AI systems choose sources in two steps: first they retrieve a shortlist of pages that match the question, then they pick the passages that best support the answer they are writing. A page gets cited when it survives both steps. To survive retrieval, it has to be crawlable, indexed by the system the engine relies on and relevant to the question. To survive selection, it has to state something specific and verifiable, agree with other credible sources and come from a site or author the wider web already treats as reliable.
No engine publishes a full formula. Google, OpenAI and Perplexity document the access rules (which crawlers to allow, which pages are eligible), and independent studies fill in the rest. Those studies point the same way: ranking well in search still helps, especially in Google AI Overviews, but brand mentions across the web, clear and quotable facts, and freshness on time-sensitive topics decide a large share of which sources make the final answer.
This guide covers the selection mechanics. To build the credibility signals themselves, see our companion guide on E-E-A-T for AI search.
What happens between a question and a cited source?
Every AI answer with citations is built by a retrieval system feeding a language model. The model does not browse the web the way a person does. It receives a small set of documents or passages, writes an answer and attaches links to the passages it relied on. The pipeline looks broadly similar across engines, even though each one implements it differently.
| Stage | What the system does | What decides the outcome | What you can influence |
|---|---|---|---|
| 1. Interpret the question | Rewrites the prompt into one or more search queries | Intent, context, follow-up questions | Covering the subtopics people actually ask about |
| 2. Fan out | Runs several related searches across subtopics | How the engine splits the question | Pages that answer each subtopic directly |
| 3. Retrieve | Pulls candidate pages from a search index | Crawl access, indexing, relevance | Robots.txt rules, indexability, server-rendered text |
| 4. Rerank and filter | Scores candidates for relevance and reliability | Authority, agreement with other sources, recency | Reputation, third-party mentions, update cadence |
| 5. Synthesize | Writes the answer from selected passages | Which passages state the needed fact most clearly | Answer-first writing, data, named sources |
| 6. Attribute | Links claims to supporting pages | Which page the model drew each claim from | Distinct, quotable statements on your page |
Google confirms the fan-out step for its own products. Its Search Central documentation on AI features (updated 10 December 2025) says AI Overviews and AI Mode may issue multiple related searches across subtopics and data sources, and that this lets them show a wider and more diverse set of supporting links than a classic results page. A page can therefore be cited for a subquestion it answers well, even if it never ranks for the original prompt.
Do ChatGPT, Google and Perplexity pull from the same sources?
No. Each engine retrieves from a different index, so the same question can produce different sources on each platform. Knowing where each engine looks tells you where you need to be visible.
| Engine | Main retrieval source | Crawler to allow | What is documented |
|---|---|---|---|
| Google AI Overviews and AI Mode | Google’s own search index | Googlebot | A page must be indexed and eligible to show with a snippet; no special markup or AI files are required |
| ChatGPT search | Third-party search providers plus OpenAI’s own crawling | OAI-SearchBot | Any public site can appear; blocking OAI-SearchBot keeps pages out of summaries and snippets |
| Perplexity | Perplexity’s own index and live page fetches | PerplexityBot | PerplexityBot surfaces and links sites in results; Perplexity recommends allowing it |
| Gemini app | Google Search index for grounding | Googlebot (Google-Extended controls Gemini grounding and training) | Google-Extended does not affect inclusion or ranking in Google Search |
OpenAI does not name every search provider it uses, but an independent test gives a strong hint. Seer Interactive (6 February 2025) checked more than 500 citations from about 100 questions and found that 87% of SearchGPT citations matched Bing’s top organic results, compared with 56% for Google. Seer calls the sample limited, so treat it as directional. The practical takeaway still holds: if your site is weak or missing in Bing, ChatGPT has fewer chances to find it.
OpenAI’s publisher FAQ also separates search from training. OAI-SearchBot controls whether content can appear in ChatGPT search answers, while GPTBot controls whether content may be used to train models. Blocking GPTBot does not remove you from ChatGPT search; blocking OAI-SearchBot does.
How much does ranking in Google predict an AI citation?
It depends heavily on the engine. For Google’s own AI features, ranking is a strong predictor. For standalone assistants, it is a weak one.
- Google AI Overviews. Ahrefs analyzed 1.9 million citations from 1 million AI Overviews (published 21 July 2025) and found that 76.1% of cited pages ranked in Google’s top 10 for the query, while 14.4% did not rank in the top 100 at all.
- Other assistants. In a separate Ahrefs study of 15,000 long-tail prompts (11 August 2025), only about 12% of URLs cited by ChatGPT, Gemini, Copilot and Perplexity ranked in Google’s top 10 for the same prompt. Perplexity was closest to Google at 28.6%; the others averaged around 8%.
- Independent research. A 2026 academic preprint by Xu, Iqbal and Montgomery (arXiv 2605.14021, not yet peer reviewed) measured 55,393 queries and found that nearly 30% of domains cited in AI Overviews did not appear on the first page of results shown next to them. The authors conclude that AI Overviews use a source selection process that is partly separate from ranking.
