Where we started
There was no problem to fix here, and it is worth saying that honestly: the existing content was fine.
The issue was different. The site had a backlog of topics worth covering, but there was no time to write them by hand. Sections like that are usually postponed to «someday» and never appear.
What we did
A preparation algorithm, not «we asked the model to write». For every piece we define in advance which question it answers, which facts it must contain and in what form — so that the answer can be quoted whole, without being torn out of context. This is the part of the work no prompt replaces: it happens before generation.
The model drafts, a human sets the structure. The model works from a prepared outline instead of deciding what to write about.
Mandatory expert review. Every piece then goes to a person acting as an auditor: checking facts, removing anything invented, fixing wording where the model drifted into generalities. Nothing is published without this step. It is the bottleneck of the process — and it cannot be removed without losing the point of the work.
Publishing in batches with a review of results. After each batch we look at what actually started appearing in search and adjust the preparation algorithm for the next one.
What changed
A section that would otherwise never have existed was filled in. Topics nobody had time for are now covered.
Some of the material started appearing in the AI answer block in search. That cannot be guaranteed — the answer is generated anew on every query — but the structure was prepared for it, and the result is visible.
We do not publish the numbers: the client is under an NDA.
Where we do not use generation
Commercial pages, case studies, service descriptions and anything containing figures or commitments are written by hand. The model works where a volume of reference material with a clear structure is needed — and only with review.