10 tips for ranking on ChatGPT that the research actually supports
There is no shortage of advice on "ChatGPT SEO". Most of it is recycled guesswork. We maintain an archive of more than 400 published GEO studies, so we did something different: we kept only the tips that have measured evidence behind them, and we put the number next to each one.
First, a framing note. "Ranking" on ChatGPT is not like ranking on Google: there is no position two you hold for weeks. There is a probability of being mentioned, a probability of being cited as a source, and both move constantly. The good news is that research now covers what pushes those probabilities up. Here are the ten levers with the strongest evidence, ordered roughly by how much they weigh.
1. Answer the question completely, on one page
1The largest citation study to date analyzed about two million citations from ChatGPT, Claude, Gemini and Google's AI answers across 10,042 URLs over six months. The strongest page-level signal was not any AEO trick: it was how well the content matched the question being asked.
The number: question-content alignment added roughly 30% more citations (β = +0.37), the biggest page-level effect measured. Checklist items like FAQ blocks, TLDRs and author bios lost their effect once domain was controlled for. Source: What Drives Citations in Production LLMs.
Practical translation: build pages where one real question gets a full answer, with the conclusion stated early. A page that half-answers three questions loses to a page that fully answers one.
2. Get into best-of listicles, or write them
2When researchers tracked 100+ brands across 100,000+ AI answers, one format towered over everything else as a citation source: the ordered "best X for Y" list.
The number: best-of listicles account for about 21% of all citations, the highest leverage of any page format. Reports, comparisons and reviews earn the highest citation rates overall; product and pricing pages earn the lowest. Sources: Kumar, arXiv:2606.20065; Török 2026.
Two moves follow. Pitch the listicles that already get cited in your niche so your brand appears in them. And publish your own honest comparison pages, because the format itself is what ChatGPT reaches for when someone asks "what is the best...".
3. Earn coverage on sites you do not own
3Your own website matters, but it is a minority shareholder in your AI reputation. A study of 167,551 URL citations used by LLMs to ground brand answers across 12 markets measured where that information actually comes from.
The number: 85.7% of citations point to sites the brand does not own. Only 14.3% go to owned properties. A separate benchmark of generative engines found an overwhelming bias toward earned media over brand content. Sources: Zatuchin, arXiv:2606.25787; arXiv:2603.12282.
Digital PR, industry publications, directories with real authority, expert roundups. If third parties do not talk about you, ChatGPT has little to quote when your category comes up.
4. Structure your pages so a machine can lift the answer
4Comparisons of pages that appear in ChatGPT answers against pages that do not show consistent structural differences. It is not decoration: retrieval systems select passages, and a clean hierarchy makes your passages liftable.
The number: 94.7% of pages visible in ChatGPT use H2 subheadings, against 74.4% of non-visible pages. Structured product pages also beat narrative ones at being selected and cited by AI agents in an 18,240-observation study. Sources: Schanbacher; Fawzy and Hegazy 2026.
Descriptive H2s that read like questions or claims, short sections, lists and tables where they fit, the key fact in the first sentences of the section rather than buried at the end.
5. Put verifiable numbers and sources in your content
5Content with concrete figures, named sources and quotable claims gets cited more, a result that has held since the original GEO experiments and keeps reappearing.
The number: adding statistics and citations raises the probability of being cited by roughly 40% in controlled tests. The flip side also holds: engines increasingly discount claims they cannot verify, and unverifiable claims lower your earned visibility. Sources: Crăciun; factor archive at geo.alexamigo.com.
One original number you measured yourself is worth more than ten adjectives. It also makes you the source other pages cite, which feeds tip 3.
6. Use schema for what it actually does
6Structured data is oversold as a magic citation switch. A controlled before-and-after experiment on an online store measured what it really changes in ChatGPT answers.
The number: Schema.org markup did not increase how often the store was cited, but its average position inside the answer jumped from 4.45 to 1.78, with visibility (nDCG@3) up 133% to 339% in discount scenarios. Field data agrees on type-specific markup: pages visible in ChatGPT carry Product schema 17.2% versus 1.8% for invisible ones, and FAQPage 6.2% versus 0.8%. Sources: Bouman 2026; Schanbacher.
So: use the schema type that matches each page, and expect better placement and attribution, not more citations out of thin air.
7. Let the bots in and keep the page technically simple
7None of the above works if ChatGPT's crawlers cannot fetch your content. This sounds obvious and keeps being the single most common failure we see in audits: CDN rules and bot protections silently blocking OAI-SearchBot or GPTBot while the team optimizes copy.
