How ChatGPT, Claude, and Perplexity Decide What to Cite

Ask ChatGPT, Claude, or Perplexity a question and you'll often get an answer with sources attached — or no sources at all, just a confident paragraph. Understanding why a model cites one page and ignores another is the core of Generative Engine Optimization.

Two different mechanisms

Not all "AI answers" work the same way under the hood:

  • Retrieval-augmented answers (Perplexity, ChatGPT Search, Google AI Overviews) actively search the web at query time, pull back a handful of pages, and generate an answer grounded in that retrieved content. Citations here map to specific pages fetched during that search.
  • Parametric answers (a base model like Claude or GPT answering from training data alone, no live search) draw on whatever the model absorbed during training. There's no live citation — the "source" is whatever content was prominent and clear enough to be memorized accurately.

Both mechanisms reward the same underlying qualities, for different reasons: retrieval systems need to find and extract your content quickly; base models need to have learned it correctly in the first place.

What retrieval-based engines look for

When an engine searches live, it's essentially running a fast relevance + extractability pass over search results:

  1. Topical match — does the page directly answer the question asked, not just mention the topic in passing?
  2. Extractable structure — can a specific paragraph or sentence be lifted as a self-contained answer? Pages buried in narrative prose with no clear takeaway are harder to quote.
  3. Freshness signals — a visible date, especially for topics where accuracy changes over time.
  4. Trust signals — domain authority still matters, but so does internal consistency: does the page contradict itself or other authoritative sources?

What shapes parametric (training-time) answers

For a model recalling facts from training rather than searching live, the dynamics are different:

  • Repetition and consistency across the web. A claim your own site makes once is weaker than the same claim appearing consistently across your site, your docs, and third-party mentions.
  • Clarity, not cleverness. Ambiguous or heavily marketing-toned language compresses badly during training. Plain, factual statements survive better.
  • Structure that survives chunking. Training pipelines break pages into chunks. A paragraph that depends on context from three paragraphs earlier may get chunked away from that context and lose its meaning.

The practical takeaway

You can't control which mechanism will answer a given user's question, so the safest strategy optimizes for both at once: write clear, dated, self-contained sections; repeat your key claims consistently instead of stating them once; and structure comparisons and definitions so a single paragraph could be lifted verbatim and still make sense.

That's precisely what GEO Score in GEO is first tests for — it sends your actual target prompts to Claude, ChatGPT, Perplexity, and Gemini and shows you, per engine, whether you were cited, where, and what was actually said.