In early October 2026, Google revised two pages of the documentation it publishes for site owners. One covers AI-generated content, the other sets out what Google means by helpful content. The page on AI-generated content now asks for every piece of it to be fact-checked by a person before it goes live. For a small business or agency that uses AI to draft articles, product listings or meta tags, the two pages now describe fairly precisely what Google expects from published content.
The changes landed in the middle of an unusually long spam update, which prompted plenty of speculation across the industry. This article goes through what the new guidance actually says and sets it against the available research on AI-written content. It then outlines a publishing workflow that meets it, with an eye on visibility in AI answers as well as in search.
Key takeaways
- Google now asks for AI-generated content to be fact-checked by hand before publication, metadata included.
- Its helpful content page also defines main content and spells out the four criteria its raters use to judge it.
- Using AI is not penalized in itself: the spam policies target pages mass-produced with nothing added, and Google warns against fake author profiles.
- The same principles apply to AI Overviews and AI Mode, since Google sets no separate optimization for them.
- A documented review process, a real author and something of your own in every piece cover most of what Google asks.
What Google changed in early October 2026
Two Search Central pages were expanded on the same day, October 1, 2026. They mostly carry over criteria Google's quality raters were already applying, rather than introducing new rules.
For the page on AI-generated content, Google's documentation changelog says it added material from its search quality rater guidelines. The stated aim is to bring the documentation in line with the talks Google gives at its developer events. The page on helpful, reliable, people-first content was expanded at the same time, with no changelog entry. The October 5 date it displays reflects a minor wording fix.
Barry Schwartz, founder of Search Engine Roundtable, picked up the changes as soon as they appeared. In a piece published on October 1, 2026, he notes that Google rarely uses the word "critical" in its documentation, which gives the fact-checking request extra weight. He also floats a possible link with the September spam update, while presenting it as speculation.
| Page revised | Main addition | What it targets |
|---|---|---|
| AI-generated content | Manual fact-checking before publication | Factual errors produced by models |
| AI-generated content | The check now extends to title tags, meta descriptions, structured data and alt text, which the page already listed | Automatically generated metadata |
| Helpful content | Definition of main content | What is actually assessed on a page |
| Helpful content | Four criteria: effort, originality, skill, accuracy | How raters judge that content |
What does the fact-checking requirement cover?
Google wants a person to check every piece of generated content for accuracy and trustworthiness before it goes live. The instruction also covers the metadata that can show up in search results, not just the body copy.
The page grounds this in how the models work. According to Google, a generative model does not retrieve facts: it predicts a likely sequence of words from its training data. It will therefore sometimes produce inaccuracies, which are known as hallucinations. The key sentence in the new version says it is critical to manually fact-check and review all AI-generated content before publishing.
Google's examples of the metadata in question are title tags, meta descriptions, structured data and image alt text. These are exactly the elements many tools generate in bulk, with nobody looking at them afterwards. A meta description that gives a product a feature it does not have is still a public mistake, even if the article itself is accurate.
The page also restates two requirements specific to ecommerce. Google Merchant Center, the tool merchants use to send their products to Google Shopping, requires AI-generated titles and descriptions to be submitted separately and flagged as AI-generated. AI-generated product images must carry metadata in the IPTC format, the photo metadata standard that grew out of the news industry. That metadata declares their origin with the TrainedAlgorithmicMedia value of the DigitalSourceType field. Where relevant, Google also suggests adding information on how the content was produced.
The four criteria Google now spells out for main content
Google defines main content as any part of a page that directly helps it achieve its purpose. It judges that content on human effort, originality, talent or skill, and accuracy.
The definition goes well beyond the body text. It includes media and interactive features, such as a calculator or an on-site search. It also covers user contributions that serve the page's purpose, content inside tabs or expandable sections, and titles and headings. A well-built price comparison tool can therefore make up most of a page's main content, while a long generic text underneath may add little.
Effort is the criterion that bears most directly on AI-assisted content. Google describes it as the amount of human work that went into the content, or into the systems that produce it. It says that pages generated automatically from data feeds, such as a product catalog, represent little to no effort. The same goes for large volumes of AI text produced without human oversight or curation. The page also points out that crediting your sources is no substitute for original effort.
