The front door, made tangible
Sample AI Visibility Audit.
This fictional specimen shows the structure of the £1,500 Visibility Audit. The practice, findings and numbers are invented to explain the report format. It shows how observations, unresolved questions and proposed fixes are presented; it is not a client audit or evidence of results.
What the audit shows
The Visibility Audit records whether a business appears when buyers ask AI engines who to use, who else appears, and which sources are shown. It combines an agreed set of client questions with checks on whether AI can read the site, whether the business is described the same way everywhere, the background code AI reads, and which other websites back the business up, then proposes a prioritised fix list. The findings below are illustrative.
Illustrative question observations
This fictional panel uses ten non-branded questions across ChatGPT, Gemini, Claude and Perplexity. The five rows below illustrate the observations a report records. In a real engagement, the questions, products, repeat dates and scope are agreed before testing. We keep the answers and links separately from our reading of them.
| Prompt | Practice named? | Businesses named | Sources shown | Priority |
|---|---|---|---|---|
| “Best cosmetic dentist in [city]” | named | Two competitors, then the practice | A local directory profile | medium |
| “Best dental implants in [city]” | absent | Five competitors | Two directories and a regional news feature | high |
| “Invisalign cost [city]” | named | The practice | Its own pricing guide | hold |
| “Emergency dentist [city] today” | absent | Three competitors | A directory and a map-style listing | medium |
| “Is teeth whitening safe?” | no local practice | Nobody local | National health publishers | scope review |
How the mention calculation works
Suppose all 40 answers in this fictional non-branded panel have been checked for business names: ten questions across four products, with one answer per combination for this example. The practice is named in 10 answers, giving 10/40 = 25%. The competitor bars use the same 40 answers. A real report also shows how many times each question was asked, keeps questions that name the practice separate, and says which answers could not be checked.
Fictional non-branded panel · 40 of 40 mentions assessed
Links to the practice’s own website are counted separately, out of the answers whose sources we could check. In this fictional example, 30 of the 40 answers could be checked and six of them link to the practice’s site: 6/30 = 20%. The other ten are reported as unknown, not as zero.
For a worked review of real law-firm pages, read our conveyancing page evidence review. It includes source links and the observation sheet; it is separate from this fictional sample.
Can AI read the site?
Google and Bing can see the site - checked through their own tools (illustrative).
A diagnostic request returned a 403 response - a lead for investigation. A request using a crawler’s name does not establish what the real crawler experienced or why a firm was absent (illustrative).
fig. - fictional examples of different evidence types. Actual crawler access needs corroboration from verified traffic or the platform’s own tools.
Search access and training permissions are separate. GPTBot and OAI-SearchBot have different roles; Anthropic also distinguishes training, search and user-requested access. Google Search uses its normal search requirements. Allowing a crawler does not guarantee a recommendation.
Is the business described the same way everywhere?
In this illustrative finding, the practice’s name differs across its website, its Google profile and a directory. We would verify the correct details and resolve material discrepancies. Public consistency checks do not establish how Google or an AI product represents the business internally.
The background code
In this fictional finding, the background code on two pages gives different answers from the text a visitor reads. That mismatch should be corrected. Links to verified profiles can make the identity claim easier to inspect; they do not prove how an AI product interprets it.
The fix list
First corrections to scope
- Investigate the 403 response with the hosting provider, verify whether real crawler requests are affected, and retest any agreed fix
- Correct material discrepancies in the practice’s name, address and contact details
- Make the background code match what the pages actually say
- Clarify the implant page, with clinical sign-off
Further investigation
- Review the directories and news sources the answers relied on; correct the practice’s profiles and look for genuine coverage
- Check that published reviews follow the platforms’ rules
- Ask the same questions again on a set date and compare names, links and accuracy separately
What happens next
The report and fix list are yours. They help you decide whether to make the changes yourself or use the Build and Watch services. Any later report shows what moved against your starting point, including nothing, where that is the honest result.
Everything in this specimen is fictional. A real audit keeps its dated evidence, so you can see exactly what changed later.
Last updated: 22 September 2026