Pre-Publish Review for Scaled AI Content: A Gate Checklist

Google's spam policies define scaled content abuse as "when many pages are generated for the primary purpose of manipulating search rankings and not helping users." The first example Google lists is "Using generative AI tools or other similar tools to generate many pages without adding value for users."

The tool is not the problem. Google's guidance on AI content says "Appropriate use of AI or automation is not against our guidelines." The policy applies "no matter how it's created," so a thousand thin pages written by humans carry the same risk as a thousand written by a model.

That changes what a review should test. Detecting AI text is beside the point. The question is whether each page earns its place. For a side-by-side view of the two approaches, see scaled AI content vs people-first content.

The stakes reach past classic search. Google states that a page must be "indexed and eligible to be shown in Google Search with a snippet" to appear as a supporting link in AI Overviews or AI Mode. A page removed from the index loses both.

OtterlyAI published one of the few public tests of mass AI publishing in 2026. It launched two new domains about capybaras, each with roughly 1,000 AI-generated posts. One site showed a fictional wildlife biologist in an author box. The other had no author.

Otterly reports that Google deindexed the author-box site after about 14 days and the other after about 90. Neither site received a manual action in Search Console. Bing kept both sites, and Microsoft Copilot produced 2,718 of the 3,934 AI citations Otterly tracked for the author-box site. ChatGPT, Gemini, AI Overviews and AI Mode almost never cited either site.

Treat this as one data point. The limits matter:

• One site per arm. The domains differed in more than the author box, so the author result cannot be isolated.
• No human-written control. The test cannot say whether Google reacted to AI text or to a thousand pages appearing at once on a domain with no history.
• A low-stakes topic. Capybara questions have little competition, so the Copilot result says little about commercial categories.
• A vendor ran it. Otterly sells AI visibility tracking. That does not make the data wrong, but replication matters.

The method and limits of the Otterly experiment are covered in full elsewhere. For a pre-publish gate, the useful lessons are narrow: removal can be silent, fake expertise did not protect the site, and losing Google's index removed the pages from Google's AI answers.

Run these steps on every batch before it goes live. A batch is any set of pages produced from one template, prompt or data feed.

1. Write down why the batch exists. Name the reader and the question each page answers. If the only honest reason is "to rank for more keywords," stop here. That is the pattern the policy describes.
2. Read pages in full. In a small batch, read every page. In a large one, read a random sample plus every page built from the thinnest input data. Templates fail first where the data is sparse.
3. Check for near-duplicates. Compare pages for shared paragraphs and swapped-in names. A city or product name dropped into identical text adds nothing for the reader.
4. Verify every fact against your source of truth. Prices, specs, dates and claims must match your pricing page and product records. Cut any statistic the model produced without a source.
5. Check that crawlers can read the page. Confirm the main content is in the raw HTML before JavaScript runs, and that no template has added noindex or nosnippet by accident.
6. Plan the release. Pushing a full batch at once onto a young domain matches the Otterly setup. A staged release is a judgement call, not a proven safeguard, but it lets you see index problems before every page is exposed.
7. Decide and record. Mark each page publish, revise or cut, and note who decided. That record is what you check against if index status drops later.

Every page in the batch should pass each item. One failure means revise or cut.

• The page answers a question real customers ask, taken from sales calls, support tickets or Search Console queries.
• It states at least one specific, verifiable fact: a price, a number, a spec or a local detail. See publishing specific, fetchable facts.
• No fact contradicts another page on your site.
• No statistic appears without a named source.
• The page is not a near-copy of another page with a name swapped.
• A real, accountable person reviewed it. No invented experts or fictional credentials.
• You would publish it even if search engines did not exist.
• The main content is in the HTML before JavaScript runs.
• No accidental noindex, nosnippet or data-nosnippet on content you want shown.
• Internal links point to real, relevant pages.
• No hidden text, cloaking, or instructions aimed at AI crawlers. These tactics are manipulation, and they put the whole site at risk.

Two items people expect are missing on purpose. Schema markup and "AI-friendly" formatting are not on the list. Google's AI optimization guide says you don't need to write in a specific way for generative AI search, and Google says E-E-A-T is not a specific ranking factor. Neither is a gate for scaled content.

A batch that passes

• Each page has a reason to exist that you can state in one sentence.
• Pages differ in substance, not just in the names they contain.
• Facts trace back to a record you control.
• Reviewers cut or rewrote a share of pages. A gate that passes everything is not checking.
• After launch, Search Console and Bing Webmaster Tools both show the pages indexed.

A batch that fails

• Hundreds of pages share the same paragraphs with one variable changed.
• Pages carry numbers nobody can source.
• The byline is a persona with credentials nobody holds.
• The plan is to publish everything on day one and see what sticks.
• After launch, impressions climb and then collapse. Otterly's Search Console data showed one site fall from 1,629 to 15 daily impressions in a day, with no message explaining why.

If you see the last pattern, check index status engine by engine. In the Otterly test, Bing's health told you nothing about Google's. For more on where automated publishing breaks down, read the quality bar for AI SEO articles.

• Every batch, before publishing. The gate is not a one-time audit. Each new template run gets the full steps.
• Weekly in the first weeks after a launch. Otterly's first site lasted about 14 days in Google. Check Search Console and Bing Webmaster Tools separately during that window.
• Whenever the inputs change. A new model, prompt, template or data source can change output quality across every page at once. Treat it as a new batch.
• Quarterly for live programmatic pages. Facts go stale. Re-verify prices, specs and claims against your records.

After the pages settle, check whether AI answers describe you correctly. The free checker on this site shows whether AI engines name you for up to three buyer questions, as a snapshot.

The gate controls what you publish. It does not control what Google indexes or what any AI engine cites. Google states that indexing and serving isn't guaranteed, even for eligible pages.

The review also cannot tell you how Google's systems weigh a batch. Otterly's sites were removed algorithmically, and Google does not say which system acted. What you can do is avoid the pattern the policy names, keep facts true, and keep pages readable to crawlers.

Volume itself is not a lever. Before scaling up output, read whether publishing more content helps AI visibility and whether Google penalizes AI content at all.

Does Google penalize content just because AI wrote it? No. Google's guidance says appropriate use of AI or automation is not against its guidelines. Its spam policies act against scaled content abuse: many pages produced mainly to manipulate rankings without helping users. The policy applies however the content was made, so human-written pages at scale with no added value carry the same risk as AI-written ones.

Will an expert author box protect scaled AI content? Nothing in the evidence suggests it will. In Otterly's test, the site with a fictional wildlife biologist in its author box was deindexed after about 14 days, and the site with no author lasted about 90. One site per arm cannot prove author boxes hurt. It does show a fake expert offered no protection, and inventing credentials is a risk on its own.

Can AI engines still cite pages Google has removed? Engines that draw on a different index can. In Otterly's test, Bing kept both sites and Microsoft Copilot kept citing them after Google removed them. Google's AI Overviews and AI Mode almost never cited the sites, which fits Google's rule that supporting links must come from indexed, snippet-eligible pages. Check each engine's index separately.

Is there a safe number of pages to publish per day? No published evidence gives one. Google's policy is about purpose and value, not a page count. The Otterly test published about 1,000 pages per new domain at once, but with no human-written control it cannot separate volume from content quality. A staged release is a reasonable precaution because it exposes problems earlier, not because a threshold is known.

Should I add schema to scaled pages before publishing? Schema is not a gate for scaled content and not a lever for AI answers. Google says structured data is not required for generative AI search, and an Ahrefs matched test found AI Overview citations fell slightly after pages added it. Existing markup does no harm. Spend review time on facts, uniqueness and crawler access instead.