1. Rank for the searches behind the answer
Assistants with web search turn a buyer's question into their own searches, then read the results. ChatGPT exposed those searches in 100% of the 25 answers we measured, and read 18.3 pages per answer on average. If you are not among those pages, you are unlikely to be named. Find the searches, check your rank for each at your buyers' location, and build or improve the one page that answers each.
2. Be where the AI looks
Most citations go to third parties, not to the brands being recommended. Across the 1,079 citations in our measurements, the mix was:
| Source type | Share of citations |
|---|---|
| Other | 63% |
| Competitor | 11% |
| Government or education | 9% |
| Your site | 8% |
| YouTube or video | 3% |
| Directory or listing | 2% |
| News or media | 2% |
| Reddit or forum | 1% |
Review sites, directories, forums and comparison articles in your category are where to get listed, reviewed or mentioned. Track which ones the engines cite for your questions, and which of those mention you.
3. Make each page answer directly
The strongest controlled evidence on content comes from Aggarwal et al. (KDD 2024): adding quotations from credible sources raised visibility in generated answers by about 41%, adding statistics by about 32%, and citing sources by about 30%. Keyword stuffing lowered it by about 10%. Put the direct answer first, then the evidence.
4. Let the crawlers in
Check robots.txt and your firewall for OAI-SearchBot, ChatGPT-User, Claude-SearchBot, PerplexityBot and Googlebot. Blocking training crawlers such as GPTBot or CCBot is a fair choice; blocking search crawlers hides you from answers. Many AI crawlers do not run JavaScript, so key facts must be in the raw HTML.
5. Keep the facts consistent
Assistants repeat what they find. If your prices, hours or service areas differ between your site and your listings, answers will too. Ask the assistants about you by name and compare what they say with your site.
What does not help
- llms.txt: no major engine is known to read it.
- Schema markup as a citation lever: a controlled study found no measurable effect on AI citations. Use it for search features, not for AI answers.
- Keyword stuffing, invented reviews and hidden text.
Measure before and after. Our AI visibility tracker reports changes only when they are statistically real; the free check shows where you stand now.
Sources: P. Aggarwal et al., "GEO: Generative Engine Optimization", KDD 2024, arXiv:2311.09735. Citation mix and search counts: Sightings's own measurements, October 2026.