Why generic AI content fails and specific AI content does not
Pages fail not because a model wrote them but because they contain nothing only that company could have said. A model given your records, your figures and your conditions produces something citable; a model given a title produces the same article everyone else got.
The industry argument about whether AI-written content works is being conducted at the wrong level. Search and assistant systems are not detecting authorship. They are detecting whether a page adds anything, and a page assembled from the general consensus adds nothing whoever typed it.
What the failing pages have in common
- Nothing checkable. No figures, no dates, no named conditions, no jurisdictions.
- No first-hand account. Nothing that happened to this company, on a real project, with a real outcome.
- A structure that surveys rather than settles: seven considerations, five best practices, three trends.
- Confident sentences about things the writer did not know, which is where fabrication enters.
- Nothing a reader could disagree with, which means nothing worth reading.
A page like this would have failed in 2015. What changed is the cost of producing it, so there are now thousands of them competing for each question, and the bar for being the one that gets read has risen accordingly.
The question is not who wrote it. It is whether anyone else could have.
What the working pages have instead
| Generic | Specific |
|---|---|
| "Migration timelines vary depending on complexity." | "Under a thousand URLs, one language: four to six weeks. Over that, or multilingual: three months." |
| "Pricing depends on your requirements." | "Between €12,000 and €40,000. What moves it: number of languages, whether the content is rewritten, and who owns the redirect mapping." |
| "We take data protection seriously." | "Processing in the EU, sub-processors listed below, thirty-day retention on conversations, no training on customer content." |
| "Results improve over time." | "We have not measured this yet and will publish the reading when we have." |
Note the last row. "We do not know yet" is a specific statement and a credible one. It reads as carefulness, it is honest, and it is a great deal better than a figure somebody invented to fill the gap.
Where a model genuinely helps
Turning a four-paragraph reply into a page with an answer at the top and headings written as questions. This is mechanical work and models are good at it.
Reading what your market searches for against what you have published, and telling you where the gaps are.
Across thirty pages written over two years by four people.
The draft is a starting point that saves the blank page, not the thing you publish.
Where it does not
Supplying the specifics. A model does not know your prices, your conditions, what happened on the project that went badly, or which of two approaches your firm actually recommends. If those are not in the input, they are not in the output, and what you have published is the consensus with your logo on it.
Questions this settles
The stated position is that the method of production is not the issue; unhelpful, unoriginal content is, however it was made. In practice that means specificity decides the outcome.
It is a reasonable thing to do and costs nothing. What matters more is that a named person reviewed it and stands behind the specifics.
Yes, routinely, when it contains something the writer actually knew. The failures are pages with nothing in them, not pages with a model in the process.