Search never measured quality. It measured things that correlated with quality back when producing content was expensive: length, structure, internal linking, freshness, apparent expertise.
Those were proxies for effort. Effort was the scarce input, so proxies for it worked well enough. Generation removed the scarcity and left every proxy standing exactly where it was, which is why the guidance that worked in 2022 started actively hurting sites.
What using AI for SEO means in practice
The phrase covers two strategies that behave very differently, and conflating them is how teams got hurt.
The first is scaled generation: produce more pages, cover more keyword variations, publish a comprehensive guide for every long-tail query. This is the strategy the tooling makes obvious, and it is the one that stopped working.
The second is using generation on the mechanical parts of content work while keeping the substance human: drafting from your own research notes, restructuring an argument you have already had, producing variants of a meta description, finding coverage gaps in an existing page.
Only the second is durable, and the reason is structural rather than a matter of which algorithm update landed when.
Why scaled content generation stopped working
If anybody can produce a comprehensive two thousand word guide on any topic for almost nothing, then a comprehensive two thousand word guide is no longer evidence of anything.
A ranking system that continued to reward it would be one update away from being embarrassing, so the correction was inevitable. The specific timing of enforcement is less interesting than the direction, which was never in doubt.
The teams that got hurt were mostly not the cynical ones. They were the ones who did exactly what the prevailing guidance said, at the scale their new tooling suddenly permitted. Doing the recommended thing a hundred times faster turned out to be a different activity from doing it.
The four things that cannot be generated
The productive question is not how to write for search. It is what you have that a model does not, and the honest answer is narrower and more valuable than most content plans assume:
- Your own data. What your product actually observes, aggregated and anonymised, with the sample size stated. Nobody else has it and it cannot be synthesised.
- Named, specific experience. Not "best practices for onboarding" but what happened when you changed yours, including the version that failed and what it cost.
- An actual position. An argument a reasonable person could disagree with. Generated text converges on consensus by construction, so taking a real position is something only an accountable author does.
- Primary work. Interviews you conducted, tests you ran, numbers you measured yourself rather than found in somebody else's post.
How AI search changes what gets cited
There is a second-order effect worth planning for. As more queries are answered inside an AI summary rather than by a click, the value of ranking shifts toward being the source a summary cites.
The things that get cited are specific claims that can be attributed: a number with a method behind it, a named position, a definition somebody wrote deliberately. Generic surveys of a topic are exactly what a summariser replaces rather than quotes.
This points the same direction as everything above. The content that survives is the content that contains something only you could have written.
The measurement trap this creates
Publishing volume is the most tempting metric here and the least informative one now. It was a defensible proxy when a post cost a day of somebody's time. It measures almost nothing when a post costs eight minutes.
The metrics that survive are the ones tied to whether a human got value: return visits, direct traffic, branded search, and whether anybody cites or links to you without being asked. These are slow to move and hard to fake, which is precisely why they still carry information.
The pattern generalises well beyond marketing, and it is the reason this blog exists. Any time a proxy was standing in for effort, and that effort becomes cheap, the proxy stops carrying information. Organisations keep reading it for years afterwards, because it is still easy to count.
