Google does not penalize content for being AI-generated. Its ranking systems reward helpful, original pages however they are made, and its spam policies target purpose and scale: using generative AI, or any other method, to produce many low-value pages mainly to rank is scaled content abuse. Use AI to assist, and have a person answer for every page.
This guide covers what Google has published on AI content since February 2023: the spam policies behind it, the who, how and why questions, what the Search Quality Rater Guidelines tell raters, and the metadata Google reads for AI images. It ends with a workflow that keeps people responsible. How AI Overviews choose sources, and how to measure AI search traffic, are separate subjects; this post is about creating the content. What to publish once answer engines absorb generic pages is covered in the next big challenge for content marketers.
Does Google penalize AI-generated content?
No, not for how it was made. On February 8, 2023, Google Search Central published its guidance on AI-generated content. Its ranking systems aim to reward original, high-quality content that shows E-E-A-T (experience, expertise, authoritativeness and trustworthiness), and Google says it focuses on the quality of content rather than how it is produced. The same post calls appropriate use of AI or automation compatible with its guidelines, and points out that automation has produced useful content for years: sports scores, weather forecasts and transcripts.
The line is purpose. Using automation, AI included, to generate content with the primary purpose of manipulating search rankings violates Google's spam policies, and Google says its spam systems, including SpamBrain, act on spam however it is produced. The post's FAQ also removes the opposite hope: using AI gives content no special gains. A page can do well if it is useful, original and trustworthy, and may not if it is not.
"Penalty" covers two different mechanisms, and it helps to keep them apart:
- Ranking systems. Google's core ranking systems assess every page. An AI-written page that adds nothing simply ranks poorly, like a thin page written by a person. No notice arrives; you see it in traffic. The former helpful content system, launched in 2022, became part of the core ranking systems in March 2024.
- Spam enforcement. Automated systems and, when needed, human reviewers act on spam policy violations. Sites that violate them may rank lower or drop out of results, and a human review can end in a manual action reported in Search Console.
What Google's spam policies say about AI content
Three spam policies announced on March 5, 2024 cover the ways AI content most often goes wrong. They arrived with the March 2024 core update. Google said that once these changes finished rolling out on April 19, 2024, search results showed 45% less low-quality, unoriginal content, against the 40% it had expected.
Scaled content abuse
Scaled content abuse is generating many pages mainly to manipulate search rankings rather than to help users: typically large amounts of unoriginal content that gives users little or no value, however it was created. It builds on Google's older policy against automatically generated content, and the announcement explains the change. Scaled production has become harder to attribute, so the policy now applies whether the pages came from automation, from people or from a mix of both.
The policy's own examples read like a list of AI content shortcuts:
- using generative AI or similar tools to generate many pages without adding value for users;
- scraping feeds, search results or other content and running it through automated transformations such as synonymizing or translating;
- stitching together content from several pages without adding value;
- spreading the output across several sites to hide how much of it there is;
- pages that make little sense to a reader but contain search keywords.
Google's May 2026 guide to optimizing for generative AI features adds a current example: writing separate content for every variation of a query, including the fan-out queries AI search generates, primarily to manipulate rankings or AI responses violates the same policy. The spam policies page, as last updated on August 28, 2026, defines spam to include attempts to manipulate generative AI responses in Google Search.
Expired domain abuse
Expired domain abuse is buying an expired domain and repurposing it mainly to manipulate rankings with content of little or no value to users, such as affiliate content on a domain a government agency once used. Google's March 2024 post notes that nobody does this by accident, and that using an old domain for a new, original site built for people is fine. Filling a bought domain with low-value AI articles in the hope that the previous owner's reputation carries them is exactly the pattern the policy describes.
Site reputation abuse
Site reputation abuse, which Google's documentation now calls the site reputation policy, applies when third-party content is published on a host site mainly to benefit from the host's established ranking signals. Enforcement began on May 5, 2024. On November 19, 2024, Google clarified that first-party involvement or oversight does not change the third-party nature of such content. Since August 30, 2026, a manual action under this policy affects only results shown outside the European Economic Area; inside the EEA, the affected section may instead be ranked separately from the rest of the site.
