How AI chooses its sources: Wikipedia, Reddit, LinkedIn, press and directories

Every generative answer is a montage of sources. Understanding which sources models prefer, and why, is the most useful part of GEO: it says where to act. The data below come from 2025-2026 studies on millions of citations, mostly in English; the proportions change by language, the mechanism does not.

The three channels an answer comes from

A model builds what it says about a company from three channels.

Training memory. Everything the model read when it was built. Repetition weighs: a fact present in Wikipedia, in ten articles and in a hundred discussions is "true" for the model, even if old. It changes only with new training runs.

Browsing. When the question requires it, the model searches the web and reads a few pages. ChatGPT and Copilot use the Bing index, Gemini the Google one, Perplexity its own. Here freshness, structure and clarity of pages count, and the fact that crawlers can read them.

Entities. Knowledge Graph, Wikidata, structured listings. They serve to identify the company and attach facts to it. An incomplete entity produces confusion and namesakes.

What they cite, in numbers

Citation studies show a long-tail distribution: a few heavily cited sources and thousands of sites cited a little.

• Wikipedia is the single most cited source by ChatGPT, with a 12-13% share of citations in the US and an almost constant presence in the top positions.
• Reddit has a similar share on ChatGPT and rises sharply on Perplexity, where user discussions are among the preferred sources for experience questions ("what is it like", "is it worth it").
• LinkedIn and YouTube appear consistently, especially for questions about people and companies.
• The press weighs for dated facts and controversies: an article from a news outlet remains a source even years later.
• Category directories dominate "which supplier" questions: for professional services, a few directories collect most citations across all engines.
• The rest, over 90%, is spread across thousands of domains, including the company's website, which however is rarely the decisive source about itself.

The differences between engines

ChatGPT with browsing cites three or four sources per answer on average and favours explanatory pages, with definitions and data. Perplexity cites much more and more often communities and recent articles. Gemini in AI Overviews draws on the Knowledge Graph and on pages with structured data. Claude browses less and relies on memory: it changes more slowly. Copilot follows Bing, with a greater weight on LinkedIn.

For a company this means the same source does not have the same value everywhere: a Reddit discussion moves Perplexity, a Wikidata entry moves Gemini, a news article moves all of them, but on different timelines.

What makes a passage citable

Regardless of the engine, picked-up passages share common traits: a fact with number, date and source; an attributed quotation; a self-contained definition; a high position on the page. Pages written to persuade, without verifiable facts, are read but rarely cited.

What it means for you

A company's source map almost always shows three things: the decisive sources are not the ones the company looks after; recent strengths exist in no third-party source; one or two dated sources weigh more than everything else. The plan starts from there: complete the entity, create or update the third-party sources the relevant audience makes models read, make new facts citable.

The GEO Assessment produces exactly this map, with URL, date and weight for every source. The free audit shows its first entries in 48 hours.