Google's spam policies name scaled content abuse: producing many pages primarily to manipulate search rankings rather than to help people.
Two words in that definition do the work, and most coverage skips both.
Primarily. The test is about intent as expressed in the output, not about volume alone. A site publishing 400 pages a year is not automatically in breach. A site publishing 400 pages that exist to catch queries is.
Spam. This sits in the spam policies, not the ranking guidance. Ranking factors move you up and down. Spam violations trigger manual actions, and manual actions do not taper. They land.
The question the policy does not ask
It does not ask who wrote the page.
Google has said repeatedly that AI-generated content is not against the guidelines when it is useful. That statement is accurate and it gets misread constantly, usually as permission. What it actually means is that authorship was never the criterion, so a human-written page of nothing is in exactly the same category as a generated one.
The criterion is whether the page contains anything a reader could not get elsewhere. That is a harder bar than "is it well written", and it is the bar most AI-assisted content fails, because a model with no access to your data, your customers or your product cannot produce a claim that is specifically yours.
What the failing pages have in common
Look at what actually gets deindexed and a pattern shows up that has nothing to do with word count.
The page restates the query and calls that an introduction. Three paragraphs establishing that the topic exists and matters, before anything is said about it.
Every section is a definition. "What is X." "Why X matters." "Benefits of X." A page structured entirely around explaining terms is a page with no position.
No numbers that came from anywhere. Statistics appear, but they are the same statistics on every competing page, traceable to the same 2019 blog post that never cited a source.
Nothing is refused. A page that recommends everything recommends nothing. Real expertise shows up as exclusion, because someone who has done the work knows which approach fails and says so.
The conclusion summarises. A closing section that repeats the headings back to the reader adds no information and signals that the page was assembled rather than written.
Every one of those survives a rewrite that swaps out the vocabulary. That is why word-level editing does not save a page from this policy.
The recovery problem
Manual actions for scaled content abuse are unpleasant to unwind, and the reason is arithmetic.
A site that published 400 thin pages has to deal with 400 thin pages. Improving them individually costs more than the pages were ever worth. Deleting them tanks the site's page count and any residual traffic. Most recoveries involve pruning hard, which means accepting that the content spend was a write-off.
The reconsideration process then requires demonstrating that the pattern has stopped, and reviewers assess the site as it stands rather than the intent behind it.
Which makes this a policy where prevention and cure differ by orders of magnitude in cost. Getting it wrong for a quarter can take a year to clear.
Publishing at volume without triggering it
Volume is not the violation. Volume of nothing is, and the difference between a site publishing 400 useful pages and a site publishing 400 thin ones comes down to three things that have nothing to do with how many were published.
One original input per page. A number you measured, a customer you spoke to, a test you ran, a thing that broke. If a page contains no input that originated with you, it has no reason to rank above the page that does.
A position, stated. Recommend something and rule something out. "Use X for teams under twenty, and do not use it above that because the permissions model does not scale" carries more authority than four paragraphs of balanced comparison.
Cut the definitional sections. If your reader searched for a comparison, they know what the thing is. The definition paragraphs exist for the word count, and the word count is the tell.
Where AI fits without causing the problem
Models are useful in this workflow, just in a narrower place than most content operations put them, which is the part that gets skipped when someone decides to scale a blog.
They are good at structure, at drafting around facts you supply, at compressing something you wrote too long. They are also good at telling you what you left out. They are not good at supplying the input, because they do not have it.
The failure mode is asking a model to produce the whole page from a keyword. What comes back is a competent summary of what is already ranking, which is the definition of a page that adds nothing.
The workable order is to bring the substance first and let the model handle the shape. It is slower per page. It also produces pages that survive a policy update, which the fast order does not.
Checking a page before you publish it
The mechanical markers correlate closely with the pages that get flagged, and they can be measured.
The free workbench scores a draft on four of them. It measures how densely the page uses the vocabulary that machine writing over-selects, how much its sentence lengths vary against the 0.60 human threshold, its em-dash rate per thousand words, and whether the opening line matches one of the stock formulas that models reach for.
None of that measures whether your page says something, which remains a judgment only you can make. What it does measure is whether the prose carries the signature of text assembled to fill a slot, and in practice the two correlate more than publishers expect.
The upstream fix is the same as everywhere else. Give the model an explicit ban-list rather than a request to write well, cap the em dashes, require sentence-length variance. Then insist that every section carry a claim a reader could check. That last constraint is the one that maps directly onto this policy, because a page where every section makes a checkable claim is a page that says something.