Case study · AI visibility
One brand. Four assistants.
Four different answers.
We put one consumer brand through ChatGPT, Claude, Gemini and Perplexity on a fixed set of 42 category questions in the same week. The blended visibility score came out at 51. That number was wrong for every one of the four assistants.
By Sri · Run date 2026-09-09 · Last reviewed: September 2026
Setup
42 questions, 30 of them unbranded.
The question set is fixed in advance and split two ways. Twelve questions name the brand. Thirty describe a need without naming anyone, the way a person who has never heard of the brand would ask. Each question goes to all four assistants. We record whether the brand is named, where in the answer it appears, and which other sources are named instead.
The split matters more than anything else in the method. A branded question measures recall of a name the person already typed. An unbranded question measures whether the brand gets named when someone wants a solution. Only the second has anything to do with acquiring a customer.
Finding one
A 55-point spread across assistants.
Same brand, same questions, same week. Perplexity named it in 73% of answers. Claude in 18%. If your buyers use Claude, a report saying "51" is telling you about a world they do not live in. A blended score is an average of four different realities, and averaging is exactly the operation that destroys the information a brand needs.
Finding two
Branded questions add 17 points that do not exist.
Nobody is lying. Both numbers are real. But a report that mixes them, weighted toward branded prompts, drifts upward every time brand awareness improves while the commercial reality stays flat. Ask any vendor what share of their prompt set names your brand. If they cannot answer, the score is not measuring what you think.
Finding three
Well documented, not recommended.
| Intent | Named |
|---|---|
| Attribute | 8/12 67% |
| Reputation | 5/8 63% |
| How to choose | 2/2 100% |
| Comparison | 2/4 50% |
| Alternatives | 2/4 50% |
| Recommendation | 2/8 25% |
| Use case | 1/4 25% |
Named in 8 of 12 attribute questions. Named in 2 of 8 recommendation questions. The assistants know what this brand's products are made of and how they behave, and they do not reach for it when someone asks what to buy. That is a specific diagnosis with a specific fix, and it is not "publish more content." It is being the source that gets cited on the questions where a purchase decision is being made.
What we did not do
No trend line.
We ran this brand once before, in July, on a different and smaller question set. The two runs are not comparable, and a "named rate fell from X to Y" line would mix a real change with a sample-size change. We only compare batch to batch on an identical set. If a report shows you a trend and cannot tell you the question set was held constant, it is not a trend.
This is what BrandPulse produces.
A fixed question set for your category, four assistants, per-assistant and per-intent results, the verbatim sentences, and the cited sources. Never only a blend.