How BrandScope measures AI visibility
Every week we ask the same set of buyer questions to ChatGPT, Claude, Perplexity, Gemini and Google AI Overviews — and we ask half of them without ever naming your brand.
That split is the whole point. Most tools only test the second kind of question, which every brand passes. The first kind is where buyers actually are when they have never heard of you.
Two kinds of question, scored separately
| Discovery (unbranded) | Brand (aided) | |
|---|---|---|
| What we send | The question only — no brand name, no domain, no hint | The question plus your brand name and domain |
| Example | "best AI visibility tools for a SaaS startup" | "How does BrandScope compare to Profound?" |
| What it measures | Whether a stranger could find you at all | Whether AI recommends you once the buyer already knows you |
| Typical result | Most brands: 0–20% | Most brands: 60–90% |
| Weight in the score | 60% (Discovery is the headline number) | Not scored — shown for context |
If you see Discovery 0% next to Brand 85%, that is not a contradiction — it is the normal state of a new product. AI knows you when prompted, and never volunteers you.
What counts as a fair denominator
An answer that names no brand at all (a pure explanation, "here is how AI search works") is not evidence that you are invisible — nobody was recommended. We exclude those answers from the denominator, the same way Profound does. So do answers where Google returned no AI Overview at all.
What stays in the denominator is every answer that recommended someone. If AI named three competitors and skipped you, that counts as a miss.
The AI Visibility Score (0–100), scoring v3
| Factor | Weight | What it measures |
|---|---|---|
| Discovery | 60% | Share of unbranded buyer questions where AI names you |
| Accuracy | 25% | When AI does name you, how often the description is correct |
| Share of Voice | 15% | Of "best/top" questions that named you, how often you rank #1 |
We removed two factors in v3: citation independence and source freshness. Both measure the quality of the evidence behind a mention, not whether buyers can find you. Keeping them inside the score let a brand with zero discovery still score 66/100. They are still shown in the report — just not folded into the headline number.
The other numbers in your report
Cited Mentions
Of the answers that named you, how many cite an independent third-party page. AI citing your own site is excluded — quoting you proves nothing about what anyone else thinks of you.
Sentiment
The tone of each mention: positive (recommends you), neutral (factual), negative (warns buyers off). Judged in the same pass as accuracy, so it costs no extra AI calls.
Average Position
When AI names you, where you land. #1 is the first brand named; #5+ means you are an also-ran.
Who AI names instead of you
On unbranded questions, how often each known competitor appears. This is the gap: buyers who never saw you.
AI Coverage
How many of the five core surfaces know you at least once. DeepSeek is collected and shown separately — its coverage of Western brands is low and unstable, so including it in the score would just halve everyone's number.
Why every percentage has a ± next to it
With ~12 unbranded answers per run, a number like 8% is not precise. We show a 95% confidence interval (Wilson score) beside each rate, so you can see how much of a change is real.
If two runs differ by less than their intervals overlap, treat it as noise. When a cell has fewer than 5 answers we say "too few to bound" instead of printing a fake margin.
Known limits — read these before trusting a number
Brand names collide
If another company shares your name, an AI answer may be about them. We check the cited source and the surrounding context, and we flag mentions we cannot attribute to your domain. This is the single biggest source of error in this category — Profound itself gets confused with Profound Logic.
API calls are not the consumer app
ChatGPT, Claude, Gemini and Perplexity are called through their APIs, not the web apps. Answers can differ by 20–40pp. Only the Google AI Overview column is a real SERP capture.
Sample size is small by design
A larger question set costs proportionally more per run. We would rather publish 12 honest answers with intervals than 60 answers at a price nobody pays.
None of our competitors show you their error bars — we do
Every number here ships with its denominator and a 95% confidence interval, and the dashboard tells you how wide the run-to-run noise band is. As far as we can tell, Profound ($99–$399/mo), Peec ($95–$495/mo) and Otterly ($29–$489/mo) all report point estimates without a margin of error. That makes their numbers look more certain than the underlying measurement allows. Pick any of them and ask "how much does this move when nothing changes?" — the dashboard will not answer.
Numbers are only comparable within one scoring version
Every run records its scoring version. We never diff a v3 number against a v1 number — that produces fake collapses like "down 47pp".
Data sources
- ChatGPT, Claude, Perplexity, Gemini via API (OpenRouter), plus Google AI Overviews via a live SERP capture.
- DeepSeek is also collected. It is shown in the per-engine breakdown but excluded from the score, because its coverage of Western brands is low and unstable.
- Raw answers — every answer behind every number is visible in Mention details, with its source URL, so any claim here can be checked.
Questions? Email support@brandscope.dev.