Google’s AI search guide is out. We turned it into a test.

Google just told the AI-search hype crowd to sit down.

After two years of “AEO consultants” pitching chunked content, fake mentions, and special markup files nobody reads, Google published their official guidance for showing up in AI Overviews and AI Mode. The whole thing collapses to one sentence: it’s still SEO.

AI Overviews and AI Mode are grounded in the same Search index, ranked by the same systems, with retrieval-augmented generation and query fan-out doing the model work under the hood. If your site shows up in Search, it can show up in AI Overviews. If it can’t, no amount of llms.txt is going to save you.

Google was explicit about what doesn’t matter:

• llms.txt files.
• “Chunking” content for the model.
• Writing in some imagined “AI voice.”
• Buying inauthentic mentions — they actively spam-filter that.

What does matter is whether your site is actually indexable. Crawl access. Page experience. Structured data. Real content written for real humans.

That’s a list. So we built it.

L4: Indexable. Twelve checks.

All twelve run from a public URL. No auth, no install, no markup you have to invent.

The cheap ones first, because most sites still miss them:

• Robots.txt isn’t blocking Googlebot.
• An explicit Google-Extended policy is declared.
• The homepage doesn’t accidentally noindex itself.
• A sitemap exists and robots.txt points to it.
• HTTPS, with http redirecting up.
• A viewport meta tag for mobile.
• A self-referential canonical URL.

The slightly less cheap ones:

• At least one JSON-LD block that actually parses.
• An Organization, WebSite, or LocalBusiness entity declared inside it.
• One h1, real h2s, no skipped heading levels.
• 80% of images carry alt text.
• 300+ words of real content on the homepage after stripping nav and footer.

Most sites pass the first batch. The JSON-LD entity check and the 300-word substance check are where things get embarrassing. A surprising number of homepages are mostly logo, hero animation, and one CTA — Google’s grounding has nothing to work with.

Send your agent, not yourself.

The dashboard works. But the faster path is to skip it entirely.

Plug your coding agent into our MCP server, point it at your domain, and let it do the work. Claude Desktop, Claude Code, Cursor, Cline, ChatGPT — anything that speaks MCP. We expose scan_url and a batch variant. The agent reads the scan, walks your codebase, opens the PR. Setup is on the MCP page.

That’s how we use our own tool. Pasting a checklist into Claude is the new Google Search.

What to do tomorrow morning.

Scan your domain. Write down the L4 score.

Fix the five-minute stuff first — canonical, viewport, alt text, the sitemap line in robots.txt. Then the structural stuff — entity schema, a declared Google-Extended policy. Then the hard one: 300 words of something only you can write.

That last one is what Google was really telling you to do, and it’s the only one your competitors will skip.

AI search is search. Google said the quiet part out loud, and the people selling “GEO packages” have a worse week ahead of them than you do.

Run the scan. Send your agent. Get back to building.