How the work is done
For agencies, researchers and anyone who wants the detail. The front page says what you get; this page says how, and where its limits are.
The five steps
Listen
Your own customers first, if you share their reviews, survey answers or messages. Then the public: research tools read forums, reviews, press and video comments in the audience's own language, and official APIs add ratings, Instagram, YouTube and the businesses' own websites and ads. We also check how a new customer finds you: on search engines, and when they put the questions buyers ask to AI assistants (Claude, and ChatGPT, Gemini or Perplexity where connected), several times each, because the answers vary. Items that only matched a word are set aside, and a sample of that screen is checked by hand.
Build personas
10–15 fictional customer profiles, one per kind of buyer heard, built on the phrases real people used (each linked to where it was said), including one or two loyal to a competitor and one sceptic.
Run the panel
Each persona answers alone first, with the options in a different order, then sees the others' answers and may change its mind. It runs on several AI models; a persona counts once whatever the model. A study that does not tell the options apart says so and draws no conclusion.
Decide
Every panel result is a guess, labelled as such, with one cheap real test sized to your own numbers. The consultant writes down their own prediction first. You choose what to act on.
Check
The real test runs: a question to your next 30 customers, a poll, two versions of a caption, your sales. The result is recorded, inconclusive ones included, and the share of guesses that held is kept.
If you have not opened yet
How people find you does not matter yet; how they find this kind of business, and your competitors, does. So the study reads the market instead of your business.
How buyers find this kind of business
For each group of buyers: what starts their search, where they look (search engines, AI assistants, directories and marketplaces, forums, word of mouth, events, ads, social networks, review sites), who they find at each step and why, and what is missing or annoying. Real searches, kept and dated so you can repeat them.
How your competitors are found
Each competitor, channel by channel: whether it comes up on the searches buyers type, how often each AI assistant names it (each assistant on its own), its listings and reviews, whether the public talks about it, the ads it runs and what it leads with. Strong, present, not seen or unknown, with the page behind each cell; then the channels they neglect.
A launch plan for the first 90 days
Where to start and the position the gaps leave open; the channels in the order to open them; three phases with milestones sized to the hours and budget you have; the riskiest assumptions first, each with a cheap test (ten conversations, a landing page with a waitlist, a pre-sale) and the result that confirms it; and when to stop or change course. Every element shows the evidence it rests on. It is a plan to test, not a promise.
How far to trust the simulated panel
Before its rankings are used, the panel is tested on posts whose real results are known: it ranks them without seeing the numbers, and its order is compared with what really happened, and with simple yardsticks such as "newest first" and the consultant's own guess. That score is shown with every study.
Scores move with the model, so only rankings are compared, and only what holds across models is reported. A few real customers answer a short interview alongside the panel, and where the two disagree, the real answers win.
What we cannot read
Closed Facebook groups, private messages, most of TikTok and Reddit, paywalled articles, and platforms without an open API. Every report lists what it could read and what it could not, and how many public items its findings rest on.
Personal data
What was said is kept, never who said it: no names, usernames or profiles. Quotes are at most 25 words and linked to their public source. Excerpts are deleted after the period agreed with the client, and anyone quoted can have their words removed. Analysis uses Anthropic's Claude models; the data is kept on a server in the EU. The questions put to other AI assistants carry no personal data, and only what their answers say about businesses is kept.