How answer engines choose their sources

When ChatGPT, Perplexity, Gemini or Google's AI Overviews answer a question, they cite a small number of web pages. Those citations are not decorative. They are the output of a selection process, and the process is documented well enough to describe. This page covers the mechanics: how content is retrieved, how sources are picked, what qualifies a page to be considered at all, and why measuring your own citations is harder than it looks.

Retrieval-augmented generation

Answer engines do not answer from model memory alone. They use retrieval-augmented generation (RAG): at question time, the system retrieves relevant documents from an index, then generates a response grounded in what it retrieved. The pages it drew on become the citations. This is why AEO is possible at all. The model's training data is fixed, but the retrieval step runs against a live index that your content can enter and improve its standing in.

Grounded in ranking systems

For Google specifically, Google's AI features guide states that AI Overviews and AI Mode select content using retrieval-augmented generation grounded in the core Search ranking systems. The retrieval step is not a separate, mysterious algorithm. The same systems that decide organic rankings decide what the generative layer gets to read. Content that ranks is content that can be cited.

Query fan-out

Google describes a technique called query fan-out: the system issues multiple related queries concurrently to gather context for a single question. The practical consequence is that your page can be retrieved for a query the user never typed. A question about pricing might fan out into queries about features, alternatives and setup. Pages that cover the adjacent subquestions of a topic have more surfaces through which retrieval can find them.

How citations appear

Google's guide notes that AI responses show prominent clickable links to the supporting pages. Being selected as a source is therefore visible brand placement at the exact moment of the answer, and the path a reader follows when they want to verify or go deeper. Google's Search Console includes a performance report for generative AI, which is the first-party way to see this traffic.

The entry ticket: indexed and snippet-eligible

To be a candidate at all, a page must be crawlable, indexed and eligible to appear as a snippet. That is the whole entry requirement. Per Google's guide, no special markup or AI-specific files are needed, llms.txt is ignored by Google Search, and structured data is not required for generative AI features. Google explicitly advises against chunking content or writing in a special format for AI. If a page cannot appear as a normal snippet, it cannot appear in an AI answer; if it can, nothing further is technically required.

How non-Google engines differ

ChatGPT, Perplexity and Claude do not sit on Google's index. They run their own crawlers and retrieval systems, so a page's standing can differ across engines, and their appetite for fresh content differs measurably. A study of 7,683 pages and 47,097 citations found that Gemini cited content updated within the past year 78 percent of the time, ChatGPT 73 percent and Perplexity 65 percent. Only 42 percent of the cited content had been recently published, which means the freshness the engines reward comes mostly from updates to existing pages rather than from new ones. If Perplexity matters to you, an older but maintained page still competes; on Gemini, a stale page is a long shot.

Why citation measurement is noisy

Checking whether an engine cites you sounds simple: ask it and look. In practice the output is unstable. Across three identical ChatGPT runs, roughly 2.2 to 2.3 percent of cited sources remain stable. In measurements of LLM brand responses, the query language accounts for 26.5 to 32.0 percent of measurement variance, while brand identity accounts for 1.5 percent. How you ask moves the result far more than who you are.

Two conclusions follow. First, a single query check provides no reliable signal; only repeated sampling across phrasings and time shows a trend. Second, treat vendor dashboards with care: no third-party tool has access to Google's internal systems, so any external citation number is a sample, not a ledger. For what this means for your content strategy, see what is AEO.