A visibility score is only useful if you can trust it. The analysis step uses AI to extract mentions from AI answers, and extraction models can hallucinate, so the pipeline is built with independent code-level checks around every claim. These are the guardrails.
Verbatim mention verification
Every mention the analysis model claims is re-checked in code against the raw answer text. If your brand name does not literally appear in the answer, the mention is discarded, no matter how confident the extraction was. This means the checker never credits inferred mentions like "tools in this category" or paraphrases of your brand. The trade-off is strictness: if AI engines refer to you by an abbreviation or an old name, those answers count as misses. Scan the brand name your market actually uses.
Evidence attached to every claim
Each counted mention ships with the exact excerpt where your brand appears, plus a per-engine mention matrix. You never have to take the score's word for it: you can open any query and read what each engine actually said.
Pinned engines, not a race
The GrowthGPT platform normally races multiple models in parallel and takes the fastest answer, which is right for generation tools. This tool deliberately does not. Each engine leg is pinned to one specific vendor's model, because the whole point is measuring what each vendor's AI says about you. A fallback that silently swapped GPT for another model would corrupt the per-engine results.
Failed queries are excluded, not counted as misses
Live model APIs occasionally fail or time out. When an engine fails to answer a query, that sample is excluded from that engine's statistics rather than scored as a zero. Your score reflects answers we actually observed, never gaps in our infrastructure. The report's per-engine coverage shows how many answers each engine contributed.
Controlled, repeatable conditions
Every query is asked with no personalization, no chat history, and no session context, so results are comparable across brands and across time. This is also why your score can differ from what ChatGPT tells you in your own account: consumer assistants mix model knowledge with browsing, memory, and your history, while the checker measures the underlying model families under identical conditions for everyone.
Known limitations
- Sampling, not a census. 25 queries across 4 engines is a structured sample of an effectively infinite query space. The score is designed for comparability and trend tracking, not as a claim to have measured every possible question.
- Models move. Vendors update models and the live web shifts, so scores drift over weeks even if you change nothing. Compare scans month over month, not hour over hour.
- Very short brand names. Names shorter than four characters are too ambiguous for strict literal matching, so for those the checker relies on the analysis model's judgment instead of the verbatim check.