How AI Engines Choose What to Cite (28,000 Answers Data)
Most advice on getting cited by AI is a guess written up as a checklist.
We tracked 28,308 answers from ChatGPT, Google AI Overviews, Gemini and Perplexity between July and October 2026, logged out, with every prompt run many times. Then we checked what we saw against the main studies published this year.
AI engines read far more pages than they credit. They cite what their own searches surface near the top, and they lean toward brands they already know. Several of the tactics sold as AI optimization barely register.
Here is what decides who gets cited, and what to do about it.
AI engines read far more than they cite
Every AI answer engine runs the same funnel:
- It decides whether to search the web or answer from memory.
- It rewrites the question into a few short searches.
- It pulls dozens of candidate pages from a search index.
- It scores those pages again and keeps short passages, not whole pages.
- It writes the answer.
- It attaches links to a small share of what it read.
Each step filters hard. Researchers who measured several ChatGPT models found it shows 6 to 8 citations drawn from 19 to 74 candidate pages (Jiang et al., October 2026).
Our own tracking shows the same gap, engine by engine.
How much of what each engine reads ends up cited
ChatGPT cited about a third of the websites it pulled in. Google's AI showed almost only what it cited.
Perplexity was tracked for four brands until mid September.
Source: THREADSLAB tracking in Peec AI, logged out, 24 July to 22 October 2026.
See the numbers
| Share of retrieved websites that were cited | |
|---|---|
| ChatGPT (90 days) | 30.8% |
| ChatGPT (last 30 days) | 25.1% |
| Perplexity (to mid September) | 18.6% |
| Google AI Overviews | 92.9% |
| Gemini | 96.9% |
A note on reading this chart. Google's surfaces mostly show only the sources they end up using, so their rate looks close to complete. ChatGPT and Perplexity expose a wider reading list, which is where the filter becomes visible.
Position inside the engine's own results matters most. In a study of 16,851 ChatGPT queries, a page in the first retrieval slot was cited 58.4% of the time. A page in slot 10 was cited 14.2% of the time (AirOps with Kevin Indig, April 2026).
Many pages are judged on their title and snippet alone. In its fast mode, ChatGPT opened no page at all in 93% of answers and worked from titles and short snippets. Pages were only opened in its slower thinking mode, and an opened page was cited 74% of the time, against 7% for a page that was retrieved but never opened (RESONEO via Search Engine Land, August 2026).
So make your title and the opening line of each section answer the question on their own. In fast mode, they may be all the engine sees.
ChatGPT now searches like an analyst
For the one brand we tracked through the whole period, ChatGPT ran a single search in 96% of the July answers where it searched at all. After its August update, that changed within a week.
ChatGPT searches per answer, week by week
One search per answer in July became three to five from mid August.
One B2B software brand tracked every week. Averages cover answers with at least one recorded search. The dip in the week of 12 October follows ChatGPT's October model release.
Source: THREADSLAB tracking in Peec AI, logged out, weeks of 13 July to 12 October 2026.
See the numbers
| Week starting | Average searches per answer |
|---|---|
| 13 Jul | 1.04 |
| 20 Jul | 1.11 |
| 27 Jul | 1.05 |
| 3 Aug | 1.95 |
| 10 Aug | 4.13 |
| 17 Aug | 4.21 |
| 24 Aug | 3.75 |
| 31 Aug | 4.26 |
| 7 Sep | 4.08 |
| 14 Sep | 3.28 |
| 21 Sep | 4.04 |
| 28 Sep | 5.32 |
| 5 Oct | 4.36 |
| 12 Oct | 2.80 |
For that brand, searches per answer went from 1.08 in July to 3.53 in mid October. Across all six brands tracked in October, ChatGPT averaged 2.08 searches in each answer where it searched, ran more than one search in 64% of those answers, and the longest ran to 25 searches.
