AI content risk audit

What actually goes wrong, and it is not the writing

Sites that scaled AI content and then declined usually share a small set of specific faults, and none of them is about prose quality.

The first is invented specifics. Language models produce statistics, study references and quotes that read entirely plausibly and do not exist. Published unchecked, these sit on your site as claims you cannot support, and anyone who checks one finds it immediately.

The second is sameness. Everyone prompting a similar model about a similar topic gets a similar answer, so a page can be original by any plagiarism test and still be functionally identical to a dozen competitors. It offers no reason to rank and, for AI search, nothing distinctive to lift.

The third is absent experience. A model can describe how something is generally done. It cannot say what went wrong when you did it, which is the part that makes content worth reading and is exactly what the guidance on experience and expertise asks for.

• Statistics, studies and quotes that do not exist
• Output near-identical to competitors doing the same thing
• No first-hand experience, because the writer had none
• Volume that outpaced anyone’s ability to check it

How we assess a page, and why we ignore detectors

AI detection tools are unreliable in both directions. They flag careful human writing as machine-written and clear machine output as human, and building decisions on them means acting on noise.

So we assess what actually determines whether a page helps or hurts. Are the factual claims true and traceable. Does the page duplicate something else on your site or say the same thing as everyone else. Does it contain anything that could only come from having done the work. Does it answer the query it targets, or circle it.

That test applies identically to a page written by a person and a page written by a model, which is the correct standard. The problem was never authorship. It was publishing without checking, and a human can do that too.

The verification pass, which is the slow part

The most valuable part of this audit is the least automatable. Every specific claim gets checked.

Statistics are traced to a primary source. Where the source does not exist, or exists and says something different, the claim is removed rather than softened. Study references get the same treatment, because a fabricated citation is worse than no citation: it looks like evidence and collapses on inspection. Quotes attributed to named people are verified as having been said.

This takes time and it is where the cost of the audit sits. It is also the part that protects you, because an unverifiable claim on your site is a liability whether or not anyone has noticed it yet, and in some sectors it is a regulatory problem rather than an embarrassment.

Fixing without simply deleting everything

Blanket removal is the wrong response and it destroys pages that were nearly fine.

Pages targeting real demand with fixable problems get rewritten: claims verified or cut, genuine experience added, duplication resolved. This is the bulk of the work and the best return, because the topic was worth covering and only the execution failed. Pages overlapping others get merged into the strongest version. Pages targeting no real demand at all get removed and redirected.

Adding experience is the step people skip and it is the one that matters most. It usually means an interview: what actually happens when you do this, what goes wrong, what you would tell a client. Half an hour of that turns a generic page into one that could not have been written by anyone else, which is the whole point.

Stopping it happening again

An audit that fixes the back catalogue and changes nothing about production buys you a year.

So the output includes a standard: what AI is used for, what must be verified before publishing, who signs off, and what a page must contain before it goes live. In practice the rule that does most of the work is simply that every specific claim has a source someone has actually opened.

We are not arguing against using these tools. Used for research, structure, editing and first drafts they are a genuine gain. The failure mode is treating output as finished, and that is a process problem with a process fix.