How Adding llms.txt Triggered AI Search Visibility

The Setup: What We Implemented

In March 2026, we deployed a full Generative Engine Optimization (GEO) stack on apexdigital.ro. The goal was simple: make our site not just rankable by Google, but citable by AI systems. Here is exactly what we did.

The llms.txt file follows the specification proposed by Jeremy Howard of Answer.AI in September 2024. It is a Markdown file placed at the site root -- like robots.txt, but for AI. The format is simple: an H1 title, a blockquote summary, then H2 sections with markdown link lists describing your site content. AI systems consume Markdown natively -- no HTML parsing overhead, no wasted tokens.

The companion llms-full.txt contains the complete site content in one file: every service description with pricing, every case study with metrics, every blog article in full, client reviews, and contact information. Think of llms.txt as the table of contents and llms-full.txt as the full book.

The Discovery: Fan-Out Queries in GSC

Less than 24 hours after deployment, something unexpected appeared in Google Search Console. New queries were showing up, but they did not look like anything a human would type.

These are not natural search queries. They are machine-generated sub-queries, the search footprint of AI systems actively decomposing user questions and retrieving information from across the web. Notice the -site: exclusion chains: each query systematically excludes Reddit, Twitter/X, YouTube, TikTok, and other major platforms. This is how AI systems avoid retrieving the same sources twice.

The -site: exclusion pattern is the signature. When an AI system like ChatGPT or Perplexity processes a user question, it breaks it into multiple sub-queries and searches the web for each. After retrieving results from one source, it excludes that source with -site:domain.com and searches again to find additional perspectives. These cascading exclusion chains prove multi-step retrieval: the AI is actively building an answer from multiple sources, and your site is one of them.

Also notable: a French-language query ("combien de sites web dans le monde sont construits avec wordpress?") appeared at position 2, confirming that AI systems query in multiple languages when constructing answers, even for our English-language blog posts.

The Analysis: What Fan-Out Queries Mean

Fan-out queries are fundamentally different from traditional search queries. They reveal how AI systems think.

• Length: They are typically 10+ words long, far longer than the 2-4 word queries humans use on Google.
• Boolean operators: They contain -site: exclusions, quotation marks, and other operators that humans rarely use in conversational search.
• Decomposition patterns: A single user prompt like "what is GEO and how do I optimize for AI search?" generates 3-5 separate fan-out queries, each targeting a different facet of the answer.
• Source diversity: The exclusion chains prove the AI is deliberately seeking multiple sources to cross-reference and synthesize its answer.

This is the new discovery layer. Traditional SEO gets you into Google's index. GEO gets you into the retrieval pool that AI systems draw from when constructing answers. The distinction matters because AI-generated answers are increasingly replacing the ten blue links -- and if your site is not in the retrieval pool, you are invisible in this new paradigm.

The Evidence: Industry Context

Our experience is not an isolated data point. The shift to AI-mediated search is accelerating across every metric.

• 527% growth in AI-referred website sessions year-over-year in 2025 (Similarweb).
• ChatGPT surpassed 3.5 billion monthly visits -- it is now a primary search interface, not just a chatbot.
• Perplexity processes 15 million queries per day, each generating multiple fan-out searches against the open web.
• Google AI Overviews now appear in a growing percentage of search results, synthesizing answers from multiple sources before users ever see the blue links.
• 844,000+ websites have adopted llms.txt, including Anthropic, Cloudflare, Stripe, GitBook, and Mintlify.

The timeline from our implementation to first visibility was measured in days, not months. This aligns with research showing that AI systems re-crawl and re-index content rapidly -- especially when technical barriers (blocked crawlers, missing structured data) are removed.

The Service We Built From This

This experience became the foundation for our AI Search Optimization (GEO) service. The core promise: we make your brand visible and citable in AI-generated search answers across ChatGPT, Perplexity, Google AI Mode, Gemini, and Claude.

Phase 1: AI Visibility Audit (Week 1-2)

We manually query 15-30 prompts relevant to your business across ChatGPT, Perplexity, and Gemini. These are conversational prompts that mirror how real buyers talk to AI -- not keyword-style queries. We document whether your brand appears, whether your site gets cited, and who shows up instead.

We audit your technical AI crawlability: robots.txt for blocked AI crawlers, JavaScript rendering issues, CDN/firewall rules that return 403s to bots, and whether llms.txt exists. We pull GSC data with a 10+ word query regex filter to check for existing fan-out queries, and check GA4 for referral traffic from chatgpt.com, perplexity.ai, and gemini.google.com.

The deliverable is a report showing your current AI visibility score, which competitors dominate AI answers in your space, which technical barriers exist, and what content gaps prevent citation.

