How to measure AI visibility: rates, not scores
“¿Cómo se mide la visibilidad en IA?” Careful with this one, because the market will offer you a score. A dashboard with a number from 0 to 100, a position, a grade. Here is the problem with scores: they hide their inputs. You cannot audit them, you cannot reproduce them, and when the number moves you will not know why. Measurement you cannot take apart is decoration.
The alternative is older and better: rates. A count of a defined event over a defined population, printed with its denominator. Everything you need to measure AI visibility fits in that format.
The four rates
Ask a bank of real buyer questions to several AI engines, several times each, and read every response. For each brand, count four things:
- Mentioned (mencionada). The brand appears in an answer to a question that never named it. The floor: presence in the category conversation.
- Recommended (recomendada). The assistant actively suggests it. Being listed among eight is not being told to buy it.
- Named as the best (señalada como la mejor). The answer distinguishes it instead of grouping it. The strongest and rarest form of presence.
- With a purchase path (con ruta de compra). The answer says where to buy it. The closest honest proxy to purchase intent, with no analytics needed.
Each one is a rate with its denominator. From our invented sample, a sunscreen brand measured against Bloqsol and Dermalux: Mentioned 26%, 103 of 396 responses. Recommended 11%, 44 of 396. Named as the best 3%, 12 of 396. With a purchase path 7%, 28 of 396.
Marca inventada, competidores inventados, cifras inventadas. La estructura es real; los números no son de nadie.
A percentage with no population behind it is a decoration. 26% of what? Of 396 responses, each one read. Now you can audit it.
The three layers, never averaged
The bank has three layers, and they answer different questions:
- Unbranded questions never name the brand. They measure spontaneous presence, the only honest way to know if you appear on your own.
- Brand-named questions ask about you directly: what the model believes, where it sends buyers, which critique arrives attached.
- Competitive questions map who owns the conversation, head-to-head and in comparisons that happen without you.
Each layer gets its own numbers. Mix them and every figure moves for reasons nobody can explain. Averages across layers are where scores go to hide.
The three assistant types
The same brand measures differently depending on how the assistant works, so the types are reported separately:
| Assistant type | Mention rate (invented sample) |
|---|---|
| Memory-only call | 9%, 12 of 132 responses |
| Search-then-answer | 22%, 29 of 132 responses |
| Google AI Mode | 31%, 41 of 132 responses |
The reading: awareness inside the model’s memory is weak; the open web is what carries the brand. Blend the three into one number and you lose exactly that insight. One more rule: assistants that show no sources are recorded as their own status, never counted as zero, because counting them as zero silently corrupts every comparison.
Marca inventada, competidores inventados, cifras inventadas. La estructura es real; los números no son de nadie.
Why this beats a score
Three practical consequences:
- You can audit it. Every figure traces to saved responses you can read.
- You can reproduce it. The same frozen bank re-run gives comparable numbers, so trends mean something.
- You can act on it. “Recommended is low, search-then-answer is carrying us, the retailer page is the top cited source” is a plan. “Our score went from 61 to 64” is a shrug.
The full method, including how the bank is built and what the study cannot tell you: /metodologia/.
Next step: The methodology