The searches changed shape too. In October, across those six brands:
- 51% included a year (2025 or 2026)
- 31% included the word official
- 15% used site: to search inside one specific website
For the brand tracked since July, searches using site: went from under 1% in July to 36% in October, and searches including official went from 4% to 25%. An independent analysis by David Konitzny found the update doubled the sources ChatGPT pulls per prompt, from about 12 to about 24 (via Peec AI, August 2026).
The shortlist often comes from memory. Data from Graphite shows that 59.6% of ChatGPT's searches contained a brand name even when the prompt named none (October 2026). Researchers at the University of Toronto found the same pattern. When ChatGPT already knows a brand, it writes that brand into its searches 67% of the time, against 11% for brands it doesn't know (Chen and Koudas, October 2026).
The model picks likely candidates first, then searches to check them. Two things follow.
- Being known matters before any search runs. Brands discussed widely across the web make the shortlist.
- Your own pages have to confirm what the model expects. Clear pricing, documentation, comparison and feature pages are what a site: search lands on.
Your Google ranking is only part of the story
If AI engines simply summarized Google, ranking would be enough. They don't.
When researchers compared the pages ChatGPT cited with Google's top 10 for the question exactly as the user asked it, only 0.1% to 0.6% matched. Even against the searches ChatGPT itself ran, 97% of the pages it retrieved sat outside Google's top 50 (Chen and Koudas, October 2026).
Google's own AI sits closer to Google search, but not by much. Only 38% of AI Overview citations rank in Google's top 10 (Ahrefs, March 2026). Measured against the top 10 results of five independent web search tools, the share was 18% for ChatGPT citations and 36.9% for Google AI Mode citations (Parse, October 2026).
The reason is those rewritten searches. Of the cited pages that ranked in Google's top 20 for any search, 32.9% ranked only for ChatGPT's rewritten searches, not for the original question. And 95% of those rewritten searches had zero volume in keyword tools (AirOps, March 2026).
Keyword research still shows you demand. It doesn't show you the searches AI engines actually run, and those are the searches you need to rank for.
Where AI citations actually land
Across 201,849 citations from 7,430 websites, most links went to company sites. Yet a brand's own website received just 4.9% of them. Competitors and other businesses took the rest of the company share.
Where AI citations land, by type of website
Company websites took 65% of all citations, but a brand's own site took 4.9%.
Company websites include brand, competitor and vendor sites.
Source: THREADSLAB tracking in Peec AI, logged out, 24 July to 22 October 2026.
See the numbers
| Share of all 201,849 citations | |
|---|---|
| Company websites | 65.1% |
| User generated platforms | 13.8% |
| Editorial and news | 8.2% |
| Reference | 5.3% |
| Government and institutions | 4.6% |
| Other | 3.1% |
Here is what the split shows.
No single platform owns AI answers. The most cited domain in our data held just 3.8% of all citations. The top 100 domains together held about half.
Each engine rewards different formats. Google's AI Overviews cited YouTube in 2,617 answers. ChatGPT pulled YouTube into 203 answers and cited it once. Facebook, Instagram and LinkedIn followed the same pattern: Google's AI cites them, while ChatGPT reads them and moves on.
| Platform | Google AI Overviews: answers citing it | ChatGPT: answers that pulled it in | ChatGPT: answers citing it |
|---|---|---|---|
| YouTube | 2,617 | 203 | 1 |
| 1,229 | 156 | 0 | |
| 924 | 91 | 0 | |
| Medium | 671 | 80 | 1 |
| 663 | 231 | 4 | |
| GitHub | 226 | 829 | 252 |
| Trustpilot | 154 | 212 | 126 |
| Capterra | 101 | 104 | 32 |
| G2 | 81 | 167 | 100 |
| Wikipedia | 44 | 140 | 3 |
ChatGPT isn't closed to outside platforms. It is selective. Review and developer platforms made it through: it cited G2 in 100 of the 167 answers where it pulled the site in, Trustpilot in 126 of 212 and GitHub in 252 of 829.
Each engine needs its own plan. For Google's AI, that means video and social profiles. For ChatGPT, it means review sites, documentation and your own pages.