Two lessons follow. First, classic SEO is still the cheapest route into Google’s AI answers, because Google’s retrieval starts from its own index. Second, ranking alone will not get you into ChatGPT or Perplexity answers. Those engines run their own queries, often phrased differently from the user’s prompt, and select from a wider set of pages. Our guide to GEO vs SEO explains how to split work between the two.
Which signals make a source look trustworthy to an AI system?
Five signal groups show up again and again in documentation and research: access, authority, consensus, freshness and extractability. None is a published ranking factor on its own, but together they explain most of the patterns studies have measured.
| Signal group | What the engine is checking | Evidence | Action |
|---|---|---|---|
| Access | Can the crawler fetch and read the text? | Vercel (17 December 2024): GPTBot, ClaudeBot and PerplexityBot fetch JavaScript files but do not execute them | Serve key content as HTML, allow the right crawlers |
| Authority | Is the site or author known for this topic? | Ahrefs: Domain Rating correlates 0.326 with AI Overview brand mentions | Topical depth, named experts, earned links |
| Consensus | Do other credible sources say the same thing about you? | Ahrefs: branded web mentions correlate 0.664, the strongest factor measured | Coverage in trade media, directories, reviews, comparisons |
| Freshness | Is the information current for a time-sensitive question? | Engines with live retrieval fetch pages at answer time | Visible update dates, real revisions, current figures |
| Extractability | Can a clear claim be lifted from the page? | GEO study (KDD 2024): quotations, statistics and citations raised visibility by roughly 30% to 40% | Answer-first sections, data with sources, short definitions |
Access is the gate. A page that is blocked, not indexed or rendered only with client-side JavaScript never reaches the later stages, no matter how good it is. Googlebot does render JavaScript, so such a site can appear in Google’s AI features yet stay invisible to other assistants.
Why does agreement across the web matter so much?
Language models are designed to give answers that are well supported, and the simplest test of support is whether several independent sources say the same thing. A claim that only appears on your own site is a weak candidate. A claim repeated by a trade publication, a review platform and a comparison article is a strong one.
The clearest data point comes from Ahrefs’ study of 75,000 brands (published 26 May 2025). It compared site metrics with how often each brand was mentioned in Google AI Overviews. Branded web mentions had a Spearman correlation of 0.664, branded anchors 0.527 and branded search volume 0.392. Classic link metrics were weaker: Domain Rating 0.326, referring domains 0.295 and backlinks 0.218. Ahrefs notes that correlation is not causation and that the sample was filtered to established sites with Domain Rating above 40, so the numbers describe relative strength, not a formula.
Seer Interactive found a weaker link for brand search volume in its own ChatGPT dataset (correlation of 0.18, published 16 April 2025), which is a useful reminder that different engines and datasets give different results. What both studies agree on is that the wider web’s view of you matters, not only your own pages.
Consensus work means consistent facts everywhere: the same name, category, locations and key claims on your site, profiles, listings and the articles that mention you. Our entity SEO guide covers how to make that consistency machine-readable.
How does freshness affect which sources get picked?
Freshness matters most when the question has a time component: prices, regulations, software versions, rankings, “best of this year” lists and news. Engines with live retrieval, such as Perplexity and ChatGPT search, fetch pages at the moment of the question, so an outdated page competes directly with current ones.
No engine publishes a freshness weight, and we found no independent study precise enough to quote. What matters is that updates are real:
- Change the substance: new figures, new examples, removed outdated advice.
- Show a visible “last updated” date and keep it honest. Changing the date without changing the page is not an update.
- Prioritize pages that answer time-sensitive questions. An evergreen definition needs fewer updates than a pricing comparison.
What content changes make a page easier to cite?
The best evidence here is the GEO paper by Aggarwal and colleagues from Princeton University and partners, accepted at KDD 2024. The researchers tested several ways of rewriting source content and measured how visible each source became in generative answers across a benchmark of queries.
- Adding quotations from relevant sources, adding statistics and citing sources were the strongest methods, with the paper reporting gains of 30% to 40% on its main visibility metric.
- Keyword stuffing performed worse than doing nothing.
- The gains were largest for sources that started lower in the result list. Sources ranked fifth more than doubled their visibility with the citation method, while top-ranked sources lost some share.
For smaller sites this is encouraging: a precise, sourced statement on a less prominent page can beat a vague one on a famous site. In practice:
- Answer first. Put the direct answer in the first one or two sentences of each section.
- Use numbers with attribution. “Ahrefs found X in a study of Y, published Z” is far easier to cite than “studies show”.
- Use question-style headings. They map naturally to the subqueries an engine runs during fan-out.
- Add tables for comparisons. Structured rows give the model clean facts to work from.
These are the same habits we apply on on-page SEO work, with more weight on clarity at the passage level.
Does being cited mean being represented accurately?
Not always. The Xu, Iqbal and Montgomery preprint mentioned above extracted 98,020 individual claims from AI Overview answers and found that 11.0% were not supported by the pages they cited. Omission, where the answer leaves out a qualifier or condition stated on the source page, was the most common failure. The researchers also found that source quality and claim accuracy were largely independent: a credible source can still be summarized badly.
