen once agreed.
How often should I re-sample?
Monthly suits most brands. Weekly is rarely justified, because the variance between two weekly readings often exceeds the change you are trying to detect, so you end up reacting to noise rather than to movement.
Can I just use a tool instead?
You can, and several good ones exist. Check three things before trusting one: whether it reports platforms separately, whether it discloses sample size and repetition count, and whether it lets you set your own prompt list rather than generating one for you.
Why not use one overall AI visibility score?
Because platform overlap is low: the two most similar platforms shared only 24.1% of cited pages in 2026 research. A blended figure averages away the actionable finding, which is that you are strong on one surface and absent from another.
What counts as a good citation rate?
There is no published benchmark worth trusting yet, and anyone quoting one should be asked for their sample. Judge yourself against your own baseline and against the named competitor set you agreed before sampling, not against an industry average.
How do I measure mentions that have no link?
You cannot measure them directly, and no tool can. What you can do is watch branded search volume alongside citation rate. An unlinked mention often produces a later branded search, so movement there is a weak but real proxy.
03Related
Continue
Diagnosing absence · How each platform differs · Our research publication standard · Have us run it
Sources
Every figure on this page comes from one of the following. Where a claim is ours rather than a source’s, it carries an evidence label instead.
- Rankability — State of AI Search, 1,645 AI citation observations across 31 topics, 2026. Source for the 24.1% cross-platform overlap, the 73.5% ChatGPT exclusive citation rate, the 22.1% vs 8.7% first-100-words finding and the 43.3% vs 15.5% query-breadth finding. Independent study
- Google Search Central — Optimizing your website for generative AI features (published 15 May 2026, updated 10 July 2026). Source for the llms.txt, chunking and AI-schema positions, and for the statement that generative features run on core ranking systems. Primary
- Searlixa sampling across client engagements. Any figure drawn from this carries an Observed label with its sample size and date. Our own
Ask an assistant
Hand this page to an assistant and ask it to summarise or challenge the argument. If it gets something wrong about us, that is itself a useful finding.