How AI engines pick sources

TL;DR — Answer engines choose sources based on a mix of pre-training exposure, real-time retrieval, structural clarity, and third-party corroboration. Nothing is fully public, but the working model is clear enough to build against.

The three moving parts

Every AI answer is shaped by three inputs:

1. What the model already knows — the pre-training corpus (the web as it was at the cut-off, plus books, forums, and code).
2. What the model just fetched — real-time retrieval, either via the engine's own search index (Google, Bing, Brave) or via an in-house crawler (OpenAI's SearchGPT, Perplexity's index, Claude's web tool).
3. How the model synthesises the two — a ranking/summarisation step that decides which brands to name, which URLs to cite, and how many.

The signals that seem to matter

Across our own analysis and the (small amount of) public writing from OpenAI, Perplexity, and Google, a consistent set of signals stands out:

• Named-brand recency. How often your brand name appears in fresh content the crawler has recently indexed.
• Third-party corroboration. Reddit threads, roundup articles, comparison posts, PR mentions — sources the engine considers "independent."
• Structural clarity. Question-shaped H2s, TL;DRs at the top, FAQ blocks, tables, clean HTML. AI engines quote from content that reads like an answer.
• Authoritative anchoring. Being cited (or linked to) from domains the engine already trusts.
• Freshness. Recently published or recently updated pages get pulled disproportionately.
• Answer completeness. Pages that fully answer a question in ~40–80 words (or one paragraph) show up in more citations than pages that require scrolling.

What we've stopped believing

• Backlink volume alone. Still helpful, but no longer the dominant signal.
• Exact-match keywords. LLMs are semantic; they don't need the phrase, they need the answer.
• Meta descriptions. Almost never quoted verbatim by AI engines.
• Long-form for its own sake. A 4,000-word article that doesn't quickly answer the question loses to a 600-word one that does.

What GeoNexo does with this

For every prompt we scan, we log which URLs each engine cited, which brands each engine named, and the source types (owned domain, third-party, forum, PDF, transcript). Over time this becomes a map of where each engine's attention actually lives for your category — and the content generation loop targets those specific formats and surfaces.