How to Audit a Site Full of AI Content: A Prioritized Recovery Checklist

Maybe you inherited the site. Maybe a past version of you got a little too excited about generating 400 posts in a weekend. Either way, you’re now staring at a content library you didn’t read, traffic that’s flat or sliding, and a nagging sense that some of these pages are doing more harm than good. The instinct is to fix everything. Don’t — that’s how you burn a month and finish nothing.

The job isn’t to scrub the “AI-ness” out of every page. It’s to find the handful of pages that matter, decide what to do with the rest, and rebuild quality where it counts. Here’s the order to do it in.

Key takeaways

• You can’t fix a whole site at once. Triage first, then spend your hours on the pages that actually move traffic or carry the most risk.
• The problem with mass AI content is rarely “it’s AI” — it’s thin depth, duplicate angles, and pages that don’t answer the query. Audit for those, not for an AI score.
• Pull data before you touch anything: Search Console impressions and clicks, indexed page count, and a crawl. You need a baseline to know if recovery worked.
• Sort every page into keep, improve, merge, or delete. Most sites end up consolidating or deleting more than they expect.
• Humanizing stiff prose is the last 10%, not the fix. Add real information first, then smooth the style.

First, get your baseline (don’t skip this)

You can’t tell if a recovery worked without a “before.” So before you edit or delete a single page, capture the numbers.

Pull these:

• Search Console, last 12 months: queries, pages, impressions, clicks. Export the Pages report.
• Indexed page count from the Pages (Index Coverage) report. Compare it to how many pages you think you have. A gap usually means thin pages Google chose not to index — useful signal.
• A full crawl with Screaming Frog, Sitebulb, or the free tier of an SEO crawler. You want titles, word counts, and a list of every URL.
• Analytics, if you have it: which pages get any human visits at all.

Now you have a map. The next steps are about reading it.

Sort every page into four buckets

Don’t grade pages on whether they “sound like AI.” Grade them on whether they earn their place. As you go through the crawl, drop each URL into one of four buckets.

Keep. It ranks, gets clicks, answers the query, and reads like a person wrote it. Leave it alone for now.

Improve. It targets a real query and gets impressions, but it’s thin, generic, or stiff. This is where your editing time pays off.

Merge. Three posts all circle the same topic from slightly different angles, splitting the authority and confusing Google about which to rank. Combine them into one strong page and redirect the rest.

Delete. No impressions, no clicks, no internal links worth keeping, no unique value. It exists because a generator made it. Remove it (return a 410, or 301 to the closest relevant page).

A quick way to sort fast: in your crawl export, flag anything under ~400 words and anything with zero clicks in 12 months. Those are your delete and merge candidates. Pages with steady impressions but a weak click-through rate are your improve list — they’re knocking on the door and just need to be better.

What “bad AI content” actually looks like

People assume the issue is detectability. It usually isn’t. Google has never said it ranks by an AI-detection score, and detectors are unreliable anyway — they spit out a probability, not a verdict. The real problems are concrete and findable:

• Thin depth. The page restates the question three ways and never answers it. No examples, no numbers, no first-hand take.
• Sameness. Every post opens the same way, hits the same beats, and lands on the same vague conclusion. A reader who clicks two of them feels it instantly.
• Hallucinated facts. Confident, specific, and wrong. Stats with no source, fake quotes, invented product features. This is a credibility and liability problem, not just an SEO one.
• Query mismatch. The page is “about” the topic but doesn’t do the thing the searcher wanted — no recipe in the recipe post, no steps in the how-to.

That last one is the heart of Google’s helpful-content guidance: content should be made for people, satisfying what they came for. Mass-produced pages built mainly to rank, with no added value, are what Google’s spam policies call scaled content abuse. The line isn’t “a model helped write it.” The line is purpose and value.

If you want to understand why these pages read flat in the first place, it’s worth knowing how detectors measure perplexity and burstiness — the same statistical evenness that trips detectors is what makes the prose feel hollow to humans, too.

A mini-scenario: 300 posts, one good month

Say you took over a personal-finance blog with 312 AI-generated posts. The crawl shows 47 pages have any clicks at all in the last year. The other 265 have a handful of impressions or none.

