Is AI Content Bad for SEO? What Google Actually Penalizes

Google does not demote a page because a model wrote it. Its guidance on generative AI content says: "Appropriate use of AI or automation is not against our guidelines." That line has been Google's position for a while, and it still holds as of October 2026.

What Google does act on is a pattern of publishing. Its spam policies describe scaled content abuse, and AI is one way to commit it, not the definition of it. We break the policy text down in Google's guidance on AI-generated content, digested.

So the useful question is not "did AI write this?" It is "does each page give a reader something they could not get from the pages already ranking?" If the honest answer is no for hundreds of pages at once, the risk is real whoever wrote them.

Google defines scaled content abuse as being "when many pages are generated for the primary purpose of manipulating search rankings and not helping users." Its first listed example is "Using generative AI tools or other similar tools to generate many pages without adding value for users." The policy applies "no matter how it's created."

Three parts of that definition matter:

• Many pages. One AI-assisted article is not the pattern. Volume is.
• Primary purpose. Pages built to capture search traffic rather than answer a reader are the target.
• Without adding value. Thin rewrites of what already exists fail this test, whether a person or a model produced them.

Enforcement can be quiet. A removal can be algorithmic, which means no message in Search Console and no named system to appeal to. The difference matters for diagnosis, and we cover it in manual action vs algorithmic deindexing.

There is a second consequence people miss. 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." Lose the index and you lose Google's AI surfaces with it.

OtterlyAI tested the volume shortcut in 2026 and published the results. It launched two new domains about capybaras, each loaded with roughly 1,000 AI-generated posts in one click. One site, capybara-world.com, showed an author box presenting a fictional wildlife biologist with more than 15 years of field experience. The other, capybaraverse.com, had no author. Otterly tracked Search Console, Bing Webmaster Tools, AI citations across seven engines and crawler visits for about four months. The write-up was last updated on 2026-08-05.

Google removed both, the fake expert first

The site with the fictional expert went first. If author signals protected scaled AI content, the order should have run the other way. That fits Google's statement that E-E-A-T "isn't a specific ranking factor."

Bing kept both, and Copilot kept citing them

Bing never deindexed either site. Combined Bing impressions reached 119,935 in July, per Otterly. Otterly tracked 100 AI citations for capybara-world.com in April and 1,666 in the first 27 days of July, while the site was gone from Google. Of 3,934 tracked citations for that site, 2,718 (69%) came from Microsoft Copilot. ChatGPT, Gemini, AI Overviews and AI Mode almost never cited the sites. ChatGPT's crawlers kept visiting, and ChatGPT citations stayed effectively at zero even while the sites ranked in Bing.

The reading is simple: an engine cannot cite what its index no longer holds. And being indexed somewhere is necessary but not enough.

Otterly ran the test in public, which is more than most AI content advice does. It still carries less weight than the headlines put on it.

• One site per arm. The author box was not the only difference. The domains differed too, so the author box result is one data point.
• No human-written control. Nobody published 1,000 human-written posts overnight for comparison. The test cannot separate a reaction to AI text from a reaction to a thousand pages appearing on a domain with no history.
• A low-stakes topic. Capybara questions are informational and uncompetitive. Copilot citing them says little about a crowded commercial category.
• Citation counts depend on prompts. Microsoft's own AI Performance report credited capybara-world.com with 42,817 citations, more than ten times the 3,934 Otterly's tracker logged. The time windows may differ, but it shows any prompt set is a sample.
• A vendor ran it. OtterlyAI sells AI visibility tracking. That does not make the data wrong. It is why replication matters.

What survives these limits: the outcome matches Google's written policy, and the index-to-citation link matches what Google and Microsoft document.

Some things sit with you. Others sit with the engines, and no tactic moves them.

In your control

• How many pages you publish, and how fast.
• Whether each page adds a fact, a view or an answer the reader can't get elsewhere.
• Whether a person checks facts, prices and claims before a page ships.
• Whether search and AI crawlers can reach your pages at all.
• Whether you claim credentials you don't have. The fake expert in Otterly's test did not save the site.

Not in your control

• Which algorithmic system removes a page, and when. Google gave neither site a warning.
• Whether an engine cites you. Bing indexing produced Copilot citations and no ChatGPT citations in the same test.
• Whether engines treat AI text differently from human text in their answers. See whether AI engines prefer AI-written content for what the evidence does and doesn't say.

Google also says you don't need to write in a specific way for generative AI search. Style tricks are not the lever. Value and access are.

If you use AI to draft content, work in this order.

1. Decide what each page is for. Write down the question it answers and what it adds. If you can't, don't publish it. The contrast is laid out in scaled AI content vs people-first content.
2. Fact-check every specific claim. Numbers, prices, dates and product names are where drafts go wrong. A wrong fact at scale is repeated hundreds of times.
3. Publish at a pace you can review. A thousand pages on a new domain is the exact pattern Google's policy describes. Use a pre-publish review for scaled AI content before anything ships in bulk.
4. Watch your index status per engine. Check Google Search Console and Bing Webmaster Tools separately. One healthy index tells you nothing about the other.
5. Check that crawlers can reach you. A robots.txt that allows a crawler while your firewall returns 403 removes you as quietly as a deindex does.
6. Measure each AI engine on its own. One engine cited Otterly's sites thousands of times while others barely did. A blended number would hide that. The free checker on this site gives a snapshot of whether engines name you for three buyer questions.

None of these steps guarantees rankings or citations. They remove the reasons you would be excluded.

These are the errors that turn a useful tool into a liability.

• Treating silence as safety. Otterly reports zero manual actions on both sites, and both were removed. No message in Search Console does not mean no problem.
• Inventing authors and credentials. A fictional expert is a fabricated claim to readers. In this test it was no protection either.
• Counting Copilot citations as a replacement for Google. Google is still the largest channel. Citations on one engine did not offset losing Google's index and its AI surfaces.
• Assuming "ChatGPT uses Bing" is a strategy. Ranking in Bing produced effectively zero ChatGPT citations in Otterly's test.
• Reaching for manipulation. Hidden text, injected instructions and cloaking pages for AI crawlers are illegitimate and carry spam risk. See whether GEO tactics can violate Google's spam policies.
• Swinging to "never use AI." The evidence doesn't support that either. For the case on the other side, read is AI-generated content good for SEO.