“What does ChatGPT say about us?”
Your client has asked. Maybe the CMO forwarded a screenshot where a competitor got recommended and they didn’t.
You can open ChatGPT and try it yourself. You will get an answer. Ask again tomorrow and you will get a different one — that is how these systems work, and it is exactly why one screenshot is not a finding. Meanwhile the tools that promise to measure it sample once, quote a single score with no error bar, blend seven engines into one meaningless average, or cost more per month than the retainer this client is on.
So the honest answer today is usually we don’t know. This is the product that changes that.
Bars are 95% confidence intervals, not point estimates. n=210 per engine.
Four things nobody credible combines
Sampled seven times, so it is a measurement
AI answers are not deterministic — ask the same question twice and the brands can change. We run every prompt on every engine seven times, 1,470 answers in a standard audit, and report a percentage with a confidence interval rather than a rank. The interval resamples prompts rather than answers, because seven answers to one question are not seven independent observations, and treating them as if they were makes the numbers look far more precise than they are.
Reported per engine, never blended
Seven surfaces, seven separate supply chains. Google's AI Mode and its AI Overviews routinely disagree with each other, which is why we count them separately when most tools fold them into one. A single blended score averages a surface your client's buyers live on with one they have never opened.
A root cause, not a dashboard
The standing complaint about this category is “great chart — now what?”. Every finding resolves to a sentence a strategist can act on: not “you are at 22% on ChatGPT” but “you are absent from comparison-shaped prompts because that supply chain is two review sites and neither carries a current listing for you”.
White-label from the first paid tier
Your logo, your colours, your domain. Not an upsell and not an enterprise call. The report is built to be handed to a client with your name on it — which is also why we are careful about what we claim inside it.
What you actually get
A PDF plus the raw data behind it — every stored answer, so any figure can be traced back to the text that produced it.
How often the brand is named, per engine, with a confidence band.
The same measurement for the competitors you nominate, on identical prompts.
Which domains and pages the engines actually drew from, split into owned, earned, social, institutional and competitive. The mix tells you whether this is winnable with your own content or not.
When the engines name your client but source someone else's page. Its corollary is fragile visibility: the share of presence resting on pages they do not control and cannot keep.
Whether each prompt is won by a comparison page, a listicle, a docs page, a forum thread or a review — which turns a source list into a brief.
Whether AI crawlers can fetch the pages, whether the title and description would get them chosen, and whether the answer survives extraction once they are.
Plus a ranked list of where a mention is realistically worth chasing.
Four steps, no access required
Brand, site, two to four competitors, category. No logins, no analytics access, no code on the client's site.
20–30 buyer-intent questions built from the category and the site. You edit the list before anything runs. No brand names inside the prompts — naming a brand in the question is how you get a flattering number instead of a true one.
Seven engines, seven samples each, every answer stored verbatim as evidence.
The white-label PDF and the raw data behind it, in under [X] days.
What this audit will not tell you
A single audit is a photograph, not a film.
It cannot tell you that you are declining. Direction needs at least two measurements. We store the baseline from your first run so a second one can answer it, but run one cannot.
Results are not broken down by buyer persona. The prompt set is built around jobs and qualifiers, not audience segments.
We do not see the internal queries an engine generates behind your question. We measure what the buyer is shown, which is the thing that decides whether your client gets recommended.
With seven samples, one prompt scoring five of seven is consistent with anything from roughly a third to nine in ten. So prompts appear as evidence — named or not, in how many samples, and who was named instead — and never as a percentage with a false decimal point on it.
If a tool tells you all of this from one run, it is guessing on your behalf.
Before you ask
Generative engine optimisation and answer engine optimisation are two names for the same work: getting a brand named and cited in AI-generated answers. Gartner calls it answer engine visibility. We say GEO; if you arrived holding AEO, you are in the right place.
Seven: ChatGPT, Google AI Mode, Google AI Overviews, Copilot, Perplexity, Gemini and Claude — on every paid tier. No engine is held back for a higher plan.
Free checkers answer from a stored database of past prompts. We run your prompts live, seven times per engine, and hand you the answers.
Because AI answers are non-deterministic. One run is an anecdote, not a measurement.
No. Anyone who does is selling something they cannot deliver. We can tell you who is being mentioned instead, which sources fed those answers, and what format of page keeps winning the prompt.
No. A brand name, a URL, competitors and a category description.
Yes, on every paid tier.
Run it on one client and see
One audit, one client, your logo on it, [PRICE]. If the report does not survive the client meeting, we have both learned something cheaply.