How to measure iGaming visibility in AI search without misleading metrics
AI answers change from one run to the next, so the method matters more than the tool. Here is the prompt-panel method we use, and the numbers we refuse to report.
To measure iGaming AI search visibility, run a fixed set of prompts on a schedule, from inside each target market, across the assistants your customers use, and record for every answer whether your brand is mentioned, whether it is linked, whether the facts are right and who else appears. Report rates and trends over the whole panel. Never report one answer as a ranking.
That last rule is the one most dashboards break, so we will start with why.
Why single answers mislead
AI assistants generate each answer fresh, influenced by wording, location, account history, model version and chance. In one study of nearly 3,000 runs, ChatGPT and Google's AI gave the same list of brands less than 1% of the time. Any measurement has to treat answers as samples, not positions.
The study, by SparkToro and Gumshoe, ran 12 prompts between 60 and 100 times each per platform in late 2025. The same brands in the same order appeared less than 0.1% of the time. When volunteers wrote their own prompts for the same need, the prompts barely resembled each other. The study was partly funded by a tracking vendor and is not peer-reviewed, but the direction matches what anyone running repeated prompts will see.
So "we rank third in ChatGPT for best Michigan casino" is not a meaningful sentence. "We were named in 9 of 20 Michigan prompts this month, up from 5" is.
Building the prompt panel
A panel is a frozen list of prompts, each tagged with a market and a type. Freezing matters: if you change prompts every month, you cannot tell whether the trend comes from your work or from your prompt edits.
We mix three types. Branded prompts ask about you directly ("is [brand] licensed in New Jersey", "how long do [brand] withdrawals take"). Category prompts ask the question a new customer would ("which online casinos are legal in Pennsylvania"). Comparison prompts pit you against named rivals. Write them the way people speak, which is usually longer and messier than a keyword.
Size depends on how many markets you track. Twenty prompts can cover one market thinly; sixty covers two properly; one hundred and twenty covers four. Below about twenty, a single odd answer swings the percentage too much to read.
How to run it
Run each prompt from a location inside its market, because answers about gambling legality depend heavily on where the user is. Use logged-out sessions where the platform allows it, or a clean account with no history. Run on the same days each month, and record the platform, model or mode (AI Mode and AI Overviews are different products), date and location with every answer.
If you can afford it, run each prompt more than once per cycle and count the share of runs in which you appear. Three runs per prompt reduce the effect of any single odd answer.
What to record for every answer
| Field | Why it matters |
|---|---|
| Mentioned (yes/no) | The headline rate: share of prompts where your brand appears |
| Linked to your domain (yes/no) | Separates visibility that can send traffic from visibility that cannot |
| Facts correct (yes/no, with notes) | Wrong licence or launch information is a compliance problem as well as a marketing one |
| Competitors named | Shows who the assistant treats as the default answer |
| Domains cited | Tells you which third-party pages shape the answer, and so which to check or correct |
| Platform, mode, date, location | Without these, nobody can reproduce or audit the result |
Numbers we will not report
We do not report an "AI ranking position", a single blended "AI visibility score" with no breakdown, or any figure without the prompt count behind it. We also do not convert mentions into estimated traffic or revenue. There is no reliable conversion: Google's and Bing's own AI reports show impressions or citations, not clicks.
And we never present a result without the raw answers available. If a client's head of compliance wants to see exactly what ChatGPT said about their Ontario licence on 3 October, they should be able to.
Adding first-party data
Prompt panels are your own sample. Platforms now offer a little data of their own, which is worth adding even though it is limited. Search Console's generative AI performance reports show AI impressions by page and country. Bing Webmaster Tools reports citations in Copilot. ChatGPT tags most referral links with utm_source=chatgpt.com. Our companion article on Search Console and Bing AI reports explains what each shows and where it falls short.
Reading the trend
Look at three months before drawing conclusions, and check the platform change log before crediting or blaming your own work. When Google or OpenAI changes a model, everyone's numbers move. If your mention rate rises while competitors' fall over the same period, something you did probably helped. If everyone moves together, the platform changed.
Start small: write twenty prompts for your most important market this week and run them once. You will learn more from reading those twenty answers than from any dashboard. If you want the panel built and run for you, that is our LLM brand visibility service.
Sources
- Search Engine Journal: AI recommendations change with nearly every query (SparkToro and Gumshoe) (30 January 2026)
- Search Engine Journal: Search Console AI reports rolled out worldwide (31 August 2026)
- Search Engine Land: Bing Webmaster Tools AI Performance report (February 2026)
- OpenAI Help: publishers and developers FAQ