Every client call I’ve had in the last few months eventually gets to the same question: “Are we even showing up in AI Overviews and ChatGPT answers?” Most of the time, nobody actually knows, including the SEOs sitting across the table.
So instead of giving another opinion, I picked 10 websites (a mix of [industry/niche, e.g. CBI/RBI advisory sites, B2B, Hospitality, Wellness sites]) and ran a manual AI search readiness audit on each one. No tool did this for me. I opened every page, checked the technical layer, read the content the way a language model would have to parse it, and scored what I found.
Here’s what actually moved the needle, what didn’t matter nearly as much as the hype suggests, and the scorecard you can steal to run this audit on your own site.
Why I Ran This Audit
Technical SEO has quietly split into two jobs: optimizing for crawlers that index pages, and optimizing for models that synthesize answers. The second one doesn’t have a Search Console yet. There’s no dashboard that tells you whether Gemini cited you in an AI Overview or whether ChatGPT pulled a stat from your page. So I built my own way to check.
I wanted honest findings, not a “10 tips for AI SEO” listicle, which is why every score below came from actually looking at the site, not running it through a single tool and copy-pasting the output.
The Scorecard Framework
I scored each site across five categories, 0-4 points each (20 points total). This is the exact framework; use it on your own site or a client’s.
| Category | What I checked | Points |
|---|---|---|
| Crawlability for AI bots | robots.txt rules for GPTBot, ClaudeBot, Google-Extended; whether key pages are JS-rendered only | 0-4 |
| llms.txt presence & quality | File exists at root, is curated (not a full sitemap dump), links to genuinely high-value pages | 0-4 |
| Answer-first content structure | Direct answers near H1/H2s, question-based headings, scannable lists over walls of text | 0-4 |
| Structured data (schema) | FAQPage, HowTo, Organization/Person schema present and valid, not just installed-and-forgotten | 0-4 |
| Citable authority signals | Original data/stats, clear author bylines, E-E-A-T signals a model could point to as a “source” | 0-4 |
A site scoring 16+ was “AI-ready.” Most weren’t close.
What I Found: Before / After Screenshots
The Findings That Actually Mattered
1. Nobody was blocking AI crawlers on purpose. They just never checked. Out of the 10 sites, 3 had robots.txt rules that hadn’t been touched since before AI crawlers existed. A few were accidentally blocking GPTBot and ClaudeBot through an old “block all bots except Googlebot” rule someone set up years ago and forgot about.
2. llms.txt is still early, but the sites that had one weren’t wasting the opportunity. An llms.txt file is a simple, Markdown-formatted file sitting at the root of a domain (yourdomain.com/llms.txt) that gives AI systems a curated map of a site’s most valuable content, rather than making a model wade through navigation menus, cookie banners, and JavaScript to figure out what matters. It’s a companion to robots.txt and sitemap.xml, not a replacement: robots.txt still governs which bots are actually allowed in, and sitemap.xml still lists everything that exists (sourced via AgileBrandGuide). Adoption is still small and no major AI provider has officially committed to honoring it, but the sites that had a clean, curated one were noticeably easier for me to navigate as a human, too, which says something.
3. Answer-first formatting beat keyword density every time. The sites scoring highest weren’t the ones stuffing the target phrase into every H2. They were the ones answering the actual question in the first 1-2 sentences under a heading, then expanding underneath. That structure lines up with how AI Overviews are built: Google synthesizes summaries from multiple sources and increasingly favors pages that resolve intent quickly over pages that just rank well (per Logic Inbound’s 2026 analysis).
4. Schema markup was either done well or not maintained at all. No middle ground. 7 of the 10 sites had FAQPage or HowTo schema installed, but when I validated it, 5 had broken or outdated markup that hadn’t been touched since the page copy changed. Schema you installed once and never revisit is close to not having it.
5. The sites with a real point of view outperformed the “safe” ones. This was the most interesting finding and the hardest to quantify. Pages with a clear author, a specific stance, and original data consistently felt more “citable” than generic, committee-written pages, even when the generic page technically covered more keywords.
Quick Self-Audit: Run This on Your Own Site
- Check robots.txt for GPTBot, ClaudeBot, Google-Extended, and PerplexityBot rules
- Confirm your most important pages render without JavaScript (view-source, not just DevTools)
- Add or clean up an llms.txt file (curated, not a full site dump)
- Rewrite your top 5 pages so the answer appears in the first two sentences under each heading
- Validate existing schema with Google’s Rich Results Test. Don’t assume it still works
- Add a real byline and at least one original data point or stat per page
If you want the full 20-point scorecard as a template you can copy, or want someone to actually run this on your site instead of DIY-ing it, that’s the next section.
Book an SEO Audit
If you read this and immediately started wondering what your own score would be, that’s usually the sign it’s worth having someone else check. I run this exact audit for a small number of clients each month.
Leave a Reply