AI for Search Content, Honestly

The obvious use of a language model in publishing is producing a lot of pages cheaply. That worked briefly and stopped, and understanding why is more useful than any list of tactics.

What actually changed

Search engines have never had a rule against AI-written content as such. Their stated position is that the origin does not matter and the quality does.

What they penalise is content produced at scale primarily to rank, regardless of how it was made. That description covers most bulk AI publishing, and enforcement against it has become considerably more effective.

Meanwhile the supply exploded. When generating a thousand words costs nothing, everyone does it, and the marginal value of another competent summary of a well-covered topic is zero.

The result: volume is no longer a strategy, because volume is free for everyone. Whatever advantage exists now comes from what a model cannot produce.

What a model cannot produce

Your own data. Numbers from your business, your customers, your experiments, your logs. Nobody else has it and no model can generate it.

First-hand experience. You used the product, ran the process, made the mistake, handled the case. Specificity of this kind is immediately visible and cannot be synthesised.

Genuine expertise applied to a hard question. Not a summary of what is known, but a judgement about something contested.

Original research, however small. A survey of thirty people, a comparison you actually ran, a test you performed.

A position. Most generated content is balanced to the point of saying nothing. Having a view, and defending it, is differentiating.

Currency with verification. Something that happened, checked, from a source.

Where AI genuinely helps in this work

Research and orientation before you write.

Structuring an argument you have already formed.

Editing — clarity, cutting, tightening. Strong use.

Drafting sections you would write the same way anyway — background, definitions, boilerplate.

Producing variations for testing headlines or descriptions.

Technical work. Structured data, schema, metadata, internal linking suggestions, identifying gaps in coverage. These are genuinely well suited and unglamorous.

Translation and localisation of existing good content.

Where it does not

Generating pages on topics you know nothing about. This is the thing that stopped working, and it now carries downside risk rather than being merely ineffective.

Producing many near-identical pages targeting keyword variations.

Replacing the expertise. If nobody involved knows the subject, the output has no source of value and no amount of editing supplies one.

The thing worth internalising

The bottleneck moved. It used to be production capacity: writing costs money, so more pages meant more budget. That constraint is gone.

The bottleneck is now having something to say. Which is exactly the thing the technology does not help with.

Publishing operations that understood this early redirected effort from volume to primary material — running tests, gathering data, interviewing practitioners, documenting real cases — and used AI for the production layer around that material.

Operations that used the technology to produce more of what they were already producing found that it stopped working, and in some cases that the whole site stopped working.

A defensible approach

Fewer pages, each with something original in it.

Every page passes a test: what is here that a model could not have generated? If the answer is nothing, do not publish it.

Use AI for the parts that are not the value — structure, editing, metadata, formatting, translation.

Verify every fact, because a fabricated statistic in a published article is a durable embarrassment.

Update rather than accumulate. Improving existing pages is generally worth more than adding new thin ones.

This is slower and it is the only version that survives contact with the current situation.

What to do with an existing thin site

For anyone holding a site built on volume that has stopped performing, the options are narrower than the recovery advice suggests.

Audit honestly. Which pages have any original substance and which are restatements of what was already available. The second category is usually the majority.

Consolidate rather than delete indiscriminately. Several thin pages on adjacent topics frequently make one good page.

Remove what cannot be fixed. Pages with nothing original, on topics you have no expertise in, are a liability rather than an asset. Retiring them properly is better than leaving them.

Improve the pages worth improving by adding what only you have — data, testing, first-hand experience.

Accept that some of it is not recoverable. A site with hundreds of generic pages and no source of original material has no route back that does not involve becoming a different kind of site.

Then publish less, better. The bottleneck is having something to say, and no amount of restructuring supplies it.