AI-written articles and Google's scaled content abuse policy: how much is safe to publish?
What the policy says, what changed in 2026, the effort test section by section, and the drift signals that precede a manual action.
Google's scaled content abuse policy prohibits generating many pages primarily to manipulate rankings rather than help users — regardless of whether a model or a person wrote them. There is no numeric ceiling on AI-written articles. Volume becomes a risk when it outruns review: when pages are published faster than anyone can verify facts, add first-hand detail and check that each page answers a real query. The safe velocity is whatever your editorial capacity can actually review, not whatever the model can produce.
- The policy targets intent and effort, not the production tool. "Many pages, primary purpose: rankings" is the test.
- The March 2026 core and spam updates hit sites with large volumes of unreviewed AI text, not sites using AI with oversight.
- Google's quality-rater guidance now explicitly lists "large amounts of generative AI text without manual oversight or curation" as little-to-no effort.
- A practical velocity rule: publish no faster than one qualified reviewer can read, fact-check and improve each page — in enterprise practice this is roughly 15–40 articles per reviewer per month.
- Risk is per-site, not per-page. A hundred thin pages can drag down the whole domain's quality assessment.
01What the policy actually says
Google's spam policies define the violation precisely, and the definition contains no reference to AI:
“Scaled content abuse is when many pages are generated for the primary purpose of manipulating search rankings and not helping users.”
The same policy lists examples: using generative tools to create many pages without value for users, scraping and rewriting, stitching feeds together, and — the one most teams forget — producing pages with "little to no value" at scale even when a person wrote them. The policy was expanded in March 2024 specifically to drop the old "automatically generated" wording so that human-made spam at scale would be covered too.
The word "primary" matters. A site that publishes 200 AI-assisted articles because 200 real customer questions exist is not in scope. A site that publishes 200 articles because a keyword tool returned 200 rows is.
02What happened in 2026
The February and March 2026 core and spam updates were the clearest enforcement of the policy so far. The pattern in the sites that lost: a high ratio of pages to authors, near-identical structure across hundreds of URLs, no author pages, no original data, and publishing dates clustered in bursts. The pattern in the sites that held: AI used for drafts, visible human editing, author bylines, and a publishing cadence that matched the size of the team.
None of this required Google to detect AI. It required Google to detect low effort, which its quality guidance now defines in exactly those terms:
“Producing large amounts of generative AI text without manual oversight or curation counts as little to no effort.”
03The effort test, section by section
| Signal | Scaled abuse pattern | Legitimate scale pattern |
|---|---|---|
| Why the page exists | Keyword list | Documented customer question or sales objection |
| Structure | Identical template, swapped nouns | Varies with the question; some pages are short |
| Facts | Model's training data, unsourced | Named source, date, checked by a person |
| Author | None, or an AI byline | Real person with a profile page |
| Publishing rhythm | 100 pages in a week | Steady cadence matching reviewer capacity |
| Internal links | Mechanical, every page to every page | Editorial, from the question to its answer |
| Updates | Never | Refresh cycle with changelog |
The left column is what a content factory looks like when the factory has no editors. The right column is what the same factory looks like with them. The difference is not the model; it is the throughput of the review layer.
04How many AI-written posts per month is safe?
There is no number from Google and anyone who gives one is guessing. What can be stated with confidence is the shape of the limit:
The review layer, done to the standard Google now asks for, takes time. Google's own guidance updated in October 2026 raises the bar further:
“critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing”
"All" includes titles, meta descriptions, structured data and alt text. A reviewer who does this properly on a 1,500-word article needs 45–90 minutes: read, verify every claim, add one first-hand observation, fix the first sentence of each section, check schema. At that pace one full-time reviewer handles roughly 15–40 articles per month depending on topic complexity. That is the ceiling for one seat — not 300.
The failure mode is not publishing 40 a month. It is publishing 40 a month with a reviewer who has capacity for 15, so that 25 go out unread. Google does not see the 15 good ones separately from the 25 bad ones; it sees one site.
05Signals that you are drifting into scaled abuse
Two "yes" answers on the wrong side of that list is the point to slow down and consolidate. Thinning a site — merging five weak pages into one strong one — is the recovery action that worked after 2026 updates, and it is also the action that makes the remaining pages citable in AI answers.
06Velocity and AI visibility are the same problem
Everything that trips the scaled abuse policy also makes a page useless for AI citation: unsourced claims, template structure, no author entity. The reverse also holds — the editorial standard that makes a page citable by ChatGPT (what the engines actually pick) is the standard that keeps a site clear of the policy. Treat velocity as an output of the review process, not an input to it, and both problems resolve together.
For the operating model that holds this in place across Legal, Brand and regional teams, see Running AI visibility through Legal, Brand and regional sign-off.
Questions people also ask
Does Google penalize AI-written articles?
No. Google's guidance since 2023 states AI content is not against its guidelines. It penalizes content produced at scale primarily to manipulate rankings, under the scaled content abuse policy, whoever or whatever produced it.
What counts as scaled content abuse?
Many pages generated primarily to rank rather than to help users: generative-AI pages without value, scraped and rewritten content, stitched feeds, and mass low-value pages even when human-written.
How many AI blog posts can I publish per month?
Google sets no number. The safe limit is the number your reviewers can fully fact-check and improve — roughly 15–40 per reviewer per month in enterprise practice. Publishing beyond review capacity is the pattern that lost traffic in the 2026 updates.
Do I have to label content as AI-generated?
Google recommends disclosure where a reader would reasonably wonder how the content was made, but does not require a label for ranking purposes. Giving AI an author byline is discouraged; use a real author.