So monitor what answers say, not only whether your link appears, and state important limits, prices and exceptions in their own sentence, close to the claim they qualify, where they are harder to drop.
How can you see which sources AI trusts in your market?
Run the questions your buyers ask and record which sources come back. This takes a few hours and tells you more than any generic checklist.
- List 20 to 40 real questions from sales calls, support tickets and search data.
- Run each one in ChatGPT, Perplexity, Gemini and Google (for AI Overviews and AI Mode).
- Record every cited domain, and note whether your brand is mentioned, cited, both or neither.
- Check whether your pages are blocked for any crawler, not indexed in Bing, or dependent on JavaScript.
- Track AI referrals. ChatGPT adds utm_source=chatgpt.com to links, and Google reports AI Overviews and AI Mode clicks inside the Search Console “Web” search type.
The pattern of cited domains is your map. If an engine keeps citing two trade publications and a review platform, that is where your consensus signals need to come from. Our guide to tracking SEO results covers the measurement setup.
Who should work with Marketer Coffee on AI citations?
Marketer Coffee is an AI SEO and GEO agency that helps brands become a source AI engines retrieve, trust and cite. Our evidence comes from other industries, and we say so plainly:
- Planika (design fireplace brand). Seven US-localized articles took a site Google had not indexed to 1,052 monthly US visits. Ahrefs recorded 401 Google AI Overview mentions, 136 Perplexity mentions and 20 ChatGPT mentions. Read the Planika case study.
- Landingi (SaaS). Roughly 5,000 sessions a month referred by AI assistants, tracked in GA4. Read the Landingi case study.
We fit companies that sell across several markets or languages, have a real expertise story to tell and want to measure AI visibility alongside revenue. We work in English, Polish, German, French, Spanish, Italian, Swedish and Danish.
Who is not a fit: a business that needs only a Google Business Profile tune-up in one city, or anyone looking for guaranteed citations. No honest agency can promise a specific AI answer.
Start with our free AI visibility audit, which shows how engines describe your brand today and which sources they cite instead of you. Details of the ongoing work are on our AI SEO service page.
Frequently asked questions about how AI chooses sources
How does ChatGPT decide which websites to cite?
ChatGPT search runs its own queries through third-party search providers and OpenAI’s crawling, then cites the pages it used for the answer. OpenAI does not publish ranking criteria. Your site must allow OAI-SearchBot, and an independent Seer Interactive test found most citations matched Bing’s top results, so Bing visibility helps.
Do I need to rank on page one of Google to be cited by AI?
For Google AI Overviews it helps a lot: Ahrefs found 76.1% of cited pages ranked in the top 10. For ChatGPT, Gemini, Copilot and Perplexity, only about 12% of cited URLs ranked in Google’s top 10 for the same prompt, so other signals matter more.
Do backlinks still matter for AI citations?
Yes, but less than brand mentions in the data available. In Ahrefs’ study of 75,000 brands, branded web mentions correlated 0.664 with AI Overview mentions, while backlinks correlated 0.218. Links still help pages rank, which feeds Google’s AI features.
Do I need special schema or an llms.txt file to be cited?
Google says you do not need new machine-readable files, AI text files or special markup to appear in AI Overviews or AI Mode. Structured data that matches visible content remains good practice, but it is not a shortcut to citations.
Why does AI cite sources that are not on the first page of results?
Engines split a question into several related searches and select passages that support specific parts of the answer. A page can be the best source for one subquestion without ranking for the original prompt.
What should you remember about how AI picks trusted sources?
AI engines cite pages they can reach, find relevant and confirm elsewhere. Access is the gate, ranking is a strong advantage in Google and a weak one elsewhere, and the wider web’s agreement about you carries more weight than any single on-page tweak. Clear, sourced, current statements are what get lifted into answers.
If you want to know which sources AI engines trust in your market and what it would take to join them, send us your site and your key questions. Book a free consultation.
Sources: Google Search Central, “AI features and your website” (updated 10 December 2025); Google Search Central, Google common crawlers documentation (Google-Extended); OpenAI Help Center, Publishers and developers FAQ; Perplexity, PerplexityBot documentation; Seer Interactive, “87% of SearchGPT citations match Bing’s top results” (6 February 2025); Seer Interactive, brand awareness and LLM visibility (16 April 2025); Ahrefs, AI Overview citations and Google rankings, 1.9 million citations (21 July 2025); Ahrefs, AI assistant citations and Google top 10, 15,000 prompts (11 August 2025); Ahrefs, AI Overview brand visibility factors, 75,000 brands (26 May 2025); Xu, Iqbal and Montgomery, “Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact”, arXiv 2605.14021 (preprint, 2026); Aggarwal et al., “GEO: Generative Engine Optimization”, KDD 2024, arXiv 2311.09735; Vercel, “The rise of the AI crawler” (17 December 2024); Marketer Coffee case studies for Planika and Landingi (other industries).
