Search engine optimisation had one great convenience: the query was visible. You could see what people typed, and you could see where you ranked.
AI assistants took that away. Nobody types “best security consultancy in Skåne” into a model and reads a list. They ask a question in their own words, and the answer names two or three companies, or none. There is no ranking page to look at.
Measure first
We built a way to measure it before trying to change it. The method is simple to state:
- Compose the questions a buyer would actually ask about your category, in their words, in several phrasings.
- Ask those questions to the assistants your buyers use.
- Score each answer: is the brand named, how is it described, which competitors appear next to it.
- Repeat on a schedule, because the answers drift.
Step 1 is where the work is. A brand that scores well on “who does X” and badly on “how do I solve Y” has a content gap that the second phrasing reveals and the first one hides. The score is only as good as the prompt set behind it.
Why composition matters more than volume
Models answer from what they can retrieve and from what they were trained on. Both reward text that states plainly what a company does, for whom, and with what evidence. They do not reward adjectives.
So the fix for a bad score is rarely “write more”. It is “write the sentence the model needs to find”: the claim, the proof, the name, in that order, on a page the crawler can reach.
What we learned building it
Running this at scale means many model calls, in several languages, with cost that adds up. We route each call to a model chosen for the job and log every call with its tokens and cost, so the measurement itself stays affordable. We keep all of it inside EU regions, because the prompts contain company names and the answers are a form of market data.
The measurement is the product of a prompt set and a schedule. Get those right and the content work becomes obvious.