The AI visibility study: where your brand stands inside AI answers, with the evidence attached
A buyer asks ChatGPT, Gemini or Perplexity for the best option in your category. The answer names three to five brands. If yours is not one of them, no ranking report will show it, and most teams find out by accident. The study answers the question directly: “¿ChatGPT conoce mi empresa?” One run tells you where your brand stands inside AI answers, with the evidence attached: the saved responses, the counts with their denominators, and the sources behind every answer.
What the study answers
Four questions, in this order:
Does the category conversation include you? Unbranded questions about your category, asked the way buyers actually ask them. If your brand appears without being named, the study counts it. If it does not appear, you finally know it.
Who gets recommended instead? The competitors the models name, including the ones your sales team does not track. Models routinely treat a premium brand as interchangeable with the private label next to it on the shelf, so the study measures the rival set the models see, not the one your team respects.
What gets said about you? When someone names your brand, what comes back: what the model believes about it, which critique arrives attached, and where it sends buyers.
Which sources move the answers? Every site the models cite, deduplicated and ordered by who can actually influence it. This is usually where the work is: most of what an assistant says about your category was written by someone else.
What you receive
The deliverable is data first, document second:
The frozen prompt bank. The instrument, in its three layers (unbranded, brand-named, competitive), documented so it can run again identically.
Every response, saved and readable. The full text of every answer from every model stays in the study. You can read what your buyers would read. No summary stands between you and the raw output.
The four outcome rates, each printed with its denominator. Mentioned: does the brand appear in an answer that never named it. Recommended: does the assistant suggest it, or just list it alongside others. Flagged as the best: is it distinguished instead of grouped. With purchase route: does the answer say where to buy it. A rate without its population behind it is decoration, so every figure in the report looks like this: Mentioned 26% (103 of 396 responses).
Share of category mentions. Your brand against each competitor, by layer and by model type. The layers are counted separately and never averaged, because mixing them is how a number stops meaning anything.
The ordered source list, with owners. Every cited domain, how often it appears, and who controls it: the regulator page you cannot edit, the retailer page you can rewrite, the earned media you can pursue.
A working session with priorities. What the numbers say, what to do first, and which team owns each action. The goal is that your people can defend these numbers in a meeting where we are not in the room.
There is no “executive summary” layer. The document exists to carry the evidence, and the session exists to turn it into decisions.
What we need from you
Little:
- One working session, the levantamiento, our kickoff research
- Category context
- A competitor list
- Your owned domains
The levantamiento asks three questions that do most of the work: does the money arrive through your website or through a shelf, what decision would you make differently based on the result, and who you compete with, including the brands the models treat as interchangeable with yours even if your team does not.
Search Console queries, call center FAQs and internal site search sharpen the aim. They are welcome and not required. They inform which themes get asked; they do not decide whether the number is trustworthy.
How the study runs
Six phases:
- Levantamiento. One session to set the scope: how the money arrives, what decision the study must inform, and the competitor set, including who the models treat as your rival.
- Bank written and reviewed. We draft the prompt bank from your category and from documented demand: search queries, autocomplete, forums, reviews. You review it. Review means checking that the questions sound like your customers and no theme is missing. It does not mean rewriting them toward the brand; that is the one change that would invalidate the result.
- Freeze and run. The approved bank freezes and becomes the instrument. It runs on the chosen models, with repetitions, and the number of observations and the cost ceiling are known before the first prompt is sent.
- Reading and interpretation. Every response is read and coded the same way: mention or recommendation, subbrand attribution, sources ordered by who can actually move them. This is the slow part, and it is where the quality lives.
- Report and working session. We walk through what the numbers say and what to do first. Each action leaves with an owner on your side.
- Second wave. After your work has had time to publish and get indexed, the same frozen bank runs again. Same instrument, so the runs compare directly. If nothing moved, the report says nothing moved. For most clients, this second run is the first monitoring cycle.
Study size is a range, not a package
A study has three knobs: questions x models x repetitions. There is no standard configuration, because the right size depends on the decision the study must inform.
The range runs from a small directional sample (a few dozen prompts, fewer models, fewer repetitions, enough to see where you stand) to a deep study (around 100 prompts, 4 or more models, 5 repetitions, enough to compare segments and models with confidence).
One illustrative configuration, labeled as such: 60 questions x 3 assistants x 3 repetitions = 540 responses read. It is an example of the arithmetic, not a default.
Cost scales with the same three knobs, which is why we quote by scope: [PRECIO POR DEFINIR]. The size is agreed in the levantamiento and priced before the first prompt is sent, so the scope conversation happens before the money is spent, not after.
Sample reports: invented brands, real structure
Two sample one-pagers show the report structure, one for each kind of brand we serve:
Solaria, consumer. A sunscreen sold in Mexican pharmacies. The one-pager carries the four rates with denominators, share of mentions by brand, mention rate by model type, the ordered source list, and a closing page on where to act first. Its source list is the moment measurement becomes a to-do list: the regulator page that cannot be displaced, the retailer page that can be rewritten this quarter without asking permission, the earned media to pursue. Read it: the consumer sample one-pager.
The B2B example. Same structure, different reality: a considered purchase with a long cycle, a buying committee that asks AI engines early, competitive loss inside evaluation questions, and reputation questions about the brand. Read it: the B2B sample one-pager.
Both samples carry the same disclaimer, verbatim:
Marca inventada, competidores inventados, cifras inventadas. La estructura es real; los números no son de nadie.
We keep the disclaimer because the structure is what matters, and an invented brand raises no confidentiality questions.
Questions
Can you guarantee ChatGPT will recommend my company?
No. Nobody can. Answers vary by user, by context and from run to run; one screenshot is an anecdote in both directions. What we commit to is serious measurement, and work on the factors you control: the sources the models cite and what they say about you.
Do I need llms.txt or special schema?
No. Google’s own documentation says optimizing for generative AI search is still SEO, and names llms.txt, chunking and special structured data as myths. What you actually need: content the models can retrieve, authority, and presence in the sources they cite. The study tells you which sources those are in your category.
Is this not just SEO?
Mostly, optimization is SEO done well. We agree, and Google agrees. What is genuinely new is measurement: whether AI systems recommend you, against whom you compete inside the answers, and which sources they trust. No traditional SEO tool reports any of that.
How long does it take?
[COMPLETAR: timeline from levantamiento to report session]
What happens after the study?
The study is the baseline. Monitoring re-runs the same frozen bank on a cadence, so you learn whether what you are doing works. A study tells you where you stand; monitoring tells you whether it is working. See the monitoring page.
How much does it cost?
It depends on the three knobs, and you will know the number before any work starts: [PRECIO POR DEFINIR]. Structure and scope are on the pricing page.
Request the study
Start with the question, not the tool. Tell us your category and the decision you need to inform. If the study fits, we scope it in the levantamiento and you see the price before the first prompt runs.