Run the AI visibility scan yourself
This is the exact instrument we run at intake: nine prompts, three assistants, one scoring rule, one worksheet. It takes about forty minutes and costs nothing. We are giving away the measurement because measurement is the cheap half. Moving the number is the part that takes months.
Three rules before you start, and all three matter
Use the same wording every time. Not similar wording, the same string. The value of this exercise is entirely in the delta between one month and the next, and a reworded prompt makes the delta meaningless.
Turn off memory and personalisation, or use a logged-out or temporary chat. A model that remembers you have asked about yourself eleven times is answering a different question than the one a stranger asks.
Save the raw answers with the date, not your summary of them. In three months you will want to compare the actual words, and a note saying "it was mostly right" is not evidence.
The nine prompts
Run all nine on ChatGPT, Claude and Perplexity. That is 27 readings. Substitute your own details in the bracketed parts and then never change the wording again.
Does the model know you exist and get you right?
Who is [full name]?What is [full name] known for?Tell me about [full name], [role] at [organisation].
Are you named when the model is asked about your field, unprompted?
Who are the leading [your field] specialists in [your city]?Who should I hire for [the specific problem you solve]?Name five people worth following on [your topic].
What does the model say when somebody is checking you out?
[full name] vs [a named competitor]: who should I work with?Is [full name] credible? What is the evidence?What do people criticise about [full name] or [your organisation]?
Tier 2 is where almost everybody discovers the problem. Tier 1 usually comes back fine, because a model can find a LinkedIn profile. Tier 2 asks the model to volunteer you when nobody has said your name, and that only happens when independent sources have already put you in that category.
The scoring rule
| Result | Score | What it actually means |
|---|---|---|
| Named, accurately, with a citation to a page you control | 2 | The strongest possible outcome. The model has a source it trusts and it resolves to you. |
| Named, accurately, no citation or a third-party citation | 1.5 | Good. The record is working. The citation is somebody else's page, which is normal and not a problem. |
| Named, but a detail is wrong | 1 | The record exists and disagrees with itself somewhere. Find the surface carrying the wrong fact. |
| Not named, but competitors are | 0 | The category question is being answered from sources that do not include you. This is the most common Tier 2 result. |
| Confused with a namesake | -1 | Worse than absence. The model is confidently wrong, and correcting it is a longer job than building from nothing. |
| Refused or said it does not know | 0 | Neutral. Common for private individuals and not in itself a failure. |
Add the 27 scores. Divide by 54. That is your share of voice, and it is a number you can compare against yourself in ninety days. It is not comparable to anybody else's, because the prompts contain your own name and field, so do not treat it as a benchmark.
What the number will not tell you, and what people wrongly do with it
- The delta over ninety days is real signal, and it is the only number here worth acting on
- Tier 2 failure with Tier 1 success means the record exists and is not categorised
- A namesake result is a specific, findable problem with a specific fix
- Citations mostly come from sources you do not own, which is what the data below shows
- That a citation is a KPI. It is a diagnostic. Nobody buys anything because a model cited a page.
- That schema markup lifts AI citation. A controlled study of 1,885 pages against 4,000 controls found AI Overviews down 4.6% after adding JSON-LD.
- That longer articles get cited more. Correlation with word count is 0.04, and over half of cited pages are under 1,000 words.
- That writing a listicle with yourself at the top helps. When a brand's own listicle was cited, the brand was left out of the recommendation 69% of the time.
Two readings of that chart. First, social and user-generated surfaces dominate almost everywhere, so content that lives only on your own site earns very little. Second, and this surprises people: the New York Times sits 35th on AI Overviews at 0.4% mention share. A tier-one press placement bought specifically for AI visibility is buying prestige, not citations. It may still be worth buying. Just not for that reason.
What to do with a bad reading
A low score is not a content problem, which is where most people go next and why most people waste the following six months.
- Tier 1 failing. The record itself is thin or scattered. Fix the entity home and the profile alignment before anything else. Nothing downstream works without this.
- Tier 1 fine, Tier 2 failing. The most common shape. You exist and you are not categorised. The fix is third-party sources that put you in the category, in public, in places already crawled. This is the part that takes months.
- Tier 3 failing. Somebody checking you out finds nothing to reassure them. Usually solved by verifiable case studies and named references rather than by more publishing.
- Namesake confusion anywhere. Stop and treat this as its own project. Adding more content while a namesake owns your name makes the collision worse, not better.