How Do AI Engines Choose Their Sources? The Complete Guide

Why does a generative AI cite Wikipedia rather than your website, or Trustpilot rather than a blog post? Understanding how ChatGPT, Gemini, or Perplexity choose their sources is the cornerstone of any effective GEO (Generative Engine Optimization) strategy: if you know what makes a source "citable," you know where to focus your efforts to increase your visibility in AI responses.

The 4 main criteria AI engines use to select sources

Whether it is RAG (Retrieval-Augmented Generation, used by Perplexity or Google AI Overview) or a response generated from the model's training completed with a web search (ChatGPT, Gemini), AI engines broadly apply the same four families of criteria when deciding which sources to cite.

1. Perceived authority of the source

An established domain that is widely cited elsewhere on the web and recognized in its field is far more likely to be picked up than a recent, poorly referenced site. This authority is built over time: link profile, press mentions, presence on recognized review platforms (Trustpilot, G2, Capterra depending on the industry).

2. Cross-source consistency

Generative AI engines try to limit the risk of citing false or isolated information. A data point confirmed by several independent sources (a figure repeated by a media outlet, a forum, and a product page, for example) is judged more reliable than a claim that appears in only one place. This is one of the reasons why a multi-platform presence (owned site, social media, customer reviews, press mentions) carries more weight than a single, however well-optimized, website.

3. Content freshness

Perplexity and Google AI Overview, both heavily oriented toward real-time search, clearly favor recent or regularly updated content, especially on fast-moving topics (pricing, news, tool comparisons). ChatGPT and Claude, more dependent on their training data, are less sensitive to immediate freshness but still take it into account when they run a web search to complete their answer.

4. Technical structure of the content

Well-structured content (Schema.org markup, a clear heading hierarchy, direct answers to questions, data presented as tables or lists) is easier for an LLM to extract and summarize than dense, unstructured text. A robots.txt file that blocks AI crawlers simply eliminates any chance of being cited, regardless of content quality.

How each AI engine chooses its sources

These four criteria are not weighted the same way across AI engines:

• ChatGPT (OpenAI): combines its training with a web search (via Bing and its own crawls) when the question requires it. Gives significant weight to community platforms (Reddit above all) and customer reviews.
• Google AI Overview and AI Mode: rely directly on Google's search index, which makes their citation logic close to classic SEO (domain authority, structure, Google Business Profile for local queries).
• Perplexity: the most freshness- and real-time-search-oriented of the four, with heavy reliance on instant web search results rather than training data alone.
• Claude and Gemini: combine training and web search depending on context, with particular attention to the coherence and editorial quality of cited sources.
• Mistral: a French engine whose coverage of French-language sources (press, government sites, French B2B platforms) is particularly relevant for brands targeting the French market.

What our study of 1.2 million AI responses reveals

To back these mechanics with data rather than intuition, BotRank analyzed more than 1.2 million responses generated by ChatGPT, Gemini, Perplexity, Google AI Overview, Mistral, Claude, and Copilot for its clients (see the full study: Top 100 LLM Sources). Three findings directly confirm the criteria described above:

• User-generated content dominates by a wide margin: Reddit (11.47% of Top 100 citations) and YouTube (9.87%) take the top two spots, ahead of Wikipedia (9.43%). LLMs give a clear premium to authentic firsthand experience over polished brand messaging.
• Review platforms carry real weight: Trustpilot ranks 4th (7.96%), followed by B2B-specialized players such as Appvizer (8th), Capterra (51st), and G2 (44th). For a query like "what's the best tool for...," the AI almost systematically consults review platforms to build its recommendation.
• General and specialized press remains a safe bet: French dailies Le Figaro (6th) and Le Monde (7th) rank in the Top 10, confirming that press relations (Digital PR) and earning a citation or backlink from a recognized media outlet remain strong trust signals for the models.

How to audit and improve your cited-source profile

Knowing the general selection mechanics is not enough: you also need to know, specifically for your brand, which sources actually influence AI responses on your target queries. That is exactly what BotRank's Source Analysis feature does, automatically identifying the domains AI engines cite when discussing your brand or industry.

In practice, three concrete actions let you act on your cited-source profile:

1. Invest in customer reviews on the platforms that matter for your industry (Trustpilot for B2C, G2 and Capterra for a software vendor, Trustfolio for an agency or B2B service provider).
2. Be present and active on Reddit and industry-specific forums, adding real value rather than self-promoting, which these communities heavily penalize.
3. Technically structure your own site (Schema.org markup, direct answers to questions, a robots.txt file that allows AI crawlers) to remain a credible primary source about your own brand, alongside third-party sources.

Frequently asked questions