The Loop Closing on Itself: How ChatGPT and Claude Started Citing Us (and Bringing Real Users)

TL;DR — what happened, in one paragraph

We built AI Exposure Tool to help products get cited by AI engines. The honest first move was to use it on ourselves. In April 2026 we audited aiexposuretool.com, found a 56/100 score, and rebuilt the AEO posture from scratch — robots.txt, llms.txt, JSON-LD, FAQPage schema, EEAT pages, the whole stack. Score climbed to 91. Then we shipped 8 deep technical reference docs at /docs/*. Two weeks later, ChatGPT started citing our docs in its answers, and real users — exactly our ICP — clicked through and signed up. Three ChatGPT-attributed signups followed (April 25, May 6, May 11); one ran 5 audits and set up project tracking on day one. Then in June it generalized: our first Claude-attributed user landed straight on our /security tool from a claude.ai conversation. Two different AI engines — ChatGPT and Claude — are now sending us real, fully-attributed users. The loop closed: the playbook the platform sells, demonstrated on the platform itself.

Act 1Auditing ourselves — 56 to 91 in 4 weeks

On March 15, 2026, we ran our own tool on aiexposuretool.com. Score: 56/100. For a tool that promises to fix AEO, the result was humbling.

The audit found:

• robots.txt did not explicitly allow AI crawlers (GPTBot, ClaudeBot, PerplexityBot, Google-Extended)
• No llms.txt — AI had no first-party brief to read
• No JSON-LD structured data — no SoftwareApplication, Organization, or FAQPage
• No FAQ schema markup
• No comparison pages
• Zero social proof visible to crawlers

We had 870 Google Search impressions and growing. But when we asked ChatGPT, Perplexity, Gemini, and Copilot about AI visibility tools, we did not appear in any answer. SEO traction without AEO traction.

Over four weeks (March 15 → April 16), we worked through every category. The score climbed, dipped on redesigns, recovered, and peaked at 91 on April 16 after a focused 4-hour fix session.

Score timeline

Every fix, with time and impact

Act 2Shipping citation-bait technical docs

A 91 score is hygiene. Hygiene is necessary but not sufficient — AI engines also need something to cite. The next move was to publish technical reference content in the exact format ChatGPT and Perplexity prefer to surface in answers: deep, comprehensive, with structured data.

Between April 18 and May 7, 2026, we shipped 8 technical reference docs at /docs/*:

• /docs/llms-txt — complete reference for the llms.txt standard
• /docs/ai-crawlers — every AI crawler grouped by owner with user-agents and IPs
• /docs/robots-txt-for-ai — robots.txt directives, templates, common mistakes
• /docs/llm-json — full spec for the JSON variant of llms.txt
• /docs/json-ld-for-aeo — schema.org JSON-LD reference filtered for AEO
• /docs/per-llm-citation-behavior — how each engine differs (ChatGPT, Perplexity, Claude, Gemini, AI Overviews, Copilot, DeepSeek)
• /docs/ai-exposure-score — every signal in the score with weights
• /docs/answer-engine-glossary — A-Z glossary of AEO/GEO terms

Every doc shipped with three JSON-LD blocks: TechArticle (or with proficiencyLevel), BreadcrumbList, and where applicable FAQPage. Every doc was internally cross-linked at the footer. Every doc led with the answer in the first paragraph (answer-first content is the single highest-citation factor).

Total time investment: roughly 18 hours. Total citation surface area created: 8 deep, structured pages aimed precisely at how ChatGPT structures answers about technical AEO topics.

Act 3ChatGPT and Claude cite us — and bring users back

On April 25, 2026, the first ChatGPT-attributed signup landed.

