How AI Answers Pick Their Sources

Retrieval, chunking, grounding and corroboration — what is publicly observable about source selection, and what is still speculation.

Nobody outside the labs knows the full recipe for how an answer engine chooses its sources, and any article claiming otherwise is guessing. What can be done honestly is narrower and more useful: separate the mechanics that are publicly observable from the folklore, and work out what each implies for the way you publish. The definitions this builds on are in What Is Answer Engine Optimization (AEO)?, and the contrast with classic ranking is in AEO vs SEO: What Changes and What Stays.

Two stages: retrieval first, generation second

Almost every answer engine works the same broad way, and the pipeline is observable in outline. A question triggers retrieval: the engine searches — usually through an index it operates or partners for — and pulls back a set of candidate documents. Those documents are split into chunks, the chunks are handed to a language model, and the model writes an answer grounded in them, with citations attached to the material it actually used.

The order matters. Generation can only quote what retrieval surfaced, so if your page is never fetched at answer time, nothing else about it matters — no amount of on-page polish repairs a retrieval failure. This is the practical difference between the two stages: retrieval is won with crawlable, well-structured, findable pages; grounding is won with passages the model can lift and believe.

Chunking: why self-contained passages win

Retrieval rarely keeps whole pages. It keeps passages — and a passage retrieved without its surrounding context has to carry its meaning alone. Sentences that depend on the paragraph above them get lifted incomplete and quietly discarded; sections that restate the question, answer it in the first sentence, and add one specific supporting fact survive the cut and, later, the quote.

This is why question-shaped headings and comparison tables keep appearing in every practical AEO guide, and why FAQ blocks fit so naturally: each question-and-answer pair is a pre-made chunk. Whether the markup itself earns its place is a separate question, taken up in FAQ Schema for AI Answers.

Trust: what the model believes about you

Domain-level versus page-level signals

Page-level quality decides which chunks get retrieved. Domain-level trust decides which domains keep showing up. Engines visibly favor sources that are authoritative for the topic at hand — official documentation for product questions, established publications for industry questions — which is why a new site can rank for a query and still lose the citation to a source the model trusts more on that subject.

Corroboration

The second mechanism is corroboration. A claim that appears on your site and nowhere else has one vote; the same claim repeated across independent sources — reviews, directories, press coverage, documentation — accumulates weight, and that weight shows up in what models assert about you. This is the unglamorous half of AEO: making the facts about your brand agree everywhere they appear, because contradiction teaches a model to hedge. We turn that into a working method in ChatGPT Recommendation Optimization: A Practical Guide.

Freshness, caching and answer drift

Answers change less often than the web does, because most engines cache. A given answer can persist for days and then change wholesale after a model update or an index refresh. Two consequences follow. Date-stamp and genuinely update pages, because grounded answers prefer material that reads as current. And never credit your own work for a metric change without first checking whether the provider shipped something that week — pinning model versions and dates in your log, part of the routine in How to Track Your Brand Mentions in ChatGPT, Gemini and Perplexity, is what makes that distinction possible.

What we will not pretend to know

The honest list of unknowns: exact ranking weights, the composition of any engine's retrieval index, how often those indexes refresh, and the thresholds that decide who is in or out. Anyone publishing precise numbers is speculating — including people who sell tools for a living. The defensible posture is behavioral: freeze a prompt set, observe the answers over months, and treat every claim about the machinery, ours included, as something your own data can overrule. When your observations start needing competitive math, Measuring Share of Voice in AI Answers supplies the formulas.

Keep reading

Part of PZTools·AEOscope·PolicyForge·LaunchKit