What moves citations, and what doesn't
Google's May 2026 guide sums up the official position: optimizing for AI search is still SEO (Google Search Central).
| Tactic | What the 2026 evidence says | Verdict |
|---|---|---|
| Getting your brand discussed across the web | Earned media, meaning coverage a brand doesn't own or pay for, made up 84% of citations across ChatGPT, Claude and Gemini (Muck Rack, May 2026). Websites the model already knew had 1.6 times the odds of being cited (Chen and Koudas). | Works |
| Titles and headings that match real questions | Pages whose headings closely matched the query were cited 41% of the time, against 29% for weak matches (AirOps, April 2026). Matching the question was also the only page feature that held up when researchers compared pages on the same site (Chen and Koudas). | Works |
| Answering early on the page | 44.2% of ChatGPT citations came from the first 30% of a page (Kevin Indig, February 2026). | Likely helps |
| A visible last updated date | Cited pages were 1.72 times as likely to show one (Suff Digital, September 2026). | Likely helps |
| Letting AI search crawlers in | OpenAI runs a search crawler that is separate from its training crawler, GPTBot. Blocking the search crawler removes you from ChatGPT search answers, while blocking GPTBot only opts you out of training. PerplexityBot, which feeds Perplexity's answers, is not used for training at all (OpenAI; Perplexity). | Required |
| Schema markup | Google says it isn't required for AI features, and it showed no lift when comparing pages on the same site (Google; Chen and Koudas). | Not a lever |
| llms.txt | Google says Google Search ignores it. In a study of server logs on 137,000 sites, 97% of llms.txt files got no requests at all, and AI search bots made about 1% of the requests the other files received (Google; Ahrefs, June 2026). | No evidence it helps |
| Adding statistics and quotes to pages | These tactics came from a 2023 lab study. A September 2026 retest found no effect across ten current engine families (arXiv, September 2026). | Expired |
How to measure AI visibility without fooling yourself
Every answer is a fresh draw. Ask the same question repeatedly and the same brand list comes back less than 1% of the time (SparkToro via Search Engine Journal, January 2026). Two ChatGPT answers to the same question, about two days apart, share only a third of their cited websites (Parse, October 2026). Data from Graphite suggests about 10 runs per prompt give a visibility estimate with an average error of about 5.6 points (April 2026).
Most ChatGPT chats carry no citations. Commercial questions usually trigger a search. Only 6.8% of all US ChatGPT conversations carried citations in May 2026 (Similarweb, 2026). On the prompts we track, two thirds of them buyer questions, ChatGPT searched the web in 76% of answers over the last 30 days. Track the questions your buyers actually ask.
Being named beats being footnoted. Inside ChatGPT, the first recommended brand gets a 10.3% to 14.6% click through rate, against about 2.3% for footnote citations (First Page Sage benchmarks, via Similarweb, October 2026).
Your analytics can undercount. For about a week in October, a bug stripped the tracking tags from ChatGPT links. Across about 500 pages, ChatGPT visits fell by roughly 31,000 a day while direct traffic rose by about 26,000 (Search Engine Roundtable, October 2026).
Use the free reports from the engines themselves. Bing Webmaster Tools now shows Copilot citations and the searches behind them (Bing, February 2026). Google Search Console reports impressions in AI Overviews and AI Mode (Google, August 2026).
A simple protocol we use:
- Fix a set of 20 to 50 real buyer questions and keep it stable.
- Run each question several times per engine, logged out, and record rates, not single results.
- Track mentions and citations separately. A brand can be named without a link.
- Report every percentage with what it is out of.
- Ask new customers where they first heard of you. Many buyers influenced by AI never click a link.
How we gathered this data
- Brands: 10 brands we track for clients and prospects. Five are B2B companies, four of them in software, and five are consumer brands. Names are withheld.
- Prompts: 507 prompts, each run many times while its brand was tracked. About two thirds are buyer questions (best options, alternatives, pricing, reviews). The rest are informational.
- Answers: 28,308 between 24 July and 22 October 2026. ChatGPT 9,561, Google AI Overviews 9,200, Gemini 6,765, Perplexity 2,782.
