By Samuel Anan | Generative AI Leader | London, UK
An AI visibility report should tell a client what was checked, what appeared and what decision the evidence supports. It should not compress a sample of changing answers into an unexplained score or imply that one favourable response is a stable ranking.
I would begin with the client's business question, document the sample in plain language and keep mentions, citations, recommendations, accuracy and business outcomes separate. That makes the work inspectable even when the conclusion is uncertain.
Open with the decision, not the dashboard
State the customer situation being examined. For example: when a London buyer asks which SaaS onboarding tools support a particular workflow, is the client's product described accurately? That is narrower and more useful than asking whether the brand is "visible in AI".
Define the part of the work being measured:
- Factual accuracy: does an answer describe the offer, availability or limitation correctly?
- Branded retrieval: what appears when the question names the business?
- Unbranded discovery: does the business appear in a relevant category question?
- Source visibility: is a page from the site shown as a source?
- Recommendation: is the brand presented as a possible choice, and on what stated basis?
Generative engine optimisation (GEO) aims to improve how accurately and usefully a brand's information can be discovered and represented in generative answers. Answer engine optimisation (AEO) focuses on making answers clear and supportable across answer-led search experiences. Neither term turns a sampled answer into proof of universal visibility.
Show exactly what the sample covers
Record the dates, services or surfaces checked, exact questions, relevant settings and number of successful runs. Log failed or missing runs rather than silently changing the denominator. If the method changed from the previous period, say so before showing a trend.
Keep the underlying answer, visible sources and capture date in an internal evidence record. A client should be able to inspect at least one complete example.
Microsoft's current AI Performance documentation for Bing Webmaster Tools is a useful reminder that citation measures have boundaries: its citation and cited-page measures describe appearances across supported experiences, while Microsoft states that they do not indicate ranking, authority, importance or a page's role within an individual answer. A first-party platform report and a manually sampled question set should therefore remain separate rows.
Google Search Console's Generative AI performance report provides a different first-party view: impressions for links shown in supported generative AI features on Google Search, with page, country, date and device dimensions. That impression record is not the same as a manual count of mentions or recommendations, so label it by platform and definition.
Keep appearance categories distinct
Use a worksheet that preserves the evidence behind each classification:
| Question | Illustrative: "Which tools help a UK SaaS team document onboarding?" |
|---|---|
| Surface | Named AI surface |
| Date | Sample date |
| Successful runs | 12 |
| Mention | 4 |
| Citation | 2 |
| Recommendation | 1 |
| Accuracy | 3 accurate, 1 outdated |
The figures above are illustrative, not observed SamuelAnan.com or client data. A single answer may belong to several categories, so the columns should not be added together as independent successes. The denominator also matters: four mentions in 12 completed runs is different from four mentions in a sample whose failed runs were removed without explanation.
An appearance can be favourable and still be wrong. If a brand is recommended using an outdated service description, report the recommendation and the error together.
Use the same saved question set, collection conditions and dated evidence in every reporting period so the comparison can be repeated and checked.
Report errors as work, not embarrassment
List consequential inaccuracies, stale sources and unsupported reasons given for a recommendation. Separate information the client controls from third-party pages it can only request to have corrected.
A practical action might be:
Inspect the cited service page, confirm the current eligibility rule with the product owner, correct inaccurate copy the company controls and repeat the defined sample on the scheduled review date.
Do not promise that correcting one page will immediately change generated answers. The useful immediate outcome is a more accurate public source and a documented follow-up.
Separate referrals from later outcomes
An AI answer observation is not a website session, and a session is not automatically an enquiry or sale. Report these in separate sections with their own sources and definitions.
Google Analytics' Traffic acquisition report is session-scoped and helps explain where website visits came from. It does not establish every influence on a later purchase or recover visits for which the source was not recorded.
The report can therefore say that sampled answers became more accurate, recorded AI referrals were unchanged and business impact is not yet established. Those statements can coexist.
End with a short decision list
Finish with no more than a few actions, each with an owner, evidence route and review date:
- Correct a verified factual error on an owned page.
- Investigate a source that repeatedly supports an outdated claim.
- Improve a useful page that does not answer the sampled buyer question.
- Continue the same sample when the period is too short for interpretation.
A wider SEO and GEO client report can hold these sections. The important point is to define the AI observations honestly so the report does not confuse visibility with traffic or business performance.
A good AI visibility report does not remove uncertainty. It shows the client where the uncertainty sits and what can reasonably be done next.
About Samuel Anan
Samuel Anan is a London-based generative AI leader working across AI search, generative engine optimisation (GEO), answer engine optimisation (AEO), workflow automation and agentic AI. He writes for startups, SaaS teams, and e-commerce businesses selling through Amazon and Shopify.