If you want ChatGPT, Perplexity or Google's AI Overviews to cite your site, the evidence points to a short list. Publish pages that already rank in conventional search. Put a complete answer high on the page and back it with named statistics, quotations and sources. Structure the page so a machine can lift sections cleanly, and keep AI crawlers unblocked. Almost everything else marketed as "GEO" either lacks evidence or has been measured and found to do nothing.
First, terms. Generative Engine Optimization (GEO) is the practice of making your content more likely to be retrieved, quoted and linked as a source by answer-generating engines. That means ChatGPT with search, Perplexity, Google's AI Overviews and AI Mode, Gemini, and Microsoft Copilot. It overlaps with SEO but optimises for being cited in an answer rather than being clicked in a results list.
This guide goes through what the published studies actually show, with dates, sample sizes and the caveats the studies themselves state. We also flag the numbers that circulate in misquoted form, because half the GEO advice you will read is built on those misquotes.
Our finding: before writing this, we snapshotted who currently ranks for "how to get cited by ChatGPT" and "best AI visibility tools" (30 August 2026). Every first-page result we logged was written by a tool vendor or an agency selling GEO services, such as Frase, Evertune, Pixis and Oltre. Not one was a neutral party publishing its own test data. That conflict of interest is the gap this site exists to fill, and it is why every claim below carries a primary source and a date.
Key Takeaways
- Citation odds track conventional rankings: Seer Interactive found a ~0.65 correlation between Google page-1 rankings and brand mentions in LLM answers (Jan 2025). GEO is not a shortcut around SEO.
- The strongest measured on-page tactics are adding quotations, statistics and cited sources, worth up to 40% more generative-engine visibility in the original Princeton GEO benchmark (KDD 2024).
- Ranked lists dominate: 63% of ~400 million citations across six LLMs pointed to listicles (Evertune, May 2026).
- Two popular tactics measurably underperform their hype: schema markup showed no meaningful citation uplift in Ahrefs' matched-control study, and no major AI platform confirms reading llms.txt.
- The prize is small but concentrated: AI referrals average ~1.08% of sessions (Conductor, 2026) yet convert far above organic in the cases measured so far.
What Decides Whether AI Search Cites You?
The single best predictor found so far is boring: already ranking in conventional search. In January 2025, Seer Interactive queried GPT-4o with 10,000 questions built from 300,000+ finance and SaaS keywords and joined the answers against SERP data. Brand mentions in LLM answers correlated with Google page-1 rankings at roughly 0.65, with Bing close behind at 0.5 to 0.6. Backlinks, the currency of classic SEO, showed weak-to-neutral impact on mentions.
On the page itself, the original academic work on this problem still holds up as the best controlled test. The Princeton-led paper that coined the term GEO (Aggarwal et al., KDD 2024) ran a 10,000-query benchmark. Optimising how content is presented improved visibility in generated answers by up to 40%. The top tactics were adding quotations, adding statistics, and citing sources. Worth stating plainly: that result comes from simulated engines and an academic visibility metric, not live ChatGPT logs. Treat 40% as a ceiling in ideal conditions, not an expectation.
Format matters more than most operators expect. Evertune's citation study (May 2026) examined the most-cited URLs across ChatGPT, Copilot, Gemini, AI Mode, AI Overviews and Perplexity, roughly 400 million citations in total, and found 63% pointed to listicles. Depending on the model, ranked lists made up 40% to 65% of the most-cited URLs. If your content honestly fits a ranked or numbered structure, that structure is doing citation work for you.
Each Engine Has Its Own Citation DNA
Optimising "for AI" as one channel is a category error; the engines source answers differently. Profound's analysis of 680 million citations (August 2024 to June 2025) found Wikipedia alone held 47.9% of ChatGPT's top-10 source share, while Reddit held 46.7% of Perplexity's. Google's AI Overviews spread citations far more evenly, with no single community source above 2.2%.
Dominant top-10 citation source, by engine — Profound, 680M citations, Aug 2024–Jun 2025
| Engine | Dominant source | Share of top-10 citations |
|---|---|---|
| ChatGPT | Wikipedia | 47.9% |
| Perplexity | 46.7% | |
| Google AI Overviews | no single community source | under 2.2% |
And that DNA moves fast. Semrush's three-month tracking study (230,000+ prompts, 100M+ citations, July to October 2025) caught a collapse in real time. ChatGPT's Reddit citations fell from roughly 60% of prompt responses in early August 2025 to about 10% by mid-September. Wikipedia fell from ~55% to under 20% in the same window, and the collapse was ChatGPT-only.