We track this by writing firstReferrer, firstReferrerDomain, utmSource, and firstLandingPath to every UserLead record in our database. When a user signs up, we know exactly where they came from. The first one had:

firstReferrer:        https://chatgpt.com/
firstReferrerDomain:  chatgpt.com
utmSource:            chatgpt.com
firstLandingPath:     /?utm_source=chatgpt.com

ChatGPT cited our content in an answer. The user clicked through. They signed up. Eleven days later, on May 6, 2026, the second ChatGPT-attributed signup arrived — this one landing directly on /docs/llms-txt:

firstReferrer:        https://chatgpt.com/
firstReferrerDomain:  chatgpt.com
utmSource:            chatgpt.com
firstLandingPath:     /docs/llms-txt?utm_source=chatgpt.com

That one is the proof point. ChatGPT got asked something about llms.txt, cited our docs page in the answer, the user clicked the citation, landed on the page, and signed up. The exact outcome we promise customers — happening to us, demonstrably, with full attribution.

It kept going. On May 11 a third ChatGPT-attributed signup arrived — this time a marketer at an 11-50 person team, a different ICP segment than the first two developers, again with firstReferrer=https://chatgpt.com/.

Then, on June 13, 2026, the loop widened beyond a single engine. Our first Claude-attributed signup landed — and it didn't come through ChatGPT at all:

firstReferrer:        https://claude.ai/
firstReferrerDomain:  claude.ai
firstLandingPath:     /security
acquisitionSource:    ai

This user asked Claude something, Claude surfaced our /security tool, and they clicked straight through to it and ran an audit — no homepage, no Google, no detour. Two distinct AI engines, ChatGPT and Claude, are now independently citing us and sending real, attributed users. That's the difference between "a ChatGPT fluke" and "a repeatable channel."

The users who came through

All four users are anonymized below for privacy. Every one is real, in the database, and squarely in our target ICP — developers, marketers and agencies who needed AEO help. Three came through ChatGPT, one through Claude.

Every one of them skipped Google entirely. They asked an AI assistant — ChatGPT or Claude — the assistant cited us, and they came directly. The single reason they exist as users is that the AEO playbook worked.

Why this is the complete circle

Most product companies talk about AEO as a thing they do for customers. We can show it as a thing the platform does to itself:

The platform is the proof. The proof is the platform.

How to replicate this for your product

This is the playbook in five lines:

1. Audit yourself. Run your AI Exposure Score. Most sites land 50-70. Whatever the gap is, fix it before doing anything else.
2. Ship the hygiene. robots.txt allows AI crawlers, llms.txt deployed, JSON-LD on every page (Organization + the relevant content type), FAQPage where you have Q&As. This alone gets you to ~85.
3. Build citation surface area. Write technical reference docs about your category in the format AI engines cite — deep, comprehensive, with TechArticle + FAQPage JSON-LD. Cross-link them. Lead every page with the answer.
4. Wait 1-2 weeks. AI engines re-crawl on a 7-14 day cycle. That's your indexing window. Don't evaluate before week 3.
5. Track who comes through. Write firstReferrer, firstReferrerDomain, and utmSource on every signup. When ChatGPT, Perplexity, or Claude cites you, you'll see it in your data immediately.

That's exactly what we did. It worked on us. It will work on your product.

Current state — what is working, what is not

What we are shipping next

Key learnings

1. Hygiene gets you to 85. Citation surface gets you cited. A perfect technical AEO posture without quality reference content gives you visibility but not citations. The docs are what ChatGPT actually quotes.
2. llms.txt + JSON-LD = instant 20+ point jump. The two highest-ROI changes any site can make. Five-minute, 12-minute fixes.
3. Redesigns kill scores. Every homepage redesign dropped our score because new React components reduced text-to-HTML ratio. Ship content alongside code.
4. Trust signals are the hardest to earn. Technical fixes take hours. Social proof takes weeks of real customer relationships. Worth investing in early.
5. Score fluctuates and that is normal. We have been between 78 and 91. Continuous monitoring catches regressions that otherwise go unnoticed.
6. Google impressions ≠ AI mentions. We had 870 Google impressions before our first AI citation. SEO and AEO are different games with different signals.
7. Track firstReferrer or you won't see the wins. Without referrer tracking, all four AI signups — three from ChatGPT, one from Claude — would have looked like "direct traffic" — invisible.

FAQ