Third-party content is not a violation by itself. Freelance articles, syndicated news, forums and other user-generated content, and advertorials written for the publication's own readers are all named as acceptable. When Google reviews a case, it looks at how much control the host exercises: presentation consistent with the site, quality in line with the main domain, clear authorship and editorial responsibility, and whether the same text appears on many other sites. A white-label vendor's AI articles dropped into a subfolder to borrow your domain's rankings fail most of those tests.
Which uses of AI are fine, and which are spam
Google's documents do not rule on tools; they rule on value and purpose. Mapped onto common ways teams use AI, they come out like this:
| How AI is used | Where it stands | What decides it |
|---|---|---|
| Research, outlines and structure; a person writes the page | Within the guidance | Google names research and structure as useful uses of generative AI |
| An AI draft rewritten by an expert who adds data and checks every fact | Within the guidance if the page helps | Quality and originality of the result, not the tool |
| Pages generated from your own structured data, such as results or schedules | Depends on the value of each page | Automated scores, forecasts and transcripts are Google's own good examples |
| Hundreds of pages for keyword or fan-out query variations | Scaled content abuse | Many pages, little value, made mainly to rank |
| Rewriting or "spinning" other sites' articles | Scaled content abuse, scraping | Automated transformations such as synonymizing add no value |
| Machine-translating freely available content into thousands of pages unchecked | Scaled content abuse | Raters are told this shows no human effort |
| AI articles on an expired domain bought for its history | Expired domain abuse | Borrowed reputation, little value to users |
| A partner's AI content in a subfolder of your established site | Site reputation policy | Third-party content placed there to use your ranking signals |
The first row comes from Google's page on using generative AI content, which calls generative AI particularly useful for research and for adding structure to original content, then warns that generating many pages without adding value may violate the scaled content abuse policy.
What "helpful, people-first" means for AI-assisted pages
Google's guide to helpful, reliable, people-first content is the self-assessment it points every creator to, AI or not. Several of its questions land exactly where unedited AI text is weakest:
- Does the page provide original information, reporting, research or analysis, and insight beyond the obvious?
- If it draws on other sources, does it add substantial value and originality instead of simply rewriting them?
- Is it written or reviewed by someone who demonstrably knows the topic, and is it free of easily verified factual errors?
- Is it mass-produced, or spread across many sites, so that individual pages get little attention or care?
The same page lists warning signs of search engine-first content, and two of them describe the careless AI workflow directly: extensive automation producing content on many topics, and mainly summarizing what others say without adding much value. It also answers anyone who prompts a model for a set length: Google has no preferred word count.
Google's May 2026 guide calls the weak version commodity content: common knowledge that anyone could have written and a generative model could easily reproduce. Its advice is to write from what you know about the subject, as a first-hand review does, instead of recycling what is already online. That is the part a model cannot supply: the test you ran, the figures from your own systems, the decision you made and why. Semantic SEO covers how to plan a topic so that each page covers its subject in depth instead of repeating its neighbours.
Stakes rise with the topic. On subjects that could significantly affect people's health, financial stability or safety, which Google calls "Your Money or Your Life" (YMYL) topics, its systems give even more weight to content with strong E-E-A-T, and trust counts most. An unreviewed AI answer about medication or taxes is the riskiest page a site can publish. How E-E-A-T and quality raters actually work is covered in how Google judges relevance and authority.
Who, how and why: disclosing AI use
Alongside its AI guidance in February 2023, Google updated the helpful content guide's who, how and why questions to cover how content is produced:
- Who created it. Make authorship self-evident: a byline where readers would expect one, leading to information about the author and the subjects they cover. Google's FAQ says an AI byline is probably not the best way to tell readers that AI was part of the process.
- How it was created. When automation substantially generated the content, readers should be able to see that, learn how it was used and understand why it was useful. Google suggests AI disclosures on content where someone might ask "How was this created?", added where readers would reasonably expect them.
