How to use AI for SEO: 3 methods we tested (and 1 to avoid)

Short answer: The safest way to use AI for SEO is to give it a job whose result you can check. Three methods fit that rule: a coding agent fixes one finding from an SEOgrinder report, AI drafts titles and meta descriptions that a person checks for length and facts, and AI writes JSON-LD that Google's Rich Results Test validates. The practice to avoid is generating many pages without adding value for users, which Google's spam policy calls scaled content abuse.

Test status (4 November 2026): This article describes a test that has not been run yet. The method is fixed below; we will add the results after the run (planned for 5 November 2026). Until then, nothing here claims a result.

Is AI content against Google's rules?

Not by itself. Google answers the question in its FAQ on AI content: "Appropriate use of AI or automation is not against our guidelines." The same answer names the limit. AI must not be used to generate content primarily to manipulate search rankings, because that is against Google's spam policies. The page on creating helpful content says it again for automation, including AI generation: producing content for the primary purpose of manipulating search rankings is "a violation of our spam policies".

So the line Google draws is the purpose, not the tool. Google's guidance on generative AI content names a use it endorses: "Generative AI can be particularly useful when researching a topic, and to add structure to original content." The same page adds a duty: "It is critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing." According to that page, the review also applies to metadata such as title elements, meta description elements, structured data and alt texts for images, because these can appear in search results. Methods 2 and 3 below are built around exactly that sentence.

AI for SEO at a glance: three methods and one warning

All three methods follow one pattern. The AI does the drafting, and something outside the AI decides whether the result is right: a new scan, a character count and a comparison with the page, or Google's own validator.

Method 1: let a coding agent fix one finding

An SEO report contains mechanical findings, such as a missing canonical tag or a missing title. They have one right answer, and you can check it by fetching the page. That makes them a good job for a coding agent such as Claude Code, Cursor or GitHub Copilot. The full workflow, with safety rules, is in our article on fixing SEO issues with an AI coding agent. Here is the short version.

Enter the address of your public page in SEOgrinder. It is free and needs no account. Press Hand over to AI in the report to open the briefing, a Markdown work order in which every task has a "Done when" condition that can be checked. To hand over a single task, press the same button on its task card. Give the briefing to your agent and tell it to do one task and stop. Read the diff, release the change, then press Check again. SEOgrinder keeps a report for up to 24 hours, and the button forces a fresh run (it works a few times per hour).

If the finding is gone from the new report, the method worked. If it is still there, the agent's own summary told you what it meant to change, not what is live. SEOgrinder reads public pages only, so a change that exists only on your laptop is invisible to it.

One limit belongs here. A fixed finding removes one technical error; SEOgrinder checks most of the technical foundation, not all of it. It does not promise a better position, and SEOgrinder does not measure rankings or whether AI assistants mention your site.

Method 2: let AI draft titles and descriptions, then check them

Drafting is where AI saves time: a model can propose three titles and three descriptions in seconds. The risk is a statement that is not on your page, such as a number, a promise or a feature. Titles and descriptions can show up in search results, so Google's fact-checking rule from above applies to them in full.

A procedure that keeps the check simple:

• Give the AI the source. Paste the page text, name the keyword and forbid anything that is not in the text. A prompt like this does it:

Write three variants of a title and three variants of a meta description for the page text that follows this message.
Use only facts that appear in the text. Add no numbers, prices, awards or promises.
Put the main keyword near the start of each title and describe what the reader gets.

• Count the characters yourself. Do not trust the length the model reports. In a terminal, printf '%s' "AI for SEO: 3 methods in our test, 1 to avoid" | wc -m prints 45.
• Check every statement against the page. If a claim from the draft is not on the page, delete it or add it to the page first.
• Choose or edit one variant. Read it as a visitor would.

Two kinds of limits matter here. The ranges 30–60 characters for titles and 110–160 characters for descriptions are the limits our own blog checker enforces. They are not Google requirements. Google's pages name no hard limit: there is no limit on how long a title element can be, and there is no limit on how long a meta description can be. In both cases the result is truncated in Google Search "as needed, typically to fit the device width". Our reading: shorter texts are less likely to be cut, so we use the ranges as a target.

For the wording, Google's page on helpful content offers two questions you can put to any draft: does the main heading or page title provide a descriptive, helpful summary of the content, and does it avoid exaggerating or being shocking in nature? Google also says it sometimes shows the meta description in the snippet when it describes the page more accurately than text from the page, so a wrong claim there can reach searchers.

Method 3: let AI write JSON-LD, then validate it

Structured data is code with documented rules, which makes it a good job for AI and for a machine check. Google recommends JSON-LD if your site's setup allows it. For breadcrumbs, Google's documentation says to define a BreadcrumbList with at least two ListItems. The required properties are itemListElement for the list, and name and position for every item. The item address is required for every breadcrumb except the last one. Other types have their own required properties, so open the page of your type in Google's documentation before you trust a draft. A BreadcrumbList for a blog page looks like this, with example.com as the site:

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "BreadcrumbList",
  "itemListElement": [
    { "@type": "ListItem", "position": 1, "name": "Home", "item": "https://example.com/" },
    { "@type": "ListItem", "position": 2, "name": "Blog", "item": "https://example.com/blog" },
    { "@type": "ListItem", "position": 3, "name": "How to use AI for SEO" }
  ]
}
</script>

Paste the code into the Rich Results Test. It has a tab for a page address and a tab for code. Google describes its purpose as showing which rich results can be generated by the structured data a page contains.

