Case study
Measuring what AI says about us, before we did anything
The day-zero prompt panel for Grafynt across ChatGPT, Gemini, Perplexity and Google AI Overviews, measured against ordinary search for the same queries, on the day this site went live.
Measured 2026-10-15Subject: Grafynt
Method
How this was run, so it can be run again.
A result nobody can reproduce is an advertisement. The next run uses this list unchanged.
- A fixed set of buying-intent questions in the wording a buyer uses, in both Bahasa Indonesia and English, not our own service names.
- Run logged out, with no session history, so personalisation could not bias the result toward us.
- Each question checked in both AI assistants and ordinary search, so a ranking problem is distinguishable from a citation problem.
- Every business named instead of us recorded, which is the output a client receives by name.
- A retrieval check on our own pages: what a live web-search API actually receives when it fetches them.
Findings
Three findings and a test.
Citation presence: none
Grafynt was named for none of the buying-intent questions, in any engine. Searching the brand name directly surfaced neither the site nor its author, and returned two unrelated companies with near-identical names instead. The site was hours old when this ran, so this is the expected result rather than a surprising one, and it is the number every later run is measured against.
Description accuracy: nothing to be accurate about
Nothing is said about Grafynt because nothing names it. The author entity does not exist publicly either, which matters more than it looks: attributable authorship is weighted in citation, and ours currently corroborates nothing.
Competitor set: established and direct
Seven agencies were named across the buying-intent questions, and they are not adjacent businesses that happened to appear. They sell the same three letters in the same city and target the same wording, several with pages built specifically for the GEO and AEO queries. This is a contested market, not an empty one, and planning as though it were empty would be the expensive mistake.
Retrieval and consistency: passing
The one part that already works. Fetched the way an assistant fetches, the site returns a complete and accurate picture: positioning, all five prices, both bundles, timelines and markets, with no JavaScript required and nothing hidden behind an interaction. Every surface agrees because every number is read from one file. The constraint is not that we are unreadable. It is that nothing points at us yet.
What happens next
Written down before we know the answer.
Published in advance so it cannot be rewritten afterwards.
- Visibility is not a switch. Retrieval-based answers can move within days of a page changing, but being named at all requires that something, somewhere, refers to you, and on day zero nothing does.
- The next run is in one month, with the same questions, logged out, in both languages. If the method changes, the comparison is worthless, so it will not.
- The honest prediction: brand-name queries should resolve first, buying-intent questions much later, and the competitor set is unlikely to shrink. Published here before we know, so it cannot be rewritten afterwards.