A buyer asks ChatGPT, Gemini or Perplexity for the best option in your category. The answer usually 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: “Does ChatGPT know my company?” 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. In our reading, models often 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:
Data
The frozen prompt bank
The instrument, in its three layers (unbranded, brand-named, competitive), documented so it can run again identically.
Data
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.
Data
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. Named as the best: is it distinguished instead of grouped. With a purchase path: 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).
Data
Share of category conversation
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.
Data
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. The list is also the handoff: if you want the sources worked rather than just named, that is the work on the sources, scoped against your own report.
Session
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:
Required
- One working session, the kickoff research
- Category context
- A competitor list
- Your owned domains
Welcome, not required
- Search Console queries
- Call center FAQs
- Internal site search
They sharpen the aim. They are welcome and not required. They inform which themes get asked; they do not decide whether the number is trustworthy.
The kickoff 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.
How the study runs
The short version, for the buying decision:
- One working session sets the scope (the kickoff)
- Bank built from your category and from documented demand
- You review it
- It freezes
- Runs on the chosen models, with repetitions
- Every response read and coded
- Report + working session
- Same bank runs again
The freeze is the pivot: everything after it is comparable only because the bank stopped changing. The report leaves each action with an owner on your side, and the same frozen bank runs again after your work has had time to land. The full six phases, with what each one requires from you: the method.
Study size is a range, not a package
A study has three knobs. There is no standard configuration, because the right size depends on the decision the study must inform.
Questions
A few dozen → around 100
Models
Fewer → 4 or more
Repetitions
Fewer → 5
Small directional sample
A few dozen prompts, fewer models, fewer repetitions. Enough to see where you stand.
Deep study
Around 100 prompts, 4 or more models, 5 repetitions. Enough to compare segments and models with confidence.
60 questions × 3 assistants × 3 repetitions 540 responses read
Standard depthMove the dials. One illustrative configuration, labeled as such: 60 × 3 × 3 = 540. It is an example of the arithmetic, not a default.
Size scales with the same three knobs, and the scope is agreed in the kickoff, before the first prompt is sent. The scope conversation happens before anything runs, 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 category conversation 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.
The B2B example
Same structure, different reality: a considered purchase with a long cycle, a buying committee that asks assistants early, competitive loss inside evaluation questions, and reputation questions about the brand.
Both samples carry the same invented-brand structure, verbatim: made-up brand, made-up competitors, made-up figures. We keep it that way because the structure is what matters, and an invented brand raises no confidentiality questions.
Questions
What do you commit to — and what does nobody guarantee?
We commit to two things: a prompt bank that is complete and useful for your brand, whatever your industry, and a run that produces useful findings. What nobody can guarantee — us included — is that the actions you take afterwards will move the numbers in a later measurement. Answers vary by user, by context and from run to run; one screenshot is an anecdote in both directions. The method is public precisely so you can judge the findings, not the promise.
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 assistants 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?
Building and freezing the prompt bank takes one week to one month, per brand or entity measured, and most of that is the back and forth to understand the category and the buyers, paced by your availability. The run itself takes about two days once the bank is frozen, and the analysis and the report about one week. After that we recommend acting on the report and measuring again one to three months later.
What happens after the study?
The study is the baseline. Next comes the work on the sources: the ordered list the report delivers, worked in priority order. Monitoring then re-runs the same frozen bank on a cadence, so the second measurement is comparable to the first and tells you whether the work worked.
Request the study
Bring the decision the study has to serve, and the category it has to serve it in. If the study fits, we scope it in the kickoff and you approve the scope in writing before the first prompt runs.