How often AI assistants name each operator
We put the same 40 standard questions to six AI assistants and Google Search — 280 answers per run — and count how often each operator is named. We repeat the run monthly; this one was read 7–8 Oct 2026. This is a presence count. It is not a quality ranking, and being named is not an endorsement.
40 questions × 7 engines, asked through APIs
On 7–8 Oct 2026 we put the same 40 questions to ChatGPT Search (gpt-5-search-api), Perplexity (sonar), Gemini (gemini-3.8-flash), Claude (claude-sonnet-5-5), Grok (grok-4.7) and DeepSeek (deepseek-flash) through their official APIs, and to Google Search through a search-results API. Each of the 280 answers is checked by software against the names and domains of every operator we track; a link whose address contains an operator's name counts as naming it. 18 of the 40 questions name an operator. When the answer names that same operator, it is counted separately — the grey +N — and not in the main number. Gemini, Claude and Grok decide for themselves whether to search the web — they searched for 19, 38 and 40 of their 40 answers. DeepSeek's API has no web search, so its answers come from what the model learned in training. Google showed no AI Overview for any of these questions, so the Google Search column counts the operators named on the first page of results.
The set is frozen and archived. Every future cycle runs the same 40 questions, so results stay comparable over time. We publish the themes and their weight, not the exact wording — published wording would let operators optimise for the measurement instead of for players.
| Cluster | Questions | What it asks |
|---|---|---|
| Category discovery | 12 | Open questions about which operators lead the category, with no brand named in the question. |
| Brand trust | 12 | Whether a specific named operator is considered legitimate and safe. |
| Withdrawals | 8 | How fast operators pay, what delays payouts, what to expect. |
| Head-to-head | 4 | Direct comparison between two named operators, or alternatives to one. |
| Industry data | 4 | Who tracks and publishes operator data — licences, payouts, complaints. |
40 questions × 7 engines = 280 answers per run, each checked against every operator.
- Answers change with the moment, the region and the account — the same question can be answered differently an hour later or from another country. Our requests are sent from a server in Finland.
- Answers come through the providers' APIs, not the apps people use, so they can differ from what a person sees in the app.
- Our July 2026 baseline was read by hand in four assistants' web interfaces. From October the run uses APIs and more assistants, so a change against July can come from the method, not only from the month.
- Google showed no AI Overview for these questions, so its column measures the first page of Google results, not an AI answer.
- DeepSeek's API cannot search the web, so its column shows what the model knows from training, not today's web.
- Each run is a snapshot of one day, repeated monthly — not continuous monitoring.
Being named is not an endorsement. This measures visibility in AI answers and search — not operator quality, safety or licensing.