We benchmarked our own AI visibility. The honest number was 0%.
· by boneLau
Short answer: our own product scored BrandScope 87/100 — "Strong AI visibility, you consistently rank #1". Then we searched our category in Google and the AI Overview did not mention us at all. The tool wasn't broken. The measurement was: we were putting the brand name inside every prompt and reporting the result as if buyers would see it. Most AI visibility scores in this category are inflated the same way. Below is the exact prompt we were sending, what three major vendors say about it, and what our numbers looked like once we removed the crutch.
What we were actually measuring
Every week our tool sends five questions to ChatGPT, Gemini, Perplexity and Google's AI Overviews. Here is the prompt we were sending for every single question:
| Field | Value |
|---|---|
| Question | What are the top AI visibility tools in 2026? |
| Focus brand | BrandScope (brandscope.dev) — AI visibility, GEO |
| Context | AI mention monitoring tool that tracks how AI assistants answer questions about your brand |
| Comparison targets | Profound, AI Directories, Goodie |
| Instruction | If the brand appears in your answer, say where and why it is recommended |
Read the last row again.
We were not asking a question. We were telling the model we existed, describing ourselves, naming our competitors, and asking it to justify a recommendation. The model complied — of course it complied.
That measures compliance, not visibility.
We call this an aided prompt: the brand name is in the prompt. What buyers actually do is unaided: they type a category question and never mention you, because they don't know you exist.
We were running 100% aided prompts and presenting the result as unaided.
Three major tools already say this in their own docs
This is not a novel insight — it is in the vendors' documentation.
Profound separates the two explicitly: "Branded prompts — prompts that mention a brand directly… Unbranded prompts — prompts that ask broader category questions and may surface a brand naturally." Its published benchmark work uses 50 unbranded prompts per category.
Peec AI frames it as a measurement property rather than a preference: "Branded prompts are decided, non-branded prompts are not. Ask about a company, and the engine mentions it run after run." Its guidance tells users to tag branded prompts separately so they don't inflate the headline score.
Otterly recommends tracking 50 to 100 non-branded bottom-of-funnel prompts per country, language and category.
The consistent guidance across the category is 80–90% unbranded prompts, with branded prompts isolated for perception analysis.
We were at 100% branded and reporting one blended number that hid it.
What we changed
Two changes, both borrowed from how Profound measures.
1. Discovery questions contain no brand context at all. For "best X tools" style questions the model now receives the question and nothing else — no brand name, no domain, no description, no competitors.
2. Answers that mention no brand are removed from the denominator. This is Profound's approach: mentions you divided by answers that mention at least one brand. If the AI answers an explanatory question without recommending any product, that is not evidence you are invisible — it is a question with no winner. Counting it against you deflates the score and makes tools incomparable. In our first clean run, 27 of 52 answers named no product at all.
Comparison and objection questions ("BrandScope vs X", "Is BrandScope worth it") keep the brand context. Those measure something different and legitimate: when a buyer already knows your name, does the AI describe you correctly? That is accuracy, not discovery.
The honest numbers
Same brand, same engines, same week.
| Before (aided) | After (unaided) | |
|---|---|---|
| Discovery — buyers who don't know you | 89% | 0 / 18 |
| Brand — buyers who already know your name | 100% | 10 / 12 |
| Mention rate, all core questions | 93% | 42% |
Discovery: zero out of eighteen. Three category questions, three engines, two samples each. Not one mention.
Brand: ten out of twelve. When the question contains our name, the AI describes us accurately and often recommends us.
Both numbers are true. They describe two different buyer situations. Collapsing them into one score is how a brand ends up with 87/100 while being invisible to every buyer who has never heard of it.
The most uncomfortable finding
Google's AI Overview did mention us. Once.
The source it cited was otterly.ai — a competitor's comparison page. Not brandscope.dev.
We were visible in Google's AI answer only because somebody else had written about us.
This is the part of GEO most tools skip, because it cannot be fixed by editing your own website. AI Overviews retrieve third-party pages. Your site is evidence, but rarely the only evidence. For a small brand, appearing in AI answers has less to do with what you publish and more to do with who else has published a page that mentions you.
It explains the asymmetry we kept seeing: our site was perfectly crawlable (llms.txt, structured data, clean canonicals) and assistants could find it when asked directly — but the AI Overview, which needs an external source pool, had nothing to cite.
What to do with this
1. Ask your tool one question: is this score aided or unaided? If the answer is "we don't distinguish", treat the number as marketing. No single number can carry both "will a stranger find me" and "does the AI describe me correctly".
2. Measure discovery separately, and expect it to be low. If you are a small brand your unaided number is probably near zero. That is not a broken tool and not a reason to panic-buy content — it means you have no third-party surface area yet.
3. Work on the pages AI actually cites. For the questions you care about, look at which domains the AI Overview cites today. Those domains are your real to-do list. In our case it was a competitor's comparison page — another post on our own site would not have fixed it.
4. Distrust any single score, including ours. We kept the blended number but demoted it. Our dashboard now shows Can buyers find you? (Discovery) and When they know your name? (Brand) side by side, with sample sizes attached. If a percentage has no n=, it is a vibe.
How this was measured
- 5 questions x 4 core answer surfaces (ChatGPT, Gemini, Perplexity, Google AI Overviews) x 2 samples, plus two Asian assistants reported separately.
- One run per week. AI answers vary run to run; roughly plus or minus 10 percentage points is normal noise.
- Google only generates an AI Overview for a subset of queries. Queries with no overview are excluded from the denominator rather than counted as misses.
- The Discovery and Brand figures above come from a single run on 2026-10-20. One data point is not a trend; we publish the trend separately.
BrandScope measures how AI assistants describe your brand across ChatGPT, Perplexity, Gemini and Google AI Overviews, verifies every citation against your own domain, and reports week-over-week change. Yes, we sell this. That doesn't make the numbers above less true — they are the reason we changed how we measure. You can [run one brand free](https://www.brandscope.dev/) and see your own discovery number before paying anything.