AI-Generated SEO at Scale: Where It Breaks and How to Stay Clear

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

Google's guidance on AI content is equally clear that the tool is not the problem: "Appropriate use of AI or automation is not against our guidelines." The policy applies to the pattern, however the pages were produced.

What this means in practice:

• A thousand thin pages on a new domain is the pattern Google describes, whether a model or a contractor wrote them.
• A smaller set of AI-assisted pages that each answer a real question with something new is not.
• The question Google is asking is "who is this page for?", not "who typed it?"

For the full reading of Google's position, see Google's guidance on AI-generated content, digested and does Google penalize AI content. The next move: before you generate anything, write down what each page gives a reader that no existing page on your site already does. If you cannot, do not publish it.

The clearest public test is OtterlyAI's 2026 experiment. Otterly launched two new domains about capybaras and loaded each with roughly 1,000 AI-generated posts in one click. One site, capybara-world.com, showed an author box presenting a fictional wildlife biologist. The other, capybaraverse.com, had no author.

Otterly reports that Google removed capybara-world.com after about 14 days and capybaraverse.com after about 90. Otterly also reports zero manual actions in Search Console on either domain. Whatever removed them was algorithmic, and Google does not say which system it was.

The collapse shows up in impressions, not in a notice. Otterly's Search Console data shows capybara-world.com falling from 1,629 to 15 daily impressions on April 9. Otterly's data shows capybaraverse.com falling from 859 to 85 daily impressions on June 26, then to single digits.

What this means

An empty Manual Actions report does not mean your site is fine. A manual action comes with a message you can respond to. An algorithmic removal does not. The difference is laid out in manual action vs algorithmic deindexing.

Read the test for what it is

It is one pair of sites on a low-competition topic, with no human-written control. OtterlyAI sells AI visibility tracking, as does 99Visibility. The finding that holds is the one that matches Google's written policy: mass pages without value, removed without warning.

Google states that to appear as a supporting link in AI Overviews or AI Mode, a page must be "indexed and eligible to be shown in Google Search with a snippet." There is no separate route in. When Google drops a page from its index, its AI surfaces lose access to it as a source.

Other engines draw on other indexes, so the outcome splits. Here is what Otterly reported for its test:

Two lessons follow. First, index status is per engine: a healthy Bing index tells you nothing about Google, and the reverse. Second, being indexed is necessary but not sufficient. "ChatGPT uses Bing" is repeated as a strategy, yet in this test ranking in Bing produced no ChatGPT citations. Platform behavior changes, so treat these splits as a snapshot of 2026, not a fixed rule.

Separate the four stages and the picture gets simpler. What a model learned in training (parametric) is frozen. Whether a crawler can reach and store your page (access) is largely yours. Whether a page gets pulled in for a query (retrieval) and whether it gets cited (generation) belong to the engine.

In your control

• What you publish, how many pages, and how fast.
• Whether a person reviews each page before it goes live.
• Whether the facts on each page are correct and consistent with the rest of your site.
• Whether robots.txt, your firewall and your CDN let each crawler in.
• Whether you watch Google Search Console and Bing Webmaster Tools, separately and often.

Not in your control

• Which algorithmic system acts, or when.
• Whether you get any notice before or after a removal.
• Whether any AI engine retrieves or cites a given page.
• How Copilot, ChatGPT, Perplexity or Claude weight what they retrieve.

No tool, tactic or vendor can promise you a citation. What you can do is avoid giving Google a reason to remove you, and make sure nothing technical blocks the engines that might use you.

Generation is safe when a person supplies the value and the model saves time. These uses stay on the right side of Google's line because each page carries something a reader needs:

• Drafting from your own material. Product docs, support tickets, spec sheets and internal data, turned into a draft that an editor checks fact by fact.
• Summarizing data only you hold. Say you run a dental clinic in Austin and track appointment wait times. A model can help write up the pattern. The numbers are what make the page worth reading.
• Translation and localization, reviewed by a fluent speaker before publishing.
• Supporting work. Outlines, briefs, title options and meta description drafts that a person rewrites.
• Updating pages you already have instead of creating near-duplicates.

