A frequent misconception: people expect ChatGPT to "look up" their site on every question. Reality splits into three distinct paths, and knowing which one produced an answer tells you what to optimize:
| Path | When it happens | What it means for you |
|---|---|---|
| Training memory | Default for most chat questions | Your site only matters if it was in crawl data before the cutoff. New sites are invisible here — no optimization trick changes that. |
| Search-augmented answer | Perplexity always; ChatGPT/Gemini when they judge fresh info is needed (or user toggles search) | This is where llms.txt, clean HTML, and crawl access pay off immediately. |
| Agent tool-call | The user's agent has a browsing or MCP tool | Machine-readable files (llms.txt, x402) decide whether the agent can understand AND transact with you. |
What this means for your FAQ page
- FAQs get fetched when phrased as questions agents expect. Use real buyer language as headings, not clever wordplay.
- Pricing pages need stable URLs and static HTML. JS-rendered prices are invisible to most retrieval paths.
- One canonical fact source beats five contradictory ones. If your llms.txt says €19 and a PDF says €25, assistants quote both — badly.
The practical takeaway
You can't control training memory. You fully control the machine-readable layer that search-augmented answers and agent tools rely on. That's why our audits grade exactly those surfaces — the part you can fix this week, not someday.
Check what assistants can read about you today
Free instant auditWeekly monitoring with fixes: €19/mo