AI visibility is measurable. It is just not measurable the way rankings are. If you import rank-tracking habits, you will produce confident numbers that mean nothing.
Why rank tracking does not transfer
A rank tracker asks one question, repeatedly, from one location, and gets a stable answer. An answer engine gives a different answer depending on:
- the model and version,
- whether web search was used,
- the date,
- the phrasing of the prompt,
- the user's history and location,
- random sampling.
There is no position one. There is a distribution of outcomes across prompts and runs.
Build a prompt set
Start with 20 to 40 prompts, not one.
Group them by intent:
- Category prompts. "Best [category] in [market]".
- Comparison prompts. "[Brand] vs [competitor]".
- Problem prompts. "How do I fix [problem]?"
- Brand prompts. "What is [brand] known for?" This tests whether the entity is understood at all.
- Buying prompts. "Who should I hire for [job] in [city]?"
Write them the way a customer would, typos and all. Do not write them the way a marketer would.
Sampling and noise
Run each prompt several times, on more than one model, on different days. Record:
- whether you were mentioned,
- whether you were recommended (mentioned is not the same as recommended),
- your position in any list,
- which competitors appeared,
- whether a source link pointed at you,
- the exact wording of the claim made about you.
Three runs per prompt is a reasonable starting point. Anything less and you are measuring the weather on one afternoon.
What to record
A simple spreadsheet beats a dashboard you do not trust.
| Field | Why it matters | | --- | --- | | Date | Movement over time is the only signal | | Model | Different models behave differently | | Prompt | The unit of analysis | | Mentioned | Binary, easy to count | | Recommended | The one that correlates with revenue | | Competitors | Who is winning the prompts you lose | | Claim accuracy | Are they describing you correctly? | | Source cited | Are they citing you or a third party? |
Interpreting movement
Two rules keep you honest.
Look at rates, not instances. "We appeared in 14 of 120 runs, up from 9" is a finding. "We appeared in ChatGPT" is an anecdote.
Separate understanding from recommendation. If models describe you accurately but never recommend you, you have a positioning problem, not a visibility problem. If they recommend you but describe you wrongly, you have an entity problem.
Reporting it to a client
Be explicit about uncertainty. A good report says:
- Here is the prompt set and why it represents the category.
- Here is the sample size and the date range.
- Here is the rate of mention and recommendation, with the change.
- Here is what competitors are doing in the prompts we lose.
- Here is what we changed, and what we expect to move.
A bad report says "AI visibility score: 72/100" with no method. The number looks precise and tells nobody anything.
What actually improves it
In rough order:
- Entity clarity. One consistent description of the business everywhere.
- Quotable content. Direct answers, specific numbers, clear definitions.
- Third-party corroboration. Reviews, directories, press, citations.
- Technical access. Let the crawlers in, including the AI ones you want.
- Patience. These systems update on their own schedule.
Measure monthly. Change one thing at a time. Otherwise you will never know what worked.