Originality means information or perspectives that are not already available on other sites. Talent or skill refers to the ability or expertise needed to give visitors a satisfying experience, which includes clear writing or a tool that works, without demanding specialist expertise on every topic. Accuracy, finally, means factual correctness on an informational page. On topics that affect people's health, finances or safety, Google expects content to be highly accurate and consistent with established expert consensus.
The same page restates the three questions Google asks of any content: who made it, how, and why. On the first, it strongly encourages bylines where readers would expect them, linked to background on the author. It also warns against fabricated profiles, such as AI-generated headshots, invented names or false credentials, and treats that kind of deception as a sign of a low-quality page.
On the third question, which Google calls perhaps the most important, it repeats that content should be made first and foremost for the people who visit the site, not to attract search traffic. We cover these trust signals in our definition of E-E-A-T.
Does Google penalize content written with AI?
Nothing in the guidance or the available research points to a penalty for using AI as such. The targets are mass production with no added value and the intent to manipulate rankings.
Google's spam policies are unambiguous on this. The page that sets them out, last updated on August 28, 2026, defines scaled content abuse as generating many pages primarily to manipulate search rankings. The policy targets unoriginal content that offers users little or no value, "no matter how it's created." Using generative AI tools to produce many pages without adding value for users is the first example listed.
The guidelines for Google's quality raters follow the same logic. According to Search Engine Journal on January 26, 2025, the version released that month gives the lowest rating to pages created with generative AI that likely contain no original content and offer users no value. Raters do not change rankings directly, but their scores are used to assess the quality of the results Google's algorithms produce.
The available data suggests AI is already widespread among well-ranked pages. Ahrefs' study published on July 7, 2025 analyzed around 600,000 pages ranking in the top 20 results for randomly selected keywords. According to its detector, 86.5% of them contained at least some AI-generated text. The correlation between that share and ranking position was 0.011, which the authors describe as effectively zero. The study relies on an automated detector, and its authors acknowledge that such tools produce false positives.
The trend is not confined to pages that rank well. The Pew Research Center published an analysis on August 20, 2026 covering around 490,000 English-language pages from Common Crawl, a public archive of the web. It found that more than a third of pages published since ChatGPT's release show signs of AI authorship. The researchers used the Open Pangram detector and point out that such tools can be wrong in both directions. Against that backdrop, a rule that punished the tool rather than the outcome would be hard for Google itself to enforce.
A spam update in the background
The new guidance was published while the September 2026 spam update was rolling out. Google has not said what that update targeted, and any link between the two remains speculation.
The update started on September 24, 2026, and finished on October 8. According to Search Engine Journal on October 8, 2026, the rollout took 13 days and 16 hours, longer than any of the year's three previous spam updates. Google described it as global and covering all languages, without saying which practices it was aimed at.
Early data is patchy. The same article reports an analysis by Diib, published on September 29, covering 32,569 sites connected to Google Analytics. In that sample, the median share of visits coming from unpaid Google search fell from 34% to 31% between the previous week and the first three days of the rollout. Diib itself calls this an early read and says its data cannot show which sites Google treated as spam.
The trade press also picked up a Google research paper on a system called SAFE, designed to detect abuse at scale. Search Engine Journal covered it on September 25, 2026, saying the system may be part of the update but offering no evidence that it is. The paper is very short and mentions early deployment results without reporting any figures. It appears to focus mainly on video. Building a strategy around the idea of a new AI-content detector being applied to websites would therefore be unwise.
The same rules apply to AI Overviews and AI Mode
Google has no separate rulebook for its AI features, so a page that classic search judges unhelpful is unlikely to be cited in an AI Overview. For other engines this is still an assumption, but it points the same way.
Google's page on AI features, last updated on December 10, 2025, states that there are no additional requirements or special optimizations needed to appear in AI Overviews or AI Mode. It recommends applying the same best practices as for search in general, and lists helpful, reliable, people-first content among them. The October criteria should therefore also weigh on which pages get selected to feed those answers. We explain how these formats work in our definition of AI Overviews.
ChatGPT, Perplexity and Claude publish no comparable guidance on content quality, and how they select sources is still poorly documented. These engines may, however, rely on existing search indexes to find the pages they cite, which could pass on some of those indexes' quality judgments. Accuracy probably matters even more there, because an AI answer that repeats a false claim damages the engine's own credibility. The research review we covered in our article on the GEO techniques that actually work points the same way. Statistics and quoted figures have a moderate effect there, provided the facts are correct.