Here’s the triage:

• The 47 with clicks go into keep or improve. You read each one. Twelve are fine. Thirty-five are thin but ranking on page two — those get rewritten with real numbers, an example, and a clear answer up top.
• Of the 265 dead pages, you find 40 clusters where 4–6 posts cover nearly the same keyword. Each cluster gets merged into one solid page; the extras 301-redirect to it.
• Whatever’s left — roughly 180 truly empty pages — gets deleted (410) or redirected if a sensible target exists.

You’ve gone from 312 weak pages to maybe 90 real ones. That’s not a loss. A smaller site full of pages that answer questions beats a huge one full of filler, and it sends a cleaner quality signal across the whole domain.

Fix the “improve” pages in the right order

For the pages worth saving, depth comes before style. Always.

1. Answer the query in the first two sentences. If someone asked “how,” give a step. Don’t warm up.
2. Add what a model can’t. A specific number, a screenshot, a real example, a mistake you made, a comparison you actually ran. This is the part that separates your page from the next AI draft.
3. Cut the padding. Delete the throat-clearing intros and the “in conclusion” wrap-ups. Anything that says nothing, goes.
4. Check the facts. Verify every stat and claim. Replace or remove anything you can’t source.
5. Then smooth the prose. Only now do you fix the robotic rhythm — the relentless even sentences, the stacked “moreover,” the parallelism that never breaks.

That last step is where a humanizer earns its keep. Once the substance is solid, you can smooth the stiff prose on the pages worth keeping so they read like a person instead of a template. Just don’t reverse the order. Humanizing a thin page gives you a thin page that reads nicely — still thin.

Track recovery against your baseline

After the cleanup, watch the numbers you captured at the start. Recrawl in a few weeks and confirm the deleted URLs are gone from the index. Watch whether impressions consolidate onto your merged pages. Give it time — Google recrawls and reassesses on its own schedule, and sitewide quality signals move slowly, often over a couple of months.

Resist the urge to publish 200 new posts to “make up” for what you deleted. That’s the exact behavior that got the site here. Build slowly, publish pages you’d be comfortable putting your name on, and let the domain’s quality reputation recover.

If you’re curious how thin AI drafts go wrong in the first place, our post archive digs into the depth gaps and the editing passes that fix them.

Frequently asked questions

Should I just delete all the AI content and start over?

Usually no. Some AI-assisted pages rank fine because they answer a real query well enough. Deleting indiscriminately throws away pages that bring traffic and the internal links pointing at them. Pull your Search Console data first, find the pages with impressions or clicks, and treat those as keep-or-improve candidates. Save deletion for pages that are thin, duplicative, and getting zero attention from search or users.

Will an AI detector tell me which pages to fix?

Not reliably. Detectors output a probability, not proof, and they’re known to be inconsistent — OpenAI retired its own classifier in July 2023 for low accuracy, and a 2023 Stanford study found detectors biased against non-native English writers. A high “AI” score doesn’t mean a page is bad, and a low score doesn’t mean it’s good. Audit for thin content, duplicate angles, and pages that don’t satisfy the query. That’s what hurts you, and it’s what you can actually measure.

How long does it take to see recovery after an AI content audit?

There’s no fixed timeline. Google has to recrawl and reassess changed or removed pages, which can take weeks, and sitewide quality signals shift slowly. If you deleted a lot of thin pages, you may see crawl budget redirected to your better pages fairly quickly. Substantial ranking recovery, especially after a quality issue, often takes a couple of months or more. Track it against the baseline you captured before starting.

Does humanizing the AI text fix the SEO problem?

It fixes one layer of it. If the writing reads stiff and robotic, smoothing it helps real readers and removes obvious tells. But humanizing doesn’t add facts, examples, or first-hand experience, and that thinness is usually the real ranking problem. Treat style cleanup as the final polish after you’ve added genuine value to the pages worth keeping.

Closing

A site full of AI content isn’t a disaster — it’s a backlog with bad sorting. Get your baseline, bucket every page into keep, improve, merge, or delete, then pour your real effort into the pages that already have a pulse. Depth first, style last.

When you reach that final polish step and need stiff drafts to read like a human wrote them, run them through PaperBleach and see the difference before you republish. Editing a whole library? Check the plans and pricing.