Be precise about what "passed" means. Google's structured data guidelines say: "Google does not guarantee that your structured data will show up in search results", and that holds even when the markup is correct according to the Rich Results Test. Passing is not enough on its own: Google's breadcrumb page says "You must follow these guidelines to be eligible to appear with breadcrumbs in Google Search", and the page has to be accessible to Google and not blocked by robots.txt, a noindex tag or a login. Passing does not create a rich result. Breadcrumbs, for example, are documented as a feature available on desktop.

The test checks the structure. It cannot tell you whether the content is true, and a model can add properties that look plausible. The same guidelines say not to mark up content that is not visible to readers of the page. So compare the markup with the page: every name, date and address in it must appear there. After the release, SEOgrinder reports pages without structured data, JSON-LD blocks it cannot read and entries that lack fields Google requires. Not every page needs markup: its report says that missing structured data is fine if no rich result type fits the page.

The practice to avoid: many pages without value

Google's spam policy defines it in one sentence: "Scaled content abuse is when many pages are generated for the primary purpose of manipulating search rankings and not helping users." One of the examples in the policy is "Using generative AI tools or other similar tools to generate many pages without adding value for users". The policy adds that this is abusive "no matter how it's created". The same list includes scraping feeds, search results or other content to generate many pages, stitching together content from different pages, and creating many pages that make little sense to a reader but contain search keywords.

The consequences are stated plainly. Sites that violate the policies "may rank lower in results or not appear in results at all", and Google detects violations through automated systems and, as needed, human review that can lead to a manual action. For content like this on your own site, the policy's instruction is short: "exclude it from Search".

This is the one practice we do not test. We will not publish pages without value to see what happens. The warning rests on Google's text alone.

How much AI is too much?

Google names no number. In the four Google documents we read for this article (the policy, the guidance on generative AI content, the helpful content page and the FAQ post) there is no percentage and no share of AI-written text that counts as too much. A rule of the form "up to X percent AI is safe" does not come from Google.

What Google gives instead are questions. The helpful content page asks, among others, "Are you using extensive automation to produce content on many topics?" and "Are you producing lots of content on many different topics in hopes that some of it might perform well in search results?" These are questions to ask yourself, not rules. It calls the "why" behind a page perhaps the most important question: content should be created primarily to help people.

The same page says that search quality raters look at effort, and that using generative AI to produce large amounts of text "without manual oversight or curation represents little to no effort". Google adds in its guidance on AI content that the raters' ratings do not directly influence ranking. Our assessment, not a Google statement: ask of every page whether it contains something a reader cannot get by summarizing other pages, such as your own measurements or screenshots, and whether a person checked the facts. If you cannot answer yes, the problem is the page, however much or little AI went into it.

Our test: what we will do on 5 November 2026

The headline says "we tested". This section is the plan for that test. It has not been run, and nothing in it is a result.

Disclosure: SEOgrinder is our tool, and we build it.

The test is planned for 5 November 2026 as three runs of about 10 minutes each, carried out by the person running the test on our own site:

1. Method 1, one page with an open finding. Scan the page in SEOgrinder and note the open finding. Hand it over to a coding agent with Hand over to AI, read the diff, release the change and press Check again.
2. Method 2, a new article of this blog. Let AI draft titles and meta descriptions from the article text. Count the characters, check every statement against the article and choose one variant.
3. Method 3, one blog page. Let AI write the JSON-LD for the page and paste the code into Google's Rich Results Test.

What we document: date and time, the versions of the tools used, the SEOgrinder report before and after run 1, the commit that contains the change, the character length of each draft from run 2, and a screenshot of the Rich Results Test from run 3.

Result: to follow. We enter it after the third run. The fourth practice, publishing many pages without value, is not part of the test.

Common mistakes

• Publishing AI drafts unread. Google calls the check "critical", and it covers titles, descriptions, structured data and alt texts.
• Trusting the model's character count. Count it yourself.
• Marking up what is not on the page. Google's guidelines say structured data must describe content that readers can see.
• Reading "passed" as "rich result guaranteed". Google states that it does not guarantee structured data will show up.
• Handing an agent the whole report at once. A big diff hides which change caused which problem. One finding at a time is slower and much safer.
• Applying a method to hundreds of pages without a sample. Try it on one page, read the result, then widen.
• Quoting a percentage as if it came from Google. The documents we read give none.

FAQ

Does Google penalize AI-written content?

Google says that appropriate use of AI is not against its guidelines. Its spam policy covers content produced for the primary purpose of manipulating search rankings, and sites that violate it may rank lower or not appear at all. The deciding factor is purpose and value for users, not the tool.

Do I have to label AI content?

Google words it as a recommendation. Its guidance says to "consider adding information on how your content was created in a way that makes sense for your audience", and its FAQ says to add such disclosures "when it would be reasonably expected".

Can I list AI as the author of an article?

Google's FAQ says that giving AI an author byline "is probably not the best way" to make clear to readers when AI is part of the content creation process.

Will these methods raise my rankings?

Nobody can promise that. Method 1 removes technical errors, method 2 changes how a page may appear in results, and method 3 validates the markup that rich results depend on. None of them guarantees a position.

Your next step

Run the free scan of your site in SEOgrinder, pick the first task of the list and try method 1 on that one finding: hand it over, read the diff, release, press Check again. If you want the safety rules for working with an agent first, read how to fix SEO issues with an AI coding agent. We will add the results of our own test to this article after 5 November 2026.