AEO, GEO and llms.txt in 2026: What Google Says Actually Works

What did Google actually say about AI search?

Google's "Guide to Optimizing for Generative AI Features on Google Search" (last updated July 2026) answers the question directly: SEO best practices still apply, because AI Overviews and AI Mode are rooted in the core ranking and quality systems.

It describes two mechanisms. Retrieval-augmented generation pulls relevant pages from the Search index to ground the answer. Query fan-out runs several related searches in parallel to cover sub-questions. Both start from pages that already rank, which is why the guide calls AEO and GEO simply SEO from Google's point of view.

To be eligible at all, a page has to be indexed and eligible to show with a snippet. If you've set nosnippet or blocked crawling, you've opted out.

Does llms.txt help SEO or AI visibility?

llms.txt was proposed by Jeremy Howard of Answer.AI in September 2024: a Markdown file at your domain root listing the pages you want language models to read. It's easy to add, which is why it spread fast.

Google's guide lists it among things you don't need, saying Google Search itself doesn't use such files. Independent studies of server logs reported that AI crawlers rarely request the file at all. The file doesn't hurt, but don't expect traffic from it. Its real use is narrower: documentation sites where coding agents are pointed at it on purpose.

What does "non-commodity content" mean for a developer blog?

Google's example contrasts a generic "7 tips" article with a first-hand account of a specific decision. For developers, commodity content is a tutorial that restates the docs, or a hot take on this morning's headline. Hundreds of pages already say the same thing, and an AI answer can generate it on its own.

Non-commodity content is what only you can write: the exact error message you hit, the config that fixed it, the trade-off you chose and why. My post on why notFound() returns 200 in Next.js is built that way, from changes in this site's code.

• Start with the direct answer, then the detail. That's good writing for people, and it's what gets quoted.
• Show real code, real config and the real error text people paste into search.
• Say where you used it, and only claim what you actually did.
• Cover the gotchas and when not to use the approach. Summaries rarely do.

What can you safely skip?

This is the guide's mythbusting list, next to what it recommends instead.

What about ChatGPT search, Copilot and other AI tools?

These tools are not Google, and several rely on Bing's index in some form. Bing recommends IndexNow for submitting new and updated URLs, and Bing, Yandex, Naver, Seznam and Yep participate. Google does not use IndexNow; for Google, keep your sitemap current and submit it in Search Console.

IndexNow is small to add. You host a key file on your domain and POST changed URLs. This site pings it whenever a post is published.

// POST new or changed URLs to IndexNow (shared with all participating engines)
await fetch("https://api.indexnow.org/indexnow", {
  method: "POST",
  headers: { "content-type": "application/json; charset=utf-8" },
  body: JSON.stringify({
    host: "example.com",
    key: process.env.INDEXNOW_KEY, // also served at keyLocation
    keyLocation: "https://example.com/indexnow-key.txt",
    urlList: ["https://example.com/blog/new-post"],
  }),
});
// 200 or 202 means accepted

The risk with AI-generated posts at scale

Google's guidance on AI-generated content is not anti-AI. It says using generative tools to create many pages without adding value for users may violate the scaled content abuse policy, which has existed since March 2024 and applies however the pages are made.

So the fix for an AI blog is not more posts. I cut this site's generator to about two posts a day, made it answer real search queries in areas I work in, required a direct answer, code where relevant and a first-hand note limited to things I've actually built, and added a quality check that saves weak drafts instead of publishing them. You can see how that pipeline works in the post on running long jobs on Vercel Hobby.

A short checklist

• Every page you care about is indexable, has a self-referencing canonical and appears in the sitemap.
• Titles and descriptions say exactly what the page answers.
• The answer is in the first paragraph; detail, code and caveats follow.
• Author, date and real experience are visible on the page, with a link to an about page.
• Missing pages return a real 404, not a 200.
• Search Console is verified, the sitemap is submitted, and you check the Generative AI performance report.
• For Bing and the tools built on it, IndexNow is wired to your publish step.

Frequently asked questions

Sources & further reading

• Google: Guide to optimizing for generative AI features on Google Search
• Google: Guidance on using generative AI content
• Google: Spam policies (scaled content abuse)
• llmstxt.org: The /llms.txt file proposal
• IndexNow documentation
• Bing Webmaster Tools: URL submission and IndexNow