Google's Guidance on AI-Generated Content, Digested

Google Search Central publishes a short page of guidance on generative AI content for site owners. Its central line is plain: "Appropriate use of AI or automation is not against our guidelines."

That sentence answers the question most marketing teams ask first. Google does not check whether a model wrote a page and demote it for that reason. The guidance sits alongside two other documents that define the actual line:

• The spam policies, which include scaled content abuse.
• The helpful content guidance, which describes people-first content.

Read together, they say the same thing from different angles. Google judges the page and the pattern of publishing, not the tool. This page digests that position and tests it against the one public experiment in the evidence base that pushed AI volume to the limit.

The generative AI guidance tells you what is allowed. The spam policy tells you where allowed use stops.

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."

So there are two tests, and neither one mentions the author:

1. Purpose. Was the page made mainly to rank, or to help someone?
2. Value. Does each page add something a reader could not get elsewhere?

A human writing 1,000 thin pages fails the same test as a model doing it. A model drafting one useful page that a person checks passes. Our digest of Google's spam policies on scaled content abuse covers the full list of examples.

OtterlyAI tested the volume shortcut in 2026 and published the results. It loaded two new domains with roughly 1,000 AI-generated posts each, about 2,000 in total. One site showed an author box presenting a fictional wildlife biologist. The other had no author.

Otterly reported these outcomes:

• Google deindexed both sites, and Otterly reports zero manual actions in Search Console on either domain.
• The site with the fictional expert lasted about 14 days in Google's index. The site with no author lasted about 90.
• capybara-world.com fell from 1,629 to 15 daily Google impressions on April 9, according to Otterly's write-up.
• capybaraverse.com fell from 859 to 85 daily impressions on June 26, per the same write-up.
• Bing kept both sites. Of 3,934 AI citations Otterly tracked for capybara-world.com, 2,718 (69%) came from Microsoft Copilot.

The Google result is the policy working as written. A thousand thin pages on a domain with no history is the pattern the scaled content abuse policy describes. The removal arrived silently. The full experiment is covered in does Google penalize AI content.

The guidance is an assertion by Google, not a measurement. It is the best available statement of Google's intent, but nobody outside Google can verify how its systems apply it.

The Otterly test has clear limits:

• One site per arm. The author box was not the only difference between the two sites, so its effect cannot be isolated.
• No human-written control. Without 1,000 human posts published overnight, the test cannot say whether Google reacted to AI text or to sudden volume. The policy targets volume.
• A low-stakes topic. Capybara questions face little competition, so the result says little about commercial categories.
• A vendor ran it. Otterly sells AI visibility tracking, as does 99Visibility. That is why replication matters.

Neither source shows that AI drafting helps a page rank or get cited. Whether AI engines favor AI-written text is a separate question, covered in do AI engines prefer AI-written content. The author box result also does not prove author boxes hurt. It only shows one did not protect thin content. Google states that E-E-A-T "isn't a specific ranking factor," so there is no author score to game.

The practical line is simple: use AI to make each page better, not to make more pages.

1. Start from a purpose. Before drafting, write down the reader and the question the page answers. If you cannot, do not publish the page.
2. Have a person check every fact. Prices, numbers, product names and claims need human review. Models get specific facts wrong, and a wrong price on your own site is a fact AI engines can repeat.
3. State specific facts plainly. A clear price, number or one-line definition on a fetchable page gives an engine a sentence it can quote.
4. Do not mass-publish. In Otterly's test, Copilot citations did not make up for losing Google, and nothing in Search Console warned either site.
5. Skip the AI-style rewrite. Google's generative AI optimization guide says "You don't need to write in a specific way just for generative AI search."
6. Watch your index status. Google's AI features documentation says 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. Deindexing removes you from both.

As of October 2026, none of this guarantees rankings or citations. It keeps you on the right side of a policy that can remove you without notice.

Do I need to disclose that content was written with AI? The evidence base does not include a Google rule requiring disclosure for ranking purposes. Google's stated test is purpose and value, not authorship. Disclosure can still be a sensible trust choice for readers, and some industries have their own rules. Decide based on your audience and any legal requirements, not on a belief that disclosure changes how Google ranks the page.

Will an expert author box protect AI-generated pages? Nothing in the evidence suggests it does. In Otterly's test, the site with a fictional expert author box was deindexed after about 14 days, and the site with no author after about 90. One site per arm cannot prove author boxes hurt. It does show a fake byline is not a shield, and inventing an expert is a credibility risk in its own right.

Does Google send a warning before removing scaled AI content? Not necessarily. Otterly reports zero manual actions in Search Console on either test domain, so the removals were algorithmic. Google does not say which system removed them. Do not rely on a Search Console notice as an early warning. Watch impressions and index coverage directly, and check Bing Webmaster Tools separately, since Bing kept both sites.

If Google deindexes a site, can it still appear in AI answers? On Google's own AI surfaces, generally not, because AI Overviews and AI Mode require indexed, snippet-eligible pages. Other engines draw on other indexes. In Otterly's test, Microsoft Copilot, which grounds on Bing, kept citing the sites after Google removed them. That is not a strategy: Google remains the largest channel, and losing it costs more than Copilot returns.

How many AI-assisted pages is too many? Google publishes no page count. Its policy turns on purpose and value: many pages generated mainly to manipulate rankings, without helping users, is the violation. Publishing at a pace your team can fact-check and justify page by page keeps you on the right side. If a page exists only to target a keyword variant, it is a candidate for removal, not publication.