Free · one page · one finding
The Gap Read
What you get
The question
The exact buyer-intent prompt we asked, verbatim, so you can paste it in yourself.
The answer
What the engine actually said, word for word, with every competitor it named.
The number
How many runs, on which engines, on what dates — and how many of them named you.
The screenshot
A full-height, dated, uncropped capture of the whole answer. Not a chart we drew.
One finding. Not eight, not a scored report, not a deck. If there is nothing there, that is what you get told — and that is a useful answer too.
The method
Why this is worth anything
- Two engines — named, because it’s only two
- ChatGPT and Gemini Flash, the free default, which is the tier a real buyer is on. We do not test Perplexity: it blocks our capture method and has produced zero captures for us. Anyone telling you they cover “all the major AI engines” is describing a wish.
- Repeated runs, because one run is noise
- Across our own corpus of 514 captures, ChatGPT gives a materially different answer to the same question 17% of the time. Gemini does it 28% of the time. So every question runs several times on both engines, and a finding that doesn’t hold across repeats gets deleted rather than softened. That rule has killed more of our own findings than it has kept.
- Neutral sessions
- Logged out and context-free on ChatGPT; personalisation off and temporary chat on Gemini. A personalised session measures what the engine says to us, which is not the thing you want to know.
- Dated screenshots, full height
- Every capture is timestamped and screenshotted whole, not cropped to the part that makes the point. The raw answer text is kept too, and you can have it.
- A recorded capture origin
- We log where each capture was taken from, because these engines localise silently and a capture without its location can be read to mean the wrong thing. We know that because it happened to us — see below.
- Read by a person
- Every headline number is checked by hand against the raw answer before it goes on a page. A pattern-matcher cannot tell “X is the best choice” from “alternatives to X include Y”, and that difference is the whole finding.
A real one, anonymised
A fintech that rebranded, and the engines that didn’t
A financial platform changed its name in late 2025. We ran its twelve category questions across both engines and counted, by hand, how often each answer used the old name versus the new one. Across 72 answers, the retired name appeared 226 times and the current one 66.
The interesting part isn’t the ratio. It’s where the ratio sits.
| When the question is… | Retired name | Current name |
|---|---|---|
| “What happened to [old name]?” — asked directly | 44 | 37 |
| About its reputation and past problems | 72 | 4 |
| About alternatives to it | 39 | 2 |
| Comparing it against two named competitors | 32 | 0 |
Asked point blank, both engines know the rebrand perfectly — one of them leads with the new name. Asked the questions a buyer actually asks on the way to a decision, the answer comes back filed under the name the company retired.
Every number above was counted by hand from the raw answers. The company is a real client-track prospect and is not named here because it has not asked to be.
And one we withdrew
A finding we killed, because it’s the same reason to trust the others
We had a national retailer that looked invisible on its core category questions and visible the moment you added a state name. It was a clean story. It was wrong.
Our captures run from Arizona, and the engines had quietly localised every “neutral” question to Phoenix — naming Phoenix shops in the answers. The retailer has no Arizona stores. So its absence wasn’t invisibility; it was the engine being correct. We found that in our own raw text, deleted the finding before it was ever sent, and added a capture-origin field so it can’t happen silently again.
We’d rather show you that than a case study. A method that only ever confirms things isn’t a method.
Questions
Two: ChatGPT and Gemini Flash, the free default tier. Those are the only two our harness has ever run. We don’t test Perplexity — it blocks our capture method and has produced zero captures — and we don’t claim coverage we don’t have.
Because these engines are unstable. Across our corpus of 514 captures, ChatGPT answers a repeated question differently 17% of the time and Gemini 28%. A single run sits inside that noise floor and proves nothing.
Nothing. No call to book, no account access, no form beyond telling us the business and the category. If we find nothing, we say so.
Yes, and you should. You won’t get the answer word for word, because of the instability above. You should see the same pattern. If you don’t, tell us — we publish corrections rather than quietly withdrawing them.
It’s the same mechanism pointed at a different mediator. SEO is discovery applied to search engines; AEO and GEO are discovery applied to answer engines. Calling it a brand new discipline is how you end up rebuilding from scratch the next time the mediator moves.
No. Nobody who understands these systems can, and anyone offering that guarantee is selling certainty they don’t have. What’s guaranteed is that the measurement is real, repeated, dated and reproducible by you.
Want one?
Tell us the business and the category. That’s the whole ask.