The publishing workflow we recommend
A workflow that meets Google's guidance comes down to five steps that a small team can run. Above all, it makes the human work and the original contribution in each piece visible.
The first step is to check every fact before publication by going back to the primary source. A figure produced by a model, or lifted from an article that quotes it without a link, should be traced to the original study page, with its date and scope. If the source cannot be found, the figure comes out. This is the rule we apply to our own articles and to those written by our content agent, whose rules bar it from inventing an expert, a quote or a statistic.
The second step covers metadata, which often slips through review. Read the title tag, meta description, alt text and structured data as carefully as the body copy, and check that they describe the page accurately. A product page whose structured data shows the wrong price or availability is as much of a problem as an error in the text.
The third step is to put a real person's name on every piece, linked to a page about their background. If you currently publish under a fictional author with an AI-generated photo, replace them with a team member who reviews the content and stands behind it. The fourth step is about original contribution. Every article should contain at least one thing that cannot be found elsewhere, such as internal data, a client case, a photo taken on site or a reasoned view from your own expert.
The fifth step is transparency. Where AI played a significant part in producing a piece, a short note can explain how it was used and who reviewed the text. That answers the "how was this created?" question Google recommends anticipating. For product catalogs, the Merchant Center fields and the IPTC metadata on generated images remain mandatory.
| Step | What you check | Google requirement it meets |
|---|---|---|
| Fact-checking | Every figure and claim traces back to a primary source | Manual fact-checking, accuracy |
| Metadata review | Title, meta description, alt text, structured data | Review of metadata shown in results |
| Real author | Byline from an identifiable person, author page | The "who" question, warning against fake profiles |
| Original contribution | Data, example or opinion not found on other sites | Effort, originality |
| Transparency | Note on AI use, Merchant Center fields, IPTC | The "how" question, ecommerce rules |
What we recommend for the weeks ahead
Start by listing the content you have published with AI assistance over the past year, beginning with anything that carries figures, prices or advice on health or money. For each piece, check the factual claims and the metadata, then correct or remove whatever does not hold up. This cleanup can pay off more than producing new articles, since it works on pages that are already indexed.
Next, look at pages produced in series, such as city landing pages, product listing variants or comparison pages spun out of a single template. These are the closest to the scaled content abuse Google describes. A page that differs from its neighbors only by a city name is worth enriching with real local information, or merging with others.
Finally, we suggest measuring the effect of these fixes rather than assuming it. An AI visibility audit run before and after you rework a group of pages shows whether they have gained in clarity, sourcing and author signals. AI citation tracking then shows whether they are starting to be cited in AI answers to the questions you monitor. A steady routine, as described in our article on publishing frequency, works better than a one-off overhaul, because Google and the answer engines keep reassessing pages over time.
This article was drafted with help from our writing agents, then checked against its primary sources and reviewed by Nicolas Meridjen.
Frequently asked questions
Does Google penalize content written with AI?
Not as such. Google's guidance targets how content is produced and what it is for, not the tool used to write it. Pages churned out at scale with nothing in them for the reader, or made mainly to manipulate rankings, fall under its spam policies whatever produced them.
What did Google change in its guidance in October 2026?
On October 1, 2026, Google added an explicit request to its page on AI-generated content: fact-check it manually before you publish. The same day, its helpful content page gained a definition of main content and four criteria for judging it: effort, originality, talent or skill, and accuracy. Google says the first change draws on its quality rater guidelines, and notes that its raters are trained to assess main content on those four criteria.
Should I disclose that an article was written with help from AI?
Google does not require it for articles, but encourages it where readers might reasonably wonder how the content was made. It suggests explaining how automation was used and why it helped. Product listings are different: Merchant Center requires AI-generated titles and descriptions to be submitted separately and labeled as such.
Do these rules also apply to AI Overviews and AI Mode?
Yes. Google says the same SEO best practices apply to its AI features, and lists helpful, reliable, people-first content among them. A page that classic search judges unhelpful has little reason to be picked for an AI Overview.
Which metadata should I review when using AI?
Google names title tags, meta descriptions, structured data and image alt text, since all of them can appear in search results. These elements are often generated automatically and rarely reviewed, yet a mistake there is visible before anyone even clicks.
