FactRelay Docs
How AI Answers Discover and Use Sources
The pipeline from a buyer's question to an assembled answer — query rewriting, retrieval, synthesis, and citation — and where brands gain or lose accuracy at each stage.
To fix wrong answers, you need a working model of how answers form. Simplified but faithful, the pipeline has four stages — and a brand can win or lose at each