Click-defence asset
Twelve checks that tell you whether you are measuring AI search visibility across every engine your buyers use, or measuring one engine and assuming the rest. Nothing you enter leaves this page.
1. You have at least ten buyer-intent prompts, written as your buyer speaks
Not keywords. Full questions, in the phrasing a person types into an assistant.
Pass: 10 or more prompts on file.
2. None of the prompts contains your brand name
A prompt that names you measures recall of a name the engine was handed. The question that matters is whether you come up unprompted.
Pass: zero prompts mention your brand.
3. The set covers category, comparison and problem shapes
"Best X for Y", "A versus B", and "how do I fix Z" pull different sources. Testing only one shape measures only one slice of the funnel.
Pass: at least two prompts of each shape.
4. You test on four engines or more
ChatGPT, Google AI Overviews, Perplexity and Gemini at minimum. The overlap data says results on one do not predict the others.
Pass: four or more engines in every run.
5. Copilot is either tested or explicitly ruled out
Copilot leans on Bing, where its overlap runs at 16.6% against 8.2% on Google. If your buyers sit in a Microsoft estate, leaving it out is a hole.
Pass: tested, or a written note saying why not.
6. You record the cited URLs, not only whether you were named
The URL tells you which page won and which of yours should have. A mention count tells you nothing you can act on.
Pass: URLs captured for every answer.
7. Your tool reports which engine each result came from
A blended "AI visibility score" hides the one thing the data says matters. Insist on the breakdown.
Pass: per-engine results, not a single index.
8. You run a positive control on every measurement run
Include a domain you know is cited for the query. If the control comes back "not cited", the run is broken and every zero in it is meaningless. This is the check almost nobody runs, and it is the one that caught our own tooling.
Pass: control returns "cited" before you read any other result.
9. You re-run the same prompts at least twice a quarter
Answers move. One run is an anecdote, and a single reading cannot tell a real change from normal variance.
Pass: two or more runs per quarter on an unchanged prompt set.
10. Every lost prompt is mapped to the page that should have won it
If no page on your site deserves to win a prompt, the gap is a content gap, not a visibility gap.
Pass: a named page for every lost prompt.
11. You know which competitor won each lost prompt, and on which URL
The winning URL is the specification. Read it before you write anything.
Pass: competitor and URL recorded for every loss.
12. Your fix list is content and authority work, not tool configuration
No tracker has ever made anyone citable. If the quarter's actions are all dashboard settings, nothing will move.
Pass: the majority of open actions change pages or earn third-party coverage.
| Checks passed | What it means | The next move |
|---|---|---|
| 0 to 4 | You are guessing. Any claim about AI visibility right now is an opinion. | Build the prompt set first. Layer 1 costs an afternoon. |
| 5 to 8 | You are measuring one engine and assuming the rest behave the same way. The overlap data says they do not. | Add engines and the positive control before you spend another euro on content. |
| 9 to 11 | Real coverage with a known gap. You can trust the direction of travel but not every number. | Close the specific failing check. It is usually 8 or 11. |
| 12 | A measurement programme rather than a dashboard subscription. | Spend the time on the content and authority work the runs point to. |