How AI engines choose sources

How do AI engines choose which sources to use?

Search-enabled AI engines choose sources in two stages. First they search and gather candidate pages. Then they pick the passages that best answer the question and write a reply based on them. The pages they drew on are shown as citations.

This is true of ChatGPT when it searches, Perplexity, Google AI Overviews and AI Mode, and Gemini or Claude when web search is turned on. When web search is off, a model answers only from what it learned in training, and names no sources.

What is retrieval-augmented generation?

Retrieval-augmented generation, usually shortened to RAG, is the method behind most AI search. The model retrieves relevant documents first, then generates an answer using them. The approach was set out in a 2020 research paper that paired a language model with a searchable index of documents.

The point of RAG is to give the model fresh, specific information it may not have learned in training. It also gives the engine something to cite. For a business, it means the pages an engine can find at the moment of the question matter a great deal.

What does grounding mean in AI search?

Grounding means tying an answer to real sources found at the time. Google describes grounding with Google Search as connecting the Gemini model to real-time web content. It says this helps reduce made-up answers and lets the model show the sources for its claims.

A grounded answer is only as good as its sources. If the pages it finds are out of date or wrong, the answer will be too. That is why the sources matter as much as your own site.

What happens between the question and the answer?

The exact method differs by engine and is not fully public. The broad steps are similar across them.

1. The question is rewritten. OpenAI says ChatGPT search rewrites your question into one or more targeted queries. Google says AI Overviews and AI Mode may use query fan-out, running several related searches across subtopics.
2. Search results are gathered. Each query returns a set of pages from a search index.
3. Passages are picked. The engine reads the pages and picks the passages that best answer each part of the question.
4. The answer is written. The model writes a reply based on those passages, adding what it already knows.
5. Citations are attached. The pages it used are linked, often next to the sentence they support.

Two things follow. A page must rank for at least one of the searches the engine runs to be read at all. And within that page, a clear passage that answers the question is more likely to be picked than a vague one.

Why do listicles and "best of" articles get cited so often?

Because they answer "which one" questions in full on a single page. When a buyer asks for the best accountants in Stirling or the best CRM for a small charity, a list article compares several options with reasons. That is exactly the shape of answer the engine is trying to write.

These articles also tend to rank well for "best" searches, so the engine finds them early. For a business, this means being included in the main comparison articles for your field can matter as much as your own pages. It also means an out of date or unfair list can shape answers about you for a long time.

Why do directories get cited?

Directories hold structured facts about many businesses in one place. They list names, addresses, services, opening hours, ratings and sometimes prices. For local questions and trade questions, that makes them a convenient and trusted source.

Directories can also be a source of errors. An old address or a service you stopped offering can be repeated in AI answers for months. Keeping your main listings accurate is basic housekeeping for AI search, the same as it is for local SEO.

Why does Reddit get cited so often in AI answers?

Reddit threads contain first-hand experience, honest opinions and comparisons from real people. Many questions ask for exactly that, such as "is this worth it" or "has anyone used this firm". Reddit threads also rank well in search for many topics, so engines find them.

There is a commercial side too. In 2024 OpenAI announced a partnership with Reddit that gives it access to Reddit's Data API, to bring more Reddit content into ChatGPT. The Reddit and AI search guide covers how to take part honestly without breaking community rules.

Why do review sites get cited?

Review sites answer judgement questions. When someone asks who is reliable, good value or best rated, the engine looks for pages that judge businesses. Review platforms do that at scale, with many voices.

Detailed, recent reviews give the engine specific things to repeat, such as "finished on time" or "clear pricing". A thin profile with a few old reviews gives it little to say. Replies to reviews also help show that a business is active and listens.

When do AI engines cite a brand's own website?

AI engines cite a brand's own website most often for facts about that brand. Examples include prices, services, locations, opening hours, product details and policies. They also cite brand pages that give genuinely useful guides or explanations.

For "which one should I choose" questions, a brand's own site is cited less often, because engines look for independent views. So your site does two jobs. It should be the clearest, most current source of facts about you. And it should hold useful guides that answer buyers' wider questions.

Do AI engines prefer big, well-known websites?

Big sites have an advantage, but they do not win every time. Well-known sites tend to rank for more searches, so engines find them more often. They also carry more trust.

Smaller sites still get cited when they answer a specific question better than anyone else. A local firm's page that sets out prices, areas covered and lead times in plain words can beat a national site with a vague overview. The same goes for a specialist blog with first-hand detail. Specific beats general, especially for questions with a place, a budget or a niche in them.

What makes a page more likely to be chosen as a source?

A few qualities come up again and again. None of them is a trick.

• It can be found. The page is crawlable by the engine's search bot and ranks for the searches the engine runs.
• It answers directly. The answer sits in the first sentence under a heading that matches the question.
• It is specific. It gives names, numbers, dates and places rather than general claims.
• It shows its sources. The original GEO research found that adding quotations, statistics and cited sources raised visibility in AI answers.
• It is current. Dated pages with recent updates are safer to use for anything that changes.
• It agrees with other sources. Facts that match across your site, listings and reviews are easier to trust.

How can you find out which sources AI engines use for your topic?

The simplest way is to ask your buyers' questions and click the citations. Do this across several engines and several runs, and note the domains that keep appearing. Group them by type, such as directories, review sites, lists, forums, news, competitor pages and your own pages.

That list becomes your to-do list. It shows which sites shape answers in your market, and so where a correct listing, a fair review or an inclusion in a comparison article would make the most difference. The ChatGPT guide shows how to turn it into steps.

How Axiom GEO helps

AI citation tracking in Axiom GEO shows the top cited domains and pages across all your tracked answers. It sorts each one by type, such as directory, review site, listicle, social, your own pages, competitor pages and news. URL mentions shows which of your own URLs get cited. Reddit insights finds the subreddits AI engines cite in your topic, using answers you have already collected.