Google's docs now spell out the effort test for AI pages

Google updated its guidance on using generative AI content on October 1, and the page now routes readers into the Search Quality Raters guidelines. The cited sections are 4.6.5 on scaled content abuse and 4.6.6 on main content created with little to no effort, little to no originality, and little to no added value. The Search Central changelog gives the reason for the edit as getting the documentation in sync with the presentations Google uses at its developer events.

None of that is a new rule. Both rater sections already existed, and by Google’s own account the docs were catching up to what it already says on stage. The difference is the address: the effort standard now sits on a dated developers.google.com page, and that is the version that settles an argument with a client or an in-house content team.

The test in 4.6.6 applies to main content, the raters’ term for the part of a page that does the job the page exists for, as opposed to the navigation, the ads, and the related-links rail. Which blocks qualify on your own templates is worth settling before someone else decides. On a programmatic location page, the main content is usually the two or three generated paragraphs under the H1, not the map, the opening hours, or the internal links below them.

Whether AI wrote the page is not the question the doc asks. Its spam line turns on scale and value together: using generative AI tools to generate many pages without adding value for users may violate the scaled content abuse policy. A single AI-assisted page carrying a real finding is not what 4.6.6 describes. Several hundred near-identical pages whose main content is model output with the city name swapped is. Thin main content costs visibility well before any enforcement arrives, since AI Overviews skip thin and stale pages sitting at the same organic rank as the ones they cite.

On accuracy the doc is blunter than most of Google’s documentation. Its reasoning is that generative models do not retrieve facts but predict a likely sequence of words from their training data, so outputs may contain inaccuracies. The instruction attached to it is to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing. “All” is the operative word, and a sample does not meet it.

On disclosure the doc stops at a suggestion, that publishers consider adding information on how their content was created in a way that makes sense for their audience. The only hard technical requirement on the page is for e-commerce: AI-generated images must carry IPTC DigitalSourceType metadata set to TrainedAlgorithmicMedia, and AI-generated product titles and descriptions must be specified separately and labeled as AI-generated.

What to do

• Mark the main content block on each template, then judge that block on its own. A rich page with a thin generated middle fails 4.6.6 no matter how much else sits around it.
• Decide per template rather than per page. The spam policy fires on many pages without added value, so the count of near-duplicates matters more than any single URL’s word count.
• Make the factcheck a named step with an owner, since the doc asks for all AI-generated content to be reviewed before publishing, not after.
• If you sell products, set DigitalSourceType on AI-generated imagery at export time and label AI-written titles and descriptions separately in the product data.

Watch out for

Attribution gets hard after the fact. Spam updates may roll out over two weeks rather than days, so a site that loses traffic during one has a hard time proving which pages caused it. Cleaning up thin main content now is cheaper than diagnosing it mid-rollout.