The short version
- Questions come from buyers. We agree 10 buyer situations with you and ask each one four ways.
- No company names in the questions. We ask what a buyer would type, then see who gets named.
- Five tools, recorded. Every answer and every source it showed is saved.
- Counts, not impressions. A finding is a number of answers out of a stated total.
- Limits stated. A study describes what was recorded. It doesn't claim to know why a model chose a firm.
Step 1: intake
You send your website, who you sell to, two or three competitors and the markets you care about. We reply by email within one working day with the scope and the price.
Step 2: buyer situations
A buyer situation is one decision a real buyer faces. "Which commercial siding contractor should an out-of-town developer consider?" is a situation. "Siding contractors" is not.
We agree 10 situations with you before we record anything, and write four questions for each. The questions are worded the way a buyer would type them, with no company names, so the answer shows who AI brings up on its own.
Step 3: recording
Each question goes to ChatGPT, Gemini, Claude, Perplexity and Google AI Overviews. That gives about 20 answers per situation and about 200 for a study.
- We record in fresh sessions with no chat history.
- We save the full answer and every source it showed.
- We note the mode each product was in, because the same product can answer differently with search on or off.
Step 4: counting
Three things are easy to confuse, so we count them separately.
| Term | What it means |
|---|---|
| Mention | The answer names a firm. |
| Recommendation | The answer puts the firm forward as one to consider for that situation. |
| Citation | The answer links to or shows a page as a source. A firm can be recommended without its own site being cited. |
Every finding is a count out of a stated total, such as "recommended in 15 of 50 answers". We don't turn small samples into percentages that look more certain than they are.
Step 5: tracing the sources
For each recommendation we follow the answer back to what the tool leaned on. Often it isn't the firm's own website. It's a directory profile, a "best of" list, a review page or a competitor's page.
The tools differ a lot here. In one study, Perplexity showed the firm's own website as a source in 58 of 66 recommendations, and Gemini in 12 of 112.
Step 6: reading
Every answer is read by a person before it reaches you. That's how we catch the things a count misses: a firm described as the business it used to be, a tool saying it couldn't verify a price, or two companies with similar names being mixed up.
Step 7: the file
You receive a private study site with every question and answer, a PDF, and the fixes ranked by what to change first. We walk you through it in chat. See what the study includes.
How we handle answers that change
A single AI answer can change from one day to the next, so we never rest a finding on one answer. Each situation is asked four ways across five tools.
In one study we asked the same questions twice, a day apart. Whether the firm appeared stayed the same in 24 of 25 pairs. The order of the shortlist moved around much more. So we treat who appears as the stable signal, and position as noise.
Monthly tracking repeats your questions three times a month, with two runs per tool, to separate real movement from noise.
What a study can't tell you
- It isn't a market survey. The answers are repeated observations of agreed questions, not a sample of all your buyers.
- It can't see inside the model. We report what was said and what was cited. We don't claim to know why a tool chose one firm over another.
- It's a dated record. AI products change their models and their search behind the scenes. A study shows what was true when it was recorded.
- A recommendation isn't an endorsement. Being named by an AI tool says nothing about the quality of a firm's work.
We separate three levels in every report: what we observed directly, what we think it means, and causes we haven't tested.
How market research differs
Market research uses the same habits, with different sources: public data, pricing pages, filings, product histories and, where needed, buyer interviews. Every figure is tied to its source, and the study ends in a go, pivot or not-yet verdict. See market research.
Questions
AI gives a different answer every time. Is this reliable?
Single answers vary. That's why we ask each buyer situation four ways across five AI tools, about 20 answers per situation. Patterns that hold across that many answers are the ones worth acting on. With monthly tracking, every question runs twice per tool, three times a month, so you see trends rather than one-off answers.
Why don't the questions include my company name?
Because buyers who don't know you yet don't type it. Asking about a company by name only shows whether AI has heard of it. Asking the buyer's question shows whether AI recommends it.
Which AI tools do you record?
ChatGPT, Gemini, Claude, Perplexity and Google AI Overviews.
Can you tell me why an AI tool recommended a competitor?
We can show you what the answer said and which pages it cited, and compare those pages with yours. We can't see inside the model, so we don't claim a single cause.
What do you need from me?
Your website, who you sell to, two or three competitors and the markets you care about. We agree the 10 buyer situations with you before we record anything.