The number: mobile compatibility and crawlable robots.txt measurably raise the probability of entering ChatGPT's retrieval pool, and blocking an engine's AI crawler sharply reduces retrieval even when the content remains reachable elsewhere. Pages visible in AI answers also load fast, 418 ms on average in one dataset. Sources: factor archive at geo.alexamigo.com; AI Overviews crawler studies.
Check the server logs for the real bot user agents, not what the robots.txt says in theory. Serve full HTML without requiring JavaScript for the main content.
8. Grow your reviews before your rivals do
8For anything commercial, ChatGPT leans on reputation signals, and the effect has been measured with unusual precision using pairs of identical products where only the brand and review data differed.
The number: with identical products, the recognized brand won 100% of 670 head-to-head tests. It took only +0.075 stars or 1.6 times more reviews for the unknown option to start breaking that preference. Source: Chu and Hou, arXiv:2606.17443.
That is a surprisingly small gap. A steady review generation program on the platforms of your sector is one of the few GEO levers with an experimentally measured threshold.
9. Show up where ChatGPT's citations actually live
9When ChatGPT does cite the web, the mix of what it cites is public and it is not what most content plans assume.
The number: reviews and user-generated content make up 28.9% of ChatGPT's web citations, the largest single slice, ahead of news media at 26.0%. Reddit specifically appears in about 17% of ChatGPT answers, and within non-corporate sources YouTube is cited more often than Reddit, editorial media or Wikipedia. Sources: Similarweb 2026; Kumar, arXiv:2606.20065.
Genuine participation where your customers already discuss your category, plus video content that answers buying questions, covers the two slices most plans ignore. Genuine is the operative word: astroturfing is both detectable and increasingly filtered.
10. Measure daily, not with one screenshot
10Everything above moves a probability, and that probability is noisy. Studies that repeat the same query over time all land on the same conclusion: a single check tells you almost nothing.
The number: between runs of the same query, only 30% to 50% of sources repeat (Jaccard 0.30 to 0.50 depending on the engine), and around 65% of cited sources change from one day to the next. Converging on a stable visibility estimate takes 40+ observations. Sources: Sielinski, arXiv:2607.10341; Schulte et al., arXiv:2604.07585.
Whatever you change, you can only tell if it worked by tracking the same prompts daily and reading the trend, not the snapshot. That volatility is also your opening: sources rotate constantly, so there is a seat to win every day.
What has no proven effect, so you can skip it
The honest list
- Keyword-stuffed "AI optimization" rewrites. A survey of 45 GEO studies found no technique with a stable effect on organic discovery; measured effects exist only within content the engine already retrieved.
- Generic AEO checklists as a bundle. FAQ blocks, TLDRs and author bios lost their measured effect once domain authority was controlled for, and domain authority weighs about six times more than the best page signal.
- Expecting clicks from every mention. Only 6.8% of ChatGPT answers carry web citations at all. Most of the value travels as brand memory, not as a link.
- Copying your Google playbook as-is. Brand size sets your baseline: household names appear in 73% of relevant answers, mid-market brands in 44%, niche brands in 11%. The tips above move you within your bracket and, over time, between brackets, but not overnight.
Questions we get
How long until any of this shows up in ChatGPT?
Retrieval-based mentions can move in days once crawlers can reach the new content, because ChatGPT browses. Mentions that come from the model's training knowledge move on training cycles, months, which is one more reason to focus on the retrieval side.
Do these tips work for other engines too?
Mostly, with different weights. But do not extrapolate blindly: Claude answers from prior knowledge 43% to 52% of the time versus 8% to 29% for ChatGPT, and engines agree on which brand to recommend only about a third of the time. Measure each engine separately.
Which tip should I start with?
Check tip 7 first, because a blocked crawler nullifies everything else and takes an afternoon to fix. Then tip 2, because listicle presence is the highest-leverage single placement the data shows.
Where do these numbers come from?
From the public GEO research archive we maintain at geo.alexamigo.com, which catalogs 400+ published studies, and from the Similarweb 2026 AI search report. Every figure in this article links back to a named study.
See how often ChatGPT mentions you today
Bee LLM tracks your brand daily across ChatGPT and four other engines, shows who gets cited instead of you and which sources feed each answer. Free plan, no credit card.
Start freeSources: What Drives Citations in Production LLMs (~2M citations, 10,042 URLs); Kumar, arXiv:2606.20065; Zatuchin, arXiv:2606.25787; Bouman 2026 (Radboud University); Chu and Hou, arXiv:2606.17443; Sielinski, arXiv:2607.10341; Schulte et al., arXiv:2604.07585; Martinez, arXiv:2607.14035; Similarweb, 2026 Generative AI Landscape. To keep reading: half of small brands never appear in AI answers and AI cites deep pages but visitors land on your homepage.