To get cited by ChatGPT, make your pages the easiest credible source for the model to build an answer from: add statistics, quote named sources, cite your own references, and write plainly. Each of those has a measured effect size behind it — and at least one has since failed a replication attempt. This guide covers the wins and the caveats.
The stakes: in a 2026 G2 survey of 1,000+ software buying decision makers, 51% said they now start product research with an AI chatbot more often than with Google. If your page isn't among an answer's sources, someone else gets that consideration. (How the engine picks those sources is its own article; this one is about making your page win.)
The three tactics with the strongest measured effect
The 10,000-query benchmark that founded the field tested nine page modifications. Three each lifted a source's share of AI answers by roughly 30–40% on its position-weighted metric:
Relative change in a source’s position-weighted share of AI answers, 10,000-query founding benchmark (KDD 2024). Keyword stuffing shown at its live Perplexity result. Later replications found weaker effects — treat as upside bets, not guarantees.
1. Cite your own sources
Pages that link their claims to the survey, standard, or dataset behind them captured measurably more of the answer. The study's most counterintuitive finding: giving credit gets you credit. A model prefers material it can verify, and visible sourcing reads as verifiable.
Better still for challengers: the gains were largest for pages in the fifth retrieved slot — up to +115% — while the top source's share fell about 30%. Citations are disproportionately powerful for whoever isn't already winning.
In practice: every factual claim on a key page should name where it comes from. "Response time matters" is unciteable. "Median first-response time in our 2025 support benchmark was 4.1 hours," linked to the benchmark, is exactly what an AI answer lifts.
2. Add statistics
"Most customers renew" becomes "87% of customers renewed in 2025." Numbers make an answer specific, and the model can only use numbers that exist in its sources. If you don't have public numbers, publish some — usage data, a customer survey, a benchmark you ran. Original numbers are a real moat: every answer that uses them has to cite you.
3. Add quotations
Direct quotes from named people and reports boosted visibility comparably — and in the study's live Perplexity test, quotation addition was the single best method (+22% over baseline). A named practitioner quote works twice: evidence for readers, quotable material for the model. Keep it real; a fabricated quote is a liability everywhere.
A fourth, quieter result: fluency edits alone — tightening awkward sentences, changing nothing else — improved visibility 15–30%. Plain writing is a tactic, not a nicety.
What newer research changed
Those numbers are from 2024, and two later studies complicate them. A NeurIPS 2025 benchmark found most of these edits ineffective on its tasks — some actively negative — with retrieval rank mattering more. An MIT/Columbia study found 15 common rewriting heuristics all beaten by iteratively optimized rewriting. Taken together: these tactics are high-upside bets, not sure things; fundamentals come first; and the only way to know what works on your pages is to measure. (The full research timeline is in the GEO overview.)
The tactic that measurably backfires
Keyword stuffing performed worse than doing nothing. A generative engine isn't matching keywords; it's reading your page and judging whether it's a credible basis for an answer, and padded pages read like exactly what they are. The numbers — and the full sorting of which SEO habits transfer — are in SEO vs. GEO.
The recipe for a citable page
- Answer the question in the first two sentences. Retrieval favors pages that plainly address the query; skimming humans reward the same thing.
- One claim, one number, one source — per paragraph. Dense, verifiable, liftable.
- One buyer question, one page. A page spread across a dozen questions competes against pages built for exactly one.
- Cut the padding. Every filler paragraph dilutes your citable material.
Before any of this: confirm the engines can actually reach you — robots.txt, CDN bot rules, server-rendered HTML. That checklist is in llms.txt & AI instructions.
Then verify it worked
AI answers are probabilistic; a single spot-check proves nothing. Rewrite your top 3–5 pages, then sample your real buyer questions on a schedule for four weeks and let the citation rate judge the rewrite. The DIY tracking protocol is the companion piece to this one.
Getting cited isn't a trick. It's making your page the easiest credible thing to build an answer from — and checking, on a schedule, that it's working.
Read next: How to check what ChatGPT says about you · SEO vs. GEO