design+code
In June we published what we were going to do about GEO. That article was a promise. This one is the audit of that promise, two months later, measured every week with the same protocol.
We are publishing the method and the findings, not our traffic numbers. The method is the part you can reuse.
Ten prompts, frozen since August. We do not edit them, because a panel you keep editing is a panel that always flatters you.
They split into three families. Four acquisition prompts, what a prospect would actually type when looking for a studio like ours, in French and in English. Two brand prompts, who we are and what we charge, which test accuracy rather than presence. Four content prompts, on subjects where we have already published something substantial.
Two engines per week, on rotation, so a run stays short. Each engine is only ever compared to its own previous run, never to another engine. One prompt per fresh conversation.
We record three things: cited or not, which page, and which sources were cited instead of us. That third column is the useful one. It is the list of sources you have to displace.
This is the finding that cost us the most. Several of our early readings were not measurements at all.
Every engine we tested on our own account recognised us. The temporary chat in ChatGPT now states that it may use memory. Perplexity answered an acquisition prompt in our own positioning language, as though it were briefing us on ourselves. Claude named our studio without being asked.
An engine that knows who is asking will cite you out of politeness. That is not a measurement, it is a mirror.
Run the panel signed out, or in a private window. If you cannot, throw the result away rather than count it.
Worth stating, because it looks like a bad score and is not.
On an account without web search enabled, Claude declines to name sources at all. It says so plainly rather than guessing. Zero citations for us, and zero for our competitors.
An engine that cannot retrieve is not measuring your visibility. Note it and move on to one that can.
Every citation we earned came from the same kind of page. A comparison that names competing products, with a stated rule for choosing between them.
Our article comparing three restaurant booking systems is cited first, with the full URL and a correct summary of our selection rule. Nothing else in our journal performs like it.
Pages about ourselves get cited only when the question already contains our name. That is not visibility, that is a lookup.
If you want to be cited, write the comparison your clients actually have to make, and name the alternatives.
This is the uncomfortable one.
On all four acquisition prompts, the sources cited are directories and roundups. Freelance marketplaces, review platforms, and best-agencies-in-Paris listicles. Studio websites appear rarely, and usually because a directory pointed at them first.
Meanwhile an entire ecosystem of GEO agencies cites itself on our core commercial question: Triaina, hyffen, Archipel Marketing, YATEO, iaba, Uplix, SEO.fr, Eskimoz. We are not in it.
You cannot write your way into that set from your own domain. Being listed where the engines already look is a different job from publishing, and we had been treating it as optional.
For several weeks an engine quoted a starting price for us that did not match our published rate card. We logged it as a hallucination.
It was not. It was reading our own FAQ, which contradicted our own pricing page by a wide margin. Two pages, two numbers, one of them wrong for months.
An engine will reproduce your inconsistency faithfully and confidently. Being present and wrong is worse than being absent. Audit your own pages for contradictions before you blame the model.
We have kept an llms.txt file in production since June, and we still recommend it. It costs nothing and it is easy to maintain.
But across two months of weekly measurement, we cannot attribute a single citation to it. No engine has publicly confirmed that it reads the file. Treat it as hygiene rather than as a lever, and be sceptical of anyone selling it as one.
Two things, both relative, both consistent week over week.
Roughly four fifths of our search impressions now come from two very long, question-shaped queries that produce no clicks at all. They read like questions an agent would ask on behalf of a user rather than anything a person types. We cannot prove that. We note it and we watch it.
And our impressions inside AI answers halved in a month, with the drop coming almost entirely from two pages losing their position. Everything else held. A site-wide metric moving is usually a handful of URLs moving.
If you only look at totals, you will diagnose the wrong problem.
We rewrote our FAQ. The answers were fine. The headings were house style rather than questions, and a FAQPage entry is matched on its question. The editorial headings stay visible, with the real question underneath, and that is what the structured data now carries.
We fixed the pricing contradiction, in the visible text and in the structured data, on the same day we found it.
And we added the entities that were missing. Our FAQ never contained the words for the sectors and the cities our prospects actually search with.
You need a spreadsheet and about ninety minutes a week.
Write ten prompts your buyer would actually type. Freeze them. Play them signed out, one per fresh conversation, on two engines, on a rotation. Record cited or not, which page, and who was cited instead of you.
After a month you will have something almost nobody in this market has. Evidence instead of a claim.
If you want us to build the panel for your brand and run it every week, email bonjour@dellamattia.com.