One on-page habit has a small positive signal in the evidence: stating a specific fact (a price, a number, a one-line definition) on a page an engine can fetch gives it a sentence it can quote. That is not a guarantee of anything.

What generation does not do is earn AI visibility on its own. The question is covered in will AI content automation tools improve my AI search visibility, and the side-by-side is in scaled AI content vs people-first content.

1. Check access first. Confirm robots.txt allows the retrieval bots you want (OAI-SearchBot, PerplexityBot, Claude-SearchBot, Googlebot, bingbot), and that your firewall or CDN does not return 403 to them. Access can change without you editing anything, as this question on crawler access explains. 99Visibility's free AI Crawler Checker requests a page as each AI crawler and shows which ones get turned away.
2. Set a value test per template. For each page type, define what a reader gets that they cannot get elsewhere. Use a pre-publish review for scaled AI content so every page passes before it ships.
3. Publish in batches you can actually review. Google publishes no safe page count, and no study in the evidence sets one. Pace output to your editing capacity, not your generation capacity.
4. Watch impressions, not just the Manual Actions report. In Otterly's test the only warning was a daily impression collapse. Check Search Console and Bing Webmaster Tools separately.
5. If impressions collapse with no manual action, stop publishing. Review what you added most recently and cut or merge pages that fail your value test. No published data measures recovery time, so treat any timeline as unknown.
6. Measure each engine on its own. One engine cited the test sites thousands of times while four barely did. A blended number would have hidden that.

• Adding a fictional expert. Otterly's site with an invented wildlife biologist was deindexed first, at about 14 days. One site per arm cannot prove author boxes hurt, but it is no evidence they help. Google says E-E-A-T "isn't a specific ranking factor", and an invented credential misleads readers.
• Waiting for a notice. Neither test site got a manual action. Silence in Search Console is not a clean bill of health.
• Treating Copilot citations as a win. Copilot kept citing pages Google had removed. That does not make up for losing Google's search and AI surfaces.
• Assuming one index covers everyone. Ranking in Bing produced no ChatGPT citations in the test.
• Rewriting pages to chase citations. C-SEO Bench, a NeurIPS 2025 benchmark, found most conversational SEO methods "largely ineffective" and frequently harmful to ranking in the model's context.
• Using hidden text or instructions aimed at AI crawlers. Injected content is illegitimate, it can be detected, and it puts the whole domain at risk.

Does Google penalize content because it was written by AI? Not for being AI-written. Google's guidance says appropriate use of AI or automation is not against its guidelines. It acts against scaled content abuse: many pages generated mainly to manipulate rankings without adding value for users. In OtterlyAI's experiment, Google removed two 1,000-post AI sites, and Otterly reports no manual action on either. The trigger was the volume-without-value pattern, which Google's policy targets however the pages are made.

Is there a safe number of AI pages to publish per day? No published number exists. Google does not set a page count, and no study in the evidence base measures a threshold. The policy is about purpose and value, not a quota. The practical limit is your review capacity: publish only as many pages as a person can check for accuracy and usefulness. If that number is small, the right output is small.

Can AI engines still cite a site after Google deindexes it? Engines that use a different index can. In Otterly's test, Microsoft Copilot, which grounds on Bing, kept citing the sites after Google removed them, and Perplexity and Claude cited them too. Google's own AI Overviews and AI Mode almost never did, because Google requires a page to be indexed and snippet-eligible to appear as a supporting link. ChatGPT effectively never cited them.

How do I know if I was hit by an algorithmic removal? Look at impressions, not messages. An algorithmic removal arrives with no manual action in Search Console. In Otterly's test, daily Google impressions on one site fell from 1,629 to 15 in a single day. Check whether pages still appear in the index through Search Console's URL inspection, and compare against Bing Webmaster Tools, since the two indexes can diverge completely.

Do AI engines prefer citing content written by AI? Nothing in the evidence shows that publishers can win citations by having a model write their pages. One study found engines favored text that is more predictable to the underlying model, but its causal test rewrote whole pages at once, so the effect cannot be pinned to writing style. Google says you do not need to write in a special way for its AI features.