Every team that tried to automate content in the last two years learned the same lesson in roughly the same order. The first month is euphoric: an article costs four minutes instead of four days. The third month is quiet. By the sixth month someone opens the analytics and finds a hundred pages that nobody reads, nobody links to, and nobody wants to admit writing.
The mistake was not using a language model. The mistake was pointing it at the wrong artifact.
Drafting is the cheap part
Writing a first draft has never been the bottleneck on a functioning content team. The bottleneck is knowing what the page has to contain in order to deserve the position it is aiming at. That knowledge is not in the model. It is in the ten results currently ranking, in the questions people actually type after the first query, in the internal links your site can afford to point at a new page, and in the parts of the subject your competitors have already covered better than you ever will.
A model asked to "write an article about technical SEO audits" invents a plausible average of everything it has read. A model handed the ten current results, the entities they all mention, the three questions none of them answer, and the two internal links that should exist writes something considerably more useful. Same model, different input.
What a brief has to contain
We settled on five parts, and we throw away briefs that are missing any of them.
- The target and the gap. One query, its current position, and one sentence on why the page is not winning yet. If nobody can write that sentence, the page should not be on the backlog.
- Entities and subtopics. Extracted from the pages that rank, not from the model's memory. Each one carries the count of competing pages that cover it, so a writer can see what is table stakes and what is a differentiator.
- Questions to answer. Pulled from related searches and from the follow up queries in your own Search Console data.
- Internal links, both directions. Which existing pages should link to this one, and which pages this one should point at. This is the part every generated outline forgets, and it is the part that decides whether a new page gets crawled and understood.
- Sources. Every claim in the brief links to the page it came from. A writer who cannot verify a claim in one click will either repeat it blindly or ignore the whole document.
Why grounding matters more than model quality
A brief is verifiable. Open the sources, check the entity list against the results, disagree with the priority. A draft is not verifiable in any practical sense: to check it you have to rewrite it, at which point the model saved you nothing.
That asymmetry is the whole argument. Put the model where its output can be audited in thirty seconds by a person who knows the subject, and it becomes genuinely useful. Put it where auditing costs more than doing the work yourself, and you have automated the production of a mess.
The workflow that survived
Cluster the queries. Score the clusters by realistic upside. Generate a brief for the top of the list. Have a human argue with the brief for ten minutes. Then write, with or without help, and publish something a person chose to publish.
The model is in the loop the entire time. It just never has the last word.