The measurement mistake to avoid
AI visibility screenshots are easy to sell and difficult to audit. A brand may be named without a link, cited as a factual source without being recommended, or recommended from information supplied by a third party. Combining those events into one “AI share of voice” number hides what actually changed.
Google has expanded links in AI search while leaving important reporting limitations. This makes a layered measurement model more useful than pretending that every exposure can be attributed like a conventional click.
Use five observable stages
| Stage | Question | Evidence to retain |
|---|---|---|
| Retrieval | Was a controlled or third-party source used? | Cited URL, source domain and capture date |
| Citation | Was the source visibly linked or attributed? | Answer capture and exact citation position |
| Mention | Was the brand or product named accurately? | Prompt, response and factual assessment |
| Recommendation | Was the brand presented as a suitable option? | User context, qualifiers and competing brands |
| Visit | Did the exposure produce a measurable session? | Landing page, referrer where available and conversion path |
Accuracy sits across every stage. A positive-looking mention with the wrong market, license or payment information is a brand risk, not a success.
Build a prompt set that can be repeated
Create prompts from real decision journeys: discovery, comparison, verification, troubleshooting and branded navigation. Include market and language context where relevant. Record the system, model or surface, signed-in state if material, location, date and prompt wording.
Do not report one run as a stable result. Answers can vary, and personalisation or product updates may change the source set. A scheduled sample reveals direction; it does not prove universal visibility.
Connect AI evidence to normal search data
Review AI observations beside indexed coverage, rankings, branded search demand, direct traffic, assisted conversions and lead quality. Search Engine Journal’s demand SEO framework similarly argues for broader patterns when click-level attribution is incomplete.
For iGaming, add market accuracy and compliance review. A mention can create harm if it implies availability where the product is restricted or cites outdated bonus terms. Human review remains part of measurement, not an optional editorial step.
What a credible report looks like
A useful monthly report shows the fixed prompt sample, the share of answers by stage, which domains were cited, factual errors, new source gaps and the actions taken. It also states what cannot be concluded. That is less theatrical than a single visibility score and far more useful for deciding where to invest.
Explore AI Citations for the evidence and source workflow, or read why AEO and GEO remain part of SEO.