"Does ChatGPT know about our product?"
I asked it. It didn't. Our competitors showed up. We were completely invisible to AI. Apparently, we'd been building a product for years — and forgot to introduce ourselves to the machines.
After digging into why, I discovered a fundamentally different game from SEO — one that most engineers haven't started playing yet.
AI Runs on Training Data, Not Search Indexes
Google crawls your site, indexes it, and ranks it in search results. This is real-time. If you rank #1 in SEO, you show up.
ChatGPT and Claude work differently.
What AI "knows" is shaped by both its training data and what it can retrieve at inference time. How often your product was mentioned across different contexts — and whether AI can find it during retrieval — determines your visibility.
Google SEO:
Publish → Crawl → Index → Appear in results
AI (LLM):
Mentions across the web → Included in training data
→ Model "knows" it → Appears in responses
You can rank #1 on Google and still be invisible to AI. These are two completely different systems.
You don't have an SEO problem. You have a visibility problem in AI systems.
The Research: What Makes AI Cite You
The Princeton paper "GEO: Generative Engine Optimization" (accepted at KDD 2024) empirically analyzed what influences AI responses. Key findings:
- In controlled experiments, adding citations increased inclusion rates by up to +115% under certain conditions
- Content with statistics saw up to +76% improvement
- Authoritative sources are prioritized
- Smaller sites benefit more from optimization
Google uses link quantity and quality for PageRank. AI uses the quality and quantity of contextual mentions to build knowledge. Fundamentally different games.
Three Scenarios You're Probably In
Scenario 1: Big Players Show Up, You Don't
Enterprise products have thousands of mentions — news articles, tech blogs, comparison posts, user reviews. AI has ingested all of this. When someone asks "compare tools in category X," the big names appear automatically.
Startups and smaller products simply have fewer mentions. Fewer chances for AI to "learn" about you.
Scenario 2: AI Knows You, But Gets It Wrong
Training data includes outdated and incorrect information. Your product description might be wrong, or based on a version you deprecated two years ago. You're "known" but not accurately known. It's like having a Wikipedia page written by someone who used your product once in 2019.
Scenario 3: You're Buried in Competitor Roundups
"Top 5 tools for X" articles feature your competitors. AI learns these articles and associates the category with those competitor names. You're not even in the conversation. You didn't lose the game — you weren't invited to play.
LLMO: What Engineers Can Actually Do
The approach to this problem is called LLMO (Large Language Model Optimization). Just as SEO optimizes for Google's algorithm, LLMO optimizes for how LLMs build knowledge.
In practice, LLMO operates across three layers: training (what the model learns), retrieval (what it can access at runtime), and representation (how easily your content is used in generated responses).
1. Deploy llms.txt
llms.txt is a guide file for AI systems reading your site. Think of it as robots.txt for AI — except instead of saying "don't look here," you're saying "please look here."
# llms.txt
## About
- Product: [Your Product]
- Category: [Category]
- Key Features: [Description]
## Reference Pages
- /about
- /features
- /docs
Some AI systems have started experimenting with llms.txt-style guidance, though adoption is still emerging. Setup cost is near zero. Specification
2. Implement JSON-LD Structured Data
AI processes structured data more efficiently than raw HTML. Implementing Organization, Product, and FAQ schemas helps AI understand your product accurately.
<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "SoftwareApplication",
"name": "Your Product",
"description": "A concise description AI can quote directly",
"applicationCategory": "DeveloperApplication",
"operatingSystem": "Web"
}
</script>
Many engineers already implement JSON-LD for SEO. It doubles as LLMO.
3. Create Citable Content
If citation quality and quantity matter in training data, create content that others want to reference:
- Technical blogs and documentation — Platforms like DEV.to and Stack Overflow are heavily used in AI training
- Original data — Benchmarks, surveys, unique research that others cite
- Appear in comparison articles — Get included in "best tools for X" roundups
Writing technical articles on DEV.to isn't just developer marketing — it's LLMO. AI systems train on this content.
4. Structure Content in Markdown
Markdown is often easier for AI systems to parse due to its clean structure. Use:
- H1-H3 headings for structure
- Bullet points for information density
- Language-specified code blocks
The READMEs and docs engineers write every day are already LLMO-optimized formats.
SEO vs LLMO: The Key Differences
| SEO | LLMO | |
|---|---|---|
| Target | Google's crawler | LLM training & inference |
| Evaluation | Link quantity & quality | Mention context & volume |
| Time to effect | After indexing (days) | Next training cycle (months) |
| Key tactics | Backlinks, meta tags | llms.txt, structured data, citable content |
| Measurement | Google Search Console | Not yet established |
The biggest difference is time to effect. SEO is a sprint — index today, rank tomorrow. LLMO is a long game — plant seeds now, harvest when the next model trains. Those who move now will be recognized by AI six months from now. Those who wait will wonder why they're still invisible.
Start Today
- Ask AI about your product — Understand your current visibility
- Deploy llms.txt — 30 minutes of work
- Add JSON-LD schemas — Works alongside existing SEO
- Write technical content — Every article is both developer outreach and LLMO
Most teams are still optimizing for search engines. Very few are optimizing for AI systems.
That gap is your opportunity.
References
- GEO: Generative Engine Optimization — arXiv:2311.09735 (Princeton University, KDD 2024)
- LLMO Framework — A systematic reference for LLMO components and tactics
- llms.txt proposal — Specification for the llms.txt standard
Originally published in Japanese on Qiita. This English version has been adapted for a global audience.
和訳はこちら（Qiita）をご覧ください。
Top comments (1)
I had the same moment — ranking well on Google but completely invisible in AI answers.
After that, I started writing more practical docs + small DEV posts, and slowly saw my product getting mentioned correctly. Not fast like SEO, but it works over time.
Feels like we’re all early in this shift — those who start now will benefit later.