The practical lesson is not any point value. It is that citation mixes are re-weighted in weeks, so a strategy bolted to one source pattern rots quickly. For example, an operator who spent August 2025 seeding Reddit threads to chase ChatGPT citations watched that channel lose most of its weight within six weeks. Cite the volatility, not the snapshot.
The Traffic Math: Fewer Clicks, Better Visitors
Be honest about what winning looks like. AI answers suppress clicks overall. Ahrefs' updated 300,000-keyword analysis (February 2026) measured the damage. When an AI Overview is present, the desktop CTR of the #1 organic result fell from 0.073 in December 2023 to 0.016 in December 2025. That is a 58% relative drop, and every position from 1 to 10 lost between 58% and 19%. Pew Research's behavioural study of 68,879 real searches (March 2025) points the same way. Users clicked a traditional result on just 8% of searches that showed an AI summary, versus 15% without. Links inside the summary itself were clicked on roughly 1% of visits.
Share of searches with a click on a traditional result — Pew Research, 68,879 searches, Mar 2025
| Search result page | Clicked a traditional link |
|---|---|
| Without an AI summary | 15% |
| With an AI summary | 8% |
| Link inside the AI summary itself | ~1% |
So why chase citations at all? Because the smaller stream converts disproportionately, at least in every case measured so far. Conductor's 2026 benchmark across 13,770 domains and 3.3 billion sessions puts AI referrals at 1.08% of sessions on average (2.8% in IT), with ChatGPT driving 87.4% of them. The value side is where it gets interesting. For example, Ahrefs reported that AI search produced 0.5% of its own site's traffic but 12.1% of signups. Semrush's modelling values an AI-search visitor at about 4.4x a traditional organic visitor for its vertical. Both numbers are honest anecdotes rather than laws: one company's site, one vertical's model. But the direction is consistent, people who arrive from an AI answer arrive pre-qualified.
There is also a harder truth from the news industry. Similarweb data reported by TechCrunch (July 2025) showed ChatGPT referrals to news sites growing 25x year-on-year. Over the same period, organic search traffic to those sites fell by more than 600 million visits. Citations are replacing a fraction of what AI answers remove. That is exactly why being the cited source, rather than the paraphrased one, is worth fighting for.
The Playbook: Six Changes That Track the Evidence
- Win conventional rankings first. The 0.65 correlation (Seer) makes page-1 Google and Bing presence the closest thing GEO has to a prerequisite. Everything you already know about search intent and content quality still applies.
- Front-load a complete answer. Engines lift self-contained passages. Put the direct answer in the first screen of the page and make each H2 section resolve its own question without needing the rest of the article.
- Add statistics, quotations and named sources. The three top tactics in the KDD 2024 GEO benchmark, worth up to 40% visibility in controlled conditions. Date every number and link the primary source; an unverifiable statistic is a liability, not a signal.
- Use ranked lists where they are honest. With 63% of citations pointing at listicles (Evertune), a genuine comparison or ranked structure earns extraction. Do not force it onto content that is not a list; forced structure reads as spam to humans and adds nothing.
- Keep AI crawlers unblocked, deliberately. Cloudflare's network data (July 2025) shows GPTBot's share of crawl traffic rising from 2.2% to 7.7% in a year, with PerplexityBot's requests up over 150,000%. Meanwhile ~14% of top-10K domains now block at least one AI bot in robots.txt. Blocking is a legitimate licensing stance for publishers, but if citations are your goal, an allowed crawler is table stakes. Check your CDN's bot settings too, such as Cloudflare's AI-bot toggle, which some plans enable by default.
-
Monitor weekly, per engine. Given the ChatGPT re-weighting episode above, a monthly check is too slow. Ask the engines your target questions directly, log who they cite, and watch your referral analytics for
chatgpt.com,perplexity.aiand Copilot referrers.