Phase 2: Technical Foundation (Week 2-3)

We create and deploy llms.txt and llms-full.txt. We update robots.txt to explicitly allow all major AI crawlers. If your hosting setup blocks bots (common with LiteSpeed and aggressive security plugins), we whitelist AI crawler IP ranges and user agents.

We implement or fix structured data: LocalBusiness, Product, Organization, FAQ, HowTo, Review schemas. Clean, comprehensive schema directly improves how AI systems interpret and cite your content. We set up measurement infrastructure -- GA4 segments for AI referral traffic, GSC regex filters for fan-out query monitoring, and optional GEO tracking tools like Peec AI for ongoing visibility scoring.

Phase 3: Content Optimization (Week 3-6)

This is the highest-impact phase. We restructure your highest-value pages following AI-extractable patterns: direct answers in the first paragraph (answer-first format), descriptive headings aligned with fan-out query language, embedded statistics and citations (content with verifiable data gets up to 40% more AI visibility), and conversational FAQ sections that match how people prompt AI.

For bilingual clients, there is a critical extra step. Research shows that even when users prompt in Romanian or French, ChatGPT conducts at least one background search in English in about 78% of cases. We ensure strategic English content exists on the pages most likely to be cited.

Phase 4: Off-Site Authority (Ongoing)

AI systems synthesize information from multiple sources. If your brand appears consistently across authoritative third-party sites, you are far more likely to be mentioned in AI answers. We identify which sources AI engines cite for your key topics, optimize your presence on platforms like Reddit, Quora, and YouTube (disproportionately cited by AI systems), and ensure accuracy across directories, Google Business Profile, and reference platforms.

Phase 5: Monitoring and Optimization (Monthly)

Every month, we re-run visibility prompts across AI engines, track GSC fan-out query patterns, report on AI referral traffic in GA4, and deliver a visibility scorecard with competitor comparison, sentiment analysis, and prioritized next actions. Content is refreshed quarterly -- AI systems have a documented recency bias, preferring recently updated content over older material.

What This Means for Your Business

The shift from search engines to answer engines is not hypothetical. It is happening now, and the businesses that act first will own the AI-generated recommendations in their space. Here is the minimum viable action list:

1. Implement llms.txt and llms-full.txt. Give AI systems a structured summary of your business. This is the single highest-leverage technical change you can make.
2. Allow AI crawlers in robots.txt. Many sites accidentally block GPTBot, ClaudeBot, and PerplexityBot. Check yours today.
3. Restructure content for extractability. Lead with answers, not buildup. Use descriptive headings. Embed verifiable data.
4. Add comprehensive structured data. JSON-LD schema on every page: Organization, Service, FAQ, Article. AI systems rely on schema to understand entities.
5. Monitor GSC for fan-out queries. Filter for queries with 10+ words or -site: operators. These are your AI visibility signals.

The five steps above are the minimum viable action list. The complete framework behind this result lives in the 2026 GEO Audit: seven sections, 47 specific checks, and the order we run them on every client engagement.

Frequently Asked Questions

What are AI fan-out queries?

Fan-out queries are machine-generated sub-queries that AI systems create when decomposing a user question into multiple web searches. They appear in Google Search Console as unusually long queries (10+ words) with boolean operators and -site: exclusion chains. They are not typed by humans -- they are the search footprint of AI systems retrieving information to construct answers.

Does llms.txt guarantee AI citations?

No. llms.txt removes a barrier -- it makes your content significantly easier for AI systems to parse and understand. But citation depends on content quality, relevance, authority, and structured data. Combined with proper schema, answer-first content, and AI crawler permissions, it substantially improves your chances. Think of it like making your store accessible versus guaranteeing foot traffic.

How long does it take to see results?

Technical implementations (llms.txt, robots.txt, structured data) can be deployed in 1-3 weeks. In our case, AI fan-out queries appeared in GSC within days of deploying llms.txt. Content optimization takes 3-6 weeks to build citation momentum. Ongoing monitoring and refinement is monthly. Results vary by industry, competition, and existing content quality.

What is the difference between SEO and GEO?

SEO (Search Engine Optimization) focuses on ranking in traditional search results -- Google's blue links. GEO (Generative Engine Optimization) focuses on being cited and recommended by AI systems like ChatGPT, Perplexity, and Google AI Overviews. They share foundations (structured data, quality content, authority) but GEO adds llms.txt, AI crawler permissions, answer-first content structure, and freshness signals. You need both. Our AI Search Optimization service covers the full GEO stack.

Can any website implement llms.txt?

Yes. llms.txt is a simple Markdown file placed at your site root, similar to robots.txt. Any website on any platform can implement it. The file summarizes your site content in a structured format that AI systems consume natively. Over 844,000 websites have already adopted it, including Anthropic, Cloudflare, and Stripe. The companion file llms-full.txt provides comprehensive content for deeper AI ingestion.