- Why it was created. Google calls this perhaps the most important question. Content made primarily to help people who come to your site directly is aligned with what its systems reward; content made mainly to attract search visits is not, and producing it with AI for that purpose is a spam policy violation.
A disclosure does not need to be dramatic. One line near the byline or at the end usually does it, for example: Drafted with an AI writing assistant from our engineers' notes; the author checked every figure against the sources listed below. It works because it is specific: it tells the reader what the tool did and who is accountable for the result.
Note
Disclosure can also be a legal duty. Under Article 50(4) of the EU AI Act, deployers who publish AI-generated or manipulated text to inform the public on matters of public interest must disclose it, unless the text has undergone human review or editorial control and a person or company holds editorial responsibility for it. The European Commission lists the Act's transparency rules as applying from August 2026. It is law, not a ranking signal; ask counsel how it applies to your publication.
What the Search Quality Rater Guidelines say about AI content
The Search Quality Rater Guidelines are the handbook for the people who rate search results so Google can evaluate its ranking changes. Their ratings do not directly change how any page ranks, but they show how Google defines quality. The version Google publishes as of September 2026 is dated September 11, 2025. Section 2.1 defines generative AI as a type of machine learning model that creates new content such as text, images, music and code, and calls it a helpful tool for content creation that, like any tool, can be misused. Google's generative AI guidance points to two later sections:
- Section 4.6.5, scaled content abuse. Pages made of content created at scale with no original content or added value get the Lowest rating however they were made. Raters are told to give it even when they cannot tell whether generative AI was used, if several pages on the site make them strongly suspect scale abuse.
- Section 4.6.6, little effort, originality or added value. The Lowest rating applies when all or almost all of the main content is copied, paraphrased, embedded, auto or AI generated, or reposted with little to no effort, originality or added value, even when the page credits its source.
The guidelines are just as careful in the other direction. Using generative AI tools does not, by itself, determine a page's effort or quality: the same tools can produce high-quality work and Lowest-quality work, and paraphrasing can add value, as when an expert restates a government policy in plain language. Section 3.2 defines effort as the extent to which a human being actively worked to create satisfying content, and gives a counter-example worth remembering: thousands of pages created automatically by running freely available content through translation software, with no oversight or curation, show no effort.
Raters are also taught how to recognize paraphrased pages: they hold only commonly known facts, overlap heavily with reference sites, or contain leftover chatbot boilerplate, such as a sentence announcing that the text comes from an AI language model. An editing pass should remove all three.
How to label AI images and AI product data
For text, Google asks for context readers can see. For images, it also reads metadata. Its image metadata documentation lists the IPTC Digital Source Type values it supports:
| IPTC value | What it means |
|---|---|
trainedAlgorithmicMedia | Created by a model derived from sampled content; IPTC names it "Created using Generative AI" |
compositeWithTrainedAlgorithmicMedia | Generative AI output combined with other media, as in inpainting or outpainting |
compositeSynthetic | A composite that includes at least one synthetic element |
algorithmicMedia | Created purely by an algorithm with no training data, such as a mathematical formula |
IPTC photo metadata is embedded in the image file itself, so it stays with the image as it moves from page to page. If an image also carries C2PA provenance data from an app, device or service that has adopted C2PA 2.1 or later, signed with a certificate from a certification authority on the C2PA Trust List, Google may show how the image was made, including edits with AI tools, in its About this image feature.
Google's generative AI guidance asks sites that generate content automatically to consider adding image metadata, and for shopping it is a requirement. Google Merchant Center requires every product image created with generative AI to carry the IPTC DigitalSourceType TrainedAlgorithmicMedia tag, and AI-generated titles and descriptions to be sent in the structured_title and structured_description attributes, with digital_source_type set to trained_algorithmic_media.