- Coverage: two brands were tracked for most of the period and supply just over half the answers. The other eight ran for one to three weeks each. Perplexity was tracked for four brands until mid September. Claude, Copilot and Google AI Mode were not tracked.
- Setup: logged out, mostly from a US location, tracked with Peec AI.
- Website types: based on Peec AI's classification, with brand, competitor and vendor sites grouped as company sites.
- Timing: ChatGPT changed how it searches in August. Figures on how many searches it runs, and what they contain, use 5 to 18 October. Comparisons over time use the one brand tracked throughout, whose search data starts in early July.
Frequently asked questions
Do I need to rank on Google to get cited by ChatGPT?
It helps less than most people think. Fewer than 1% of the pages ChatGPT cited were in Google's top 10 for the question as asked. ChatGPT runs its own searches, so pages that match those searches, from brands it already knows, tend to win.
Do llms.txt or schema markup help?
Neither showed a benefit in the 2026 evidence. Google says Google Search ignores llms.txt and that schema isn't required for its AI features. The time is better spent on clear pages and on being discussed across the web.
Why does ChatGPT cite different sources every time?
Each answer is a fresh sample from a wide pool of pages. The same brand list repeats less than 1% of the time, so measure rates across many runs instead of trusting one result.
Which engine should I focus on first?
ChatGPT has 1.2 billion weekly users (The Verge, September 2026), so most brands start there. Google's AI Overviews reach people on Google itself and cite a wider mix of sources, including video and social profiles. Plan for both.
The brands that win in AI answers are the ones the engines can find, check and repeat.
Sources
- Jiang et al., Understanding ChatGPT citations, evidence from production web search across model generations, arXiv, October 2026.
- AirOps with Kevin Indig, The fan out effect, what happens between a query and a citation, April 2026.
- RESONEO via Search Engine Land, Inside ChatGPT's retrieval stack, the index, cache and pages it actually reads, August 2026.
- David Konitzny via Peec AI, GPT 5.6 retrieval changes, more iterations, double the sources, August 2026.
- Graphite, Parametric Knowledge Optimization, October 2026.
- Chen and Koudas, University of Toronto, What drives AI citations? Content, retrieval and model knowledge in generative engine optimization, arXiv, October 2026.
- Ahrefs, Update: 38% of AI Overview citations pull from the top 10, March 2026.
- Parse, AI citations vs search results, 18% overlap, October 2026.
- AirOps, The influence of retrieval, fan out and Google SERPs on ChatGPT citations, March 2026.
- Google Search Central, Optimizing your website for generative AI features on Google Search, May 2026.
- Muck Rack, Earned media still drives 84% of AI citations, May 2026.
- Kevin Indig, Growth Memo, The science of how AI pays attention, February 2026.
- Suff Digital, What AI cited web pages have in common, September 2026.
- OpenAI, Overview of OpenAI crawlers, 2026.
- Perplexity, Perplexity crawlers, 2026.
- Ahrefs, We analyzed 137K sites, 97% of llms.txt files never get read, June 2026.
- Bajemon and Rochet, Scoring without the engine, validating a deterministic, manipulation resistant content score for generative engines, arXiv, September 2026.
- SparkToro via Search Engine Journal, AI recommendations change with nearly every query, January 2026.
- Parse, AI citation volatility, why one missing source is not a trend, October 2026.
- Graphite, Demystifying randomness in AI, April 2026.
- Similarweb, AI search stats 2026, market share, referral and citation data, 2026.
- Similarweb, AI search news, October 2026.
- Search Engine Roundtable, OpenAI drops UTM parameters from ChatGPT results, October 2026.
- Bing Webmaster Blog, Introducing AI Performance in Bing Webmaster Tools public preview, February 2026.
- Google Search Console Help, Generative AI performance report for Search, August 2026.
- The Verge, ChatGPT now has 1.2 billion weekly users, September 2026.
- THREADSLAB tracking in Peec AI, 10 brands, 24 July to 22 October 2026, logged out.