Auditing a site that is already full of AI content

Almost everyone who asks me about AI content has already published some. Often a lot. The question is never "should I start", it is "I have got sixty pages up, some of it is fine, some of it is embarrassing, and I do not know which is which".

That is a genuinely different problem from starting clean, and most advice does not address it. This is the order I work through it, and the first instruction matters more than all the rest.

Do not start by deleting

The instinct is to find the worst pages and remove them. It is the wrong first move, for three reasons.

• Some of those pages are quietly ranking for things you did not target and did not notice, and deleting them costs you enquiries you never knew you had.
• Deletion does not fix cannibalisation on its own. If two pages compete and you delete one, you have not consolidated anything - you have thrown away whatever authority it had rather than moving it.
• You cannot tell which pages are bad until you know what each page is supposed to own, and that is an output of the audit rather than an input.

So the sequence is: measure, then decide ownership, then merge towards it. Deletion is a last step and often not needed at all.

Step 1: find what is actually happening

Open Search Console, set the range to the last twelve months, and export the Pages report and the Queries report. That is your ground truth and it costs nothing.

Sort your pages by clicks and read down the list. You are looking for three groups:

• Pages with clicks. These are working. Whatever else you do, do not break them. Note them and leave them alone until the end.
• Pages with impressions but almost no clicks. These rank somewhere nobody looks, usually position 11 to 25. They are candidates for merging into something stronger.
• Pages with neither. These are doing nothing at all, which is information rather than a verdict - a page can be doing nothing because it is bad or because nothing links to it.

Step 2: find the cannibalisation, because it goes first

In the Queries report, look for any search term where more than one of your URLs appears. In the Search Console interface you can filter to a query and then look at the Pages tab underneath it, which shows you exactly this.

Every one of those is two pages you paid for, competing. This is the most common failure in an AI buildout and the most expensive, because it degrades pages that would otherwise be fine. The mechanism, and how to avoid recreating it, is in the cannibalisation trap.

For each case:

1. Pick the page that should win. Usually the one with the most clicks already, not the one you like best.
2. Move the genuinely useful parts of the others into it - the specifics, the questions answered, anything a customer would want.
3. Redirect the losers to the winner with a 301. Do not just delete them; the redirect is what moves the value across rather than binning it.
4. Update any internal links that pointed at the losers so they point at the winner directly, rather than through a redirect.

Step 3: fact-check everything that carries a consequence

Now go through every page for invented claims. Credentials, registrations, years in business, team size, service areas, guarantees, prices, outcomes. Read the numbers hardest, because a wrong number reads exactly as smoothly as a right one.

If you are in a regulated field, do this step before step 2 rather than after. A compliance problem sitting live is more urgent than a ranking problem, and it does not care what order your audit is in.

This is also the step where you find out how bad the context problem was. If you are finding an invention on every third page, the fix is not just correcting them - it is writing the context document before you generate anything else: what AI needs to know about your business.

Step 4: check every internal link resolves

Click every internal link on every page. On a sixty-page site this is an afternoon and it always turns something up, because a model writing page 40 has no idea pages 1 to 39 exist and will invent a plausible URL rather than leave a gap.

While you are in there, note which pages have nothing pointing at them. An orphan page will not rank however good it is, and the fix is one link from a relevant page, which takes a minute.

Why this fails so reliably is in internal linking is the thing AI gets most wrong.

Step 5: rewrite the openings of your top five

Take the five pages with the most clicks and rewrite the first two paragraphs of each by hand. Not edit - rewrite.

This is the highest-return work in the whole audit, because those pages already have the traffic. Everything else you are doing is trying to earn attention; this is converting attention you already have. Thirty minutes, five pages, and it is the only step in this list that shows up in enquiries within the month.

Step 6: build the map, so the next sixty pages are not like the first sixty

Last, and the one that makes it stick: write down every page you have, with its URL and the one search term it owns. Add the terms you want that no page owns yet. Those are your genuine gaps.

Then add a row every time you publish. That single habit is what prevents the drift, and it is the difference between an audit you do once and an audit you do every eighteen months forever. Plan the site first, generate second is the long version.

What about the pages that are just bad

There will be some that are grammatically fine, factually correct, and say nothing. No clicks, no impressions, nothing links to them, and reading one tells you nothing you could not have guessed.

My honest answer is to leave them until everything above is done. They are not costing you much - a page with no impressions is not competing with anything - and the effort of dealing with them is better spent on the five pages that already have traffic. When you do get to them, the choice is improve or consolidate, and consolidating three thin pages into one useful one is almost always better than deleting three and having nothing.

I would only actually delete a page if it is factually wrong in a way that cannot be fixed, or if it is about something you no longer do. Neither of those is common.

What the site should look like when you are done

It is worth knowing what you are aiming at, because "audit the site" is open-ended enough to never finish.

• Every page can answer, in one phrase, which search it owns - and no two pages give the same answer.
• No search term in Search Console returns more than one of your URLs.
• Every claim on the site is one you could evidence if somebody asked.
• Every internal link resolves, and every page has at least one other page pointing at it.
• Your top five pages open in your own voice rather than an average of your industry's.
• There is a sheet, and it gets a new row every time you publish.

Nothing on that list is about writing quality, and that is deliberate. A site that clears all six with mediocre prose will outperform a beautifully written one that fails half of them, because five of the six are structural and structure is what was missing.

Roughly what this costs

For a sixty-page site with a person who knows the business, this is two to three days of work spread over a fortnight. Step 2 is the longest. Step 5 is the shortest and the one you feel first.

That is the number worth holding on to if you are still deciding how fast to publish. The audit costs far more than the writing did, it arrives whether you plan for it or not, and it is the reason thirty posts a month is the red flag rather than the writing quality. The structure that stops it recurring is in using AI for your website content without wrecking it.