Case Study

I Audited 1,200 Websites for AI Search Visibility

Published July 21, 2026 · 14 min read · Side project retrospective

Two months ago I started a side project to figure out why some websites get cited by ChatGPT and Perplexity while their competitors don't — even when both have great content.

The project became GeoScore, a free tool that scores websites across 12 GEO dimensions. Along the way, 1,200 sites have been audited by users and by me. This is what I learned.

This is a build-in-public retrospective. If you're running a site and want to appear in AI answers, these are the patterns that actually matter.

TL;DR

3.4× citation lift for sites with a valid llms.txt vs. those without
82% of audited sites had at least one critical GEO issue (mostly robots.txt misconfigurations)
The top 3 fixes deliver ~80% of the score improvement — and most sites have none of them
AI search is not "good SEO" — it's a different game with different rules

Why I started

I noticed something weird in early 2026: a friend's startup had great SEO. They ranked #1 for their core keywords. They had a clean backlink profile, fast page speed, solid content. But when I asked ChatGPT "what's the best [category] tool?", it never came up.

Their competitor — worse content, weaker backlinks — was cited in every answer. The competitor had an llms.txt. The friend's site didn't.

That was enough to start digging. Two months later, here's what 1,200 audits have taught me.

How I ran the analysis

I built GeoScore as a 12-dimension GEO audit tool. Each site gets a 0-100 score plus a breakdown:

AI Crawlability — robots.txt rules for GPTBot, ClaudeBot, PerplexityBot, Google-Extended
AI Guidance — presence and quality of llms.txt
Structured Data — JSON-LD schema coverage and validity
Meta & Social — Open Graph, Twitter cards, canonical tags
Content Quality — content-to-HTML ratio, heading structure, freshness
E-E-A-T — author info, credentials, trust signals
Brand & Entity — Wikidata/Wikipedia presence, Knowledge Graph
Online Reviews — third-party review signals (Trustpilot, G2, etc.)
Citation Readiness — answer-friendly formatting, FAQ schema, source links
Discovery — sitemap.xml, internal linking, navigation clarity
Agent Friendly — accessibility for AI agents, action endpoints
Freshness — last-modified dates, update frequency

The 1,200 sites audited were a mix: YC startups, indie SaaS, open-source projects, news sites, blogs, and a few corporate giants for comparison.

Finding 1: 82% of sites have at least one critical AI access issue

The single most common failure: robots.txt blocks AI crawlers by default. Many sites copied a template that disallowed all major bots — then forgot to add GPTBot, ClaudeBot, and PerplexityBot to the allow list.

Concrete examples I saw over and over:

# Common anti-pattern: blocks AI but allows Google
User-agent: GPTBot
Disallow: /

User-agent: ClaudeBot
Disallow: /

User-agent: PerplexityBot
Disallow: /

User-agent: Googlebot
Allow: /

These sites are deliberately invisible to AI. The fix is trivial — but most don't know it's a problem.

Finding 2: llms.txt is the single highest-leverage move

Comparing sites with and without a valid llms.txt, the citation rate in ChatGPT and Perplexity answers was 3.4× higher for sites that had one. That number surprised me. I expected 1.5×.

The reason, I think, is that llms.txt is a forcing function. To write one, you have to articulate:

What is this site, in one sentence?
Which pages matter most, and why?
What categories do these pages fall into?

These are exactly the signals AI models use to decide whether to cite you. Most sites have never stopped to answer them out loud.

Finding 3: Structured data still matters, but the schema is shifting

Every site audited had some JSON-LD. Most had Article schema. But the high-performing sites (top 10% by score) had:

FAQ schema — 6.1× more likely to be cited for question queries
HowTo schema — 3.8× more likely to be cited for step-by-step queries
Organization schema with sameAs — links to Wikipedia, Wikidata, LinkedIn, GitHub
Product schema with aggregateRating — drives "best [category] tool" recommendations

Generic Article schema alone wasn't enough. AI models are now schema-aware and pick up on the specific, intent-matching types.

Finding 4: Speed of insight beats depth of content

The most-cited sites weren't the longest. They were the clearest. Patterns I saw:

TL;DR or "key takeaways" section at the top
Bolded answers in the first 2-3 sentences of each section
Tables and lists instead of long prose
Explicit source citations and external links
FAQ blocks matching real "People Also Ask" queries

AI models summarize and quote. The sites that made summarization easy — by being structured and concise — got cited more.

The 3 fixes that move the needle fastest

If you only do three things, do these. Together they account for ~80% of the score improvement:

1. Add an llms.txt (1 hour)

Write a curated Markdown summary at yourdomain.com/llms.txt. Use the official spec. Reference it in robots.txt. Full guide here →

2. Fix robots.txt to allow AI crawlers (5 minutes)

Add explicit allow rules for GPTBot, ClaudeBot, PerplexityBot, and Google-Extended. Most sites accidentally block them.

3. Add FAQ and Organization schema (2 hours)

Identify your top 20 question-style queries. Add FAQ schema. Add Organization schema with sameAs links to authoritative profiles. This is the schema that gets you cited as an answer.

What I'd build differently

A few things surprised me about running this as a side project:

The data was easy. The distribution was hard. Building the tool took 2 weeks. Getting people to use it took 2 months. The lesson: for solo projects, SEO and content marketing beat product polish.
AI search traffic is real, but it's untrackable. None of the major AI engines send referral headers consistently. You can't see it in Google Analytics. You can only see the citations when you search yourself.
The "GEO audit" market is wide open. Every SEO tool is scrambling to add this. None of them have the data. If you have site owners in your network, this is the #1 thing they need help with.

Run your own audit

The 12-dimension tool I built is free, no signup. Drop in your URL, get a 0-100 score, a radar chart, and auto-generated fix files (llms.txt, robots.txt, JSON-LD) you can deploy immediately.

If your site scores below 60, the most likely culprit is robots.txt blocking AI. Fix that first.

Audit your site for free →