To set or check the field yourself, ExifTool writes it to the image's XMP data as XMP-iptcExt:DigitalSourceType. The IPTC user guide recommends the full URI of the code as the value:
# Mark an image as created with generative AI
exiftool -XMP-iptcExt:DigitalSourceType="http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia" illustration.jpg
# Check the file your site actually serves, after the build or CDN has processed it
exiftool -XMP-iptcExt:DigitalSourceType illustration.jpg
Removing image metadata reduces file size, as Google's documentation notes, so check that your optimization step keeps this field. Schema.org also defines a digitalSourceType property for any creative work, but Google's image documentation lists the digital source type only among the embedded IPTC fields it reads, so do not rely on the markup alone.
A responsible AI content workflow
Put together, Google's documents describe a workflow in which AI does what it is good at and people own everything a reader has to trust:
- Choose the page from a real question. Start with what your audience asks: support tickets, sales calls and the queries in Search Console. Untapped SEO opportunities shows how to find questions you already appear for without answering them well. Keep one intent per page, written for people who would read it if they arrived directly.
- Use AI for research leads and structure. Treat everything a model says as a lead. Open the primary source, find the claim in its text and keep a list of the sources each page relies on.
- Write, or rewrite, with first-hand knowledge. If a model drafts, an expert rewrites and adds what only your team has: measurements, screenshots, the configuration that worked, the trade-off you chose and why. That is the non-commodity part, and the effort raters look for.
- Fact-check every claim. Check every number, date, version and name against the page it came from, date anything that changes over time, and link the source. A model states a wrong version number as fluently as a right one.
- Check the metadata as well. Google's generative AI guidance asks for accuracy, quality and relevance in titles, meta descriptions, structured data and image alt text, especially when they are generated automatically. Validate structured data before release.
- Sign and disclose. Use a real author's byline with a profile, add a short AI note where readers would want to know how the page was made, and embed IPTC metadata in generated images.
- Publish at the speed you can review. No batch goes live that nobody has read. Later, improve pages that do not help readers, or remove them if they cannot be saved; do not just change their dates, which Google lists as a warning sign.
The review step is the one that cannot be automated away. A model can produce a fluent draft in seconds; only a person who knows the subject can say whether it is true, what is missing and what the reader should do next. How drafting tools produce that text, and where they fail, is covered in what Draft AI is.
How to tell whether AI content is hurting your site
Start in Search Console. The Manual actions report lists any action a human reviewer has taken. The two most relevant to AI content are Thin content with little or no added value (scraped content, thin affiliate pages, doorways) and Major spam problems, which Google describes as aggressive spam techniques such as scaled content abuse. Pages found inconsistent with the site reputation policy get an action of their own.
No manual action does not mean no problem. Core updates and automated spam systems change rankings without sending a message, so compare the dates of any traffic drop with the Google Search Status Dashboard. Since the March 2024 changes, spam updates started on June 20, 2024, December 19, 2024, August 26, 2025, March 24, 2026, June 24, 2026 and August 18, 2026, and another began on September 24, 2026.
An AI detector is not the right test. OpenAI withdrew its own AI text classifier on July 20, 2023 because of its low rate of accuracy, and Google's guidance judges pages by their quality, not by how they were produced.
If a drop lines up with an update, Google's helpful content guide suggests auditing the pages that lost the most and the searches they lost, against its self-assessment questions. Fix the pattern, not one page:
- merge or remove AI pages that repeat each other or target near-identical queries;
- rewrite the pages worth keeping with first-hand material and checked sources;
- exclude from Search any scaled content you decide to keep for other reasons, as the spam policy advises;
- move third-party sections that exist to use your domain's signals to a separate domain, without redirecting the old URLs to it, as the manual action help recommends;
- for a manual action, fix every affected page, then request a review in the report, with examples of the bad content you removed and the good content you added. Google says most reviews take several days or weeks.
If you want help planning content like this, our SEO service builds topic clusters from search demand and intent, with briefs, sources and outlines for writers who know the subject. If you want a model inside your publishing workflow, AI & automation uses a language model only where a step needs judgment, such as reading or drafting, with a rule or a named person approving anything before it changes your systems.
