Check this page with an assistantOpens a chat asking it to summarise this article and name the evidence behind each claim.
Claude opens with the prompt on your clipboard: Anthropic does not support prefilled prompts on the web, and we would rather copy it than ship a button that drops it.
The Shape of the Failure
It rarely looks like rejection. Generated pages get indexed, pick up long-tail positions, and grow slowly for months, which reads as a content programme working. Then a core update lands and a large share of it goes at once, because the pages were always similar and get reassessed similarly.
The delay is what makes this dangerous rather than merely disappointing. Six months of apparent success is long enough to justify scaling the approach, hiring against it, and building a plan on the traffic. The correction arrives after the commitment, and it arrives as a cliff rather than a slope.
Sites that had built substantial traffic this way have reported losing most of it within a fortnight of an update. Treat the specific percentages circulating with caution, since they come from vendor analyses of self-selected cases, but the shape is not in dispute and it has repeated across several updates now.
It Is Not Detection, It Is Assessment
The common framing is that Google got better at spotting AI writing. That is not what the policies describe. Google's stated position is that it rewards quality regardless of how content is produced, while scaled content abuse, producing pages primarily to rank rather than to help, is a violation.
The distinction is practical rather than semantic. If the problem were detection, the fix would be making AI writing less detectable, and an industry of humanisers exists on that premise. If the problem is that the pages are generic and unverified, making them read more naturally changes nothing, because the assessment is about what the page offers rather than its prose style.
The evidence for the second reading is that human-written thin content fails identically and always has. Content farms predate language models by fifteen years and were caught by the same kind of reassessment. AI did not create this failure mode; it made it cheap enough that far more sites walked into it. The policy detail is in does Google penalise AI content.
What Separates Durable From Disposable
One question does most of the sorting: is there anything on this page a competitor could not generate in thirty seconds? Everything that survives has an answer, and everything that decays does not.
| Property | Decays | Holds |
|---|---|---|
| Facts | Plausible, unverified, occasionally wrong | Checked against primary sources, dated |
| Experience | Generic advice applicable to anyone | What you observed doing this specifically |
| Data | Statistics borrowed from other articles | Numbers you produced, per original research |
| Accountability | No named author, or an invented one | A real person who stands behind it, per author entities |
| Reason to exist | A keyword had volume | Someone actually asked |
| Publishing pattern | Large volumes at a constant rate | Paced with what could genuinely be reviewed |
The publishing pattern row is the one that turns individual page quality into a site-level risk. Volume that could not plausibly have been reviewed is visible in aggregate even when each page looks acceptable alone, and it is exactly what the scaled-content provisions describe.
Auditing Your Own Exposure
You do not need a detector. You need to know how much of your site consists of pages that could be regenerated by anyone, and whether those pages are carrying traffic you are counting on.
- Find the cohort, not the pages.
Sites lose groups rather than scattered individuals. Sort by publishing period, template, and author, and look for the cluster that shares a trait, per page-level authority.
- Apply the thirty-second test to a sample.
Read ten pages from the cohort and ask what is on each that could not be regenerated. If you cannot answer for eight of them, you know the size of the exposure.
- Check the facts on the ones that matter.
Pick the pages carrying real traffic and verify their claims against sources. Confidently wrong facts are the most common defect in generated content and the most damaging when found by a reader.
- Decide merge, fix, or remove per cluster.
Not per page, which does not scale. The decision framework is in the consolidation playbook.
- Slow the publishing rate to what you can review.
This is the structural fix. A rate that permits genuine verification produces a different kind of site than one set by how fast drafts can be generated.
Using AI Without Building the Exposure
The distinction that holds up is AI as a production tool versus AI as a replacement for expertise. Drafting, restructuring, summarising your own material, and first passes are the first. A page that exists entirely because generation was cheap is the second.
Practically that means the model produces the draft and a person supplies the parts it cannot: verified facts, the specific case, the thing that surprised you, and accountability. That review is not overhead to be optimised away; it is the entire difference between the two categories. The working method is in editing an AI draft into something publishable.
It also means accepting a lower ceiling on volume. If review is what makes content durable, then review capacity is the real constraint on publishing rate, and a programme sized to generation capacity is sized wrong. That is an unwelcome conclusion for anyone who bought tooling on a volume premise, and it is what the pattern keeps demonstrating.
Questions About AI Content Decay
- Why does AI-written content rank at first and then disappear?
Because fluent content clears the initial bar and accumulates positions quietly, then gets reassessed in bulk when a core update changes how quality is judged. The content did not change; the assessment did. Sites that had built traffic on volumes of generated pages have reported losing most of it within a fortnight of an update.
- Is Google detecting that content is AI-written?
Google says it rewards quality regardless of production method rather than targeting AI authorship itself. Its policies target scaled content abuse: mass-producing pages primarily to rank rather than to help. What gets caught is the publishing pattern, which is why human-written thin content fails the same way.
- What separates AI content that survives from content that does not?
Whether a person added something the model could not produce. Drafts that were fact-checked against primary sources, given specific experience, and edited by someone accountable tend to hold. Pages that are fluent, generic, and unverified are the ones that accumulate quietly and disappear together.
- Can I recover pages that lost traffic in an update?
Sometimes, and it depends on why they lost. If the pages answer a real question badly, improving them genuinely works. If they exist because a calendar needed filling and no reader ever wanted them, there is nothing to improve toward and removal or consolidation is the honest answer.
- How do I know whether my content is exposed?
Ask what is on the page that a competitor could not generate in thirty seconds. If the answer is nothing, the page is exposed regardless of who wrote it. That test is more useful than any AI detector, which measure style rather than value and produce false positives on careful human writing.