Here is the whole evidence base in one table, with the strength of each finding stated honestly:
| Tactic | Evidence | Verdict |
|---|---|---|
| Rank on Google/Bing page 1 | ~0.65 correlation with LLM mentions (Seer, 300K keywords, Jan 2025) | Strongest known predictor |
| Statistics, quotes, cited sources | Up to +40% visibility (Aggarwal et al., KDD 2024, simulated engines) | Strong, controlled, but lab conditions |
| Ranked-list structure | 63% of ~400M citations were listicles (Evertune, May 2026) | Strong correlation, use where honest |
| Unblocked AI crawlers | Eligibility requirement (Cloudflare crawl data, Jul 2025) | Table stakes |
| Schema markup for citations | No meaningful uplift, matched controls (Ahrefs, May 2026) | Measured, does not work |
| llms.txt | 5.61% adoption, zero platform confirmation (HTTP Archive, Jun 2026) | No evidence it does anything |
What the Evidence Says Doesn't Work
Schema markup as a citation tactic. Ahrefs tested this properly in May 2026: 1,885 pages that added JSON-LD between August 2025 and March 2026, compared against ~4,000 matched control pages in a difference-in-differences design. Result: AI Overviews citations moved −4.6%, AI Mode +2.4%, ChatGPT +2.2%, none meaningful. Schema still earns conventional rich results and entity clarity, keep it for those reasons, but the "schema gets you cited by AI" claim repeated across agency blogs currently has no measured support.
llms.txt as anything but a lottery ticket. HTTP Archive analysis by Casey Burridge (June 2026) shows adoption among the top 10,000 sites grew from 1.04% to 5.61% in a year. Much of that came from Shopify rolling the file out automatically to its stores. No major AI platform has confirmed it reads the file. We publish one on this site because it costs five minutes and cannot hurt; we expect nothing from it and log any evidence either way.
Chasing crawl volume as if it were traffic. Crawl-to-referral ratios reported from Cloudflare's 2025 Year in Review data run into the tens of thousands of pages crawled per visit referred back (Search Engine Journal's summary). The circulating figures vary by report and are frequently misquoted, so pin the exact number to the primary page before repeating it. Being crawled is not being cited, and being cited is not being clicked. Measure citations and referrals, not bot hits.
How to Check Whether You're Being Cited
The manual loop costs nothing. Take your ten most valuable customer questions and ask them in ChatGPT (with search enabled), in Perplexity, and in a Google query that triggers an AI Overview. Record which domains each engine cites. Repeat weekly; the volatility data above is why. In parallel, segment AI referrers in your analytics, such as chatgpt.com, perplexity.ai and copilot.microsoft.com. Track them against conversions rather than raw sessions, since the value story, per the Ahrefs and Semrush figures above, lives in conversion rate, not volume.
Purpose-built trackers (Profound, Peec, Otterly, Semrush's AI toolkit and others) automate exactly this loop at prompt scale. A neutral, hands-on comparison of these tools is our next piece of work on this site.
Frequently Asked Questions
Does schema markup get you cited by AI search?
The best available evidence says no. Ahrefs' matched-control study (May 2026) found no meaningful uplift on any engine after pages added JSON-LD: AI Overviews −4.6%, AI Mode +2.4%, ChatGPT +2.2%. Keep schema for conventional rich results and entity clarity, not as a citation lever.
Do I need an llms.txt file?
Not on current evidence. Adoption is growing (1.04% to 5.61% of top-10K sites in a year, per HTTP Archive analysis), but no major AI platform confirms reading the file. Ship one if it costs you nothing; expect nothing from it.
How long does it take to get cited?
No reliable published benchmark exists, and citation mixes re-weight in weeks (see the ChatGPT Reddit collapse of August–September 2025). Distrust fixed timelines from vendors; measure weekly.
Is AI-referred traffic worth anything at ~1% of sessions?
The measured cases say yes: Ahrefs saw 0.5% of traffic deliver 12.1% of signups, and Semrush models AI visitors at ~4.4x organic visitor value for its vertical. Both are single-case measurements, but they agree on direction: small stream, high intent.
The Bottom Line
Getting cited by AI search in 2026 is mostly disciplined fundamentals: rank, answer first, prove every claim, structure for extraction, let the crawlers in, and measure per engine every week. The two most-hyped technical tactics fare worst under measurement. Schema-for-citations has evidence against it, and llms.txt has no evidence for it. We will keep testing on this site's own pages and update this guide as the data changes; the updated date above is real. If you want the ongoing results, the about page explains our methodology, and the RSS feed is the subscription.
Originally published at Citable Press. Every statistic here links to a dated primary source; if you find one that has moved or changed, tell me and I'll correct it.
Top comments (1)
{"content":"honestly the part about structured data being a requirement for those specific llms was a wake up call, i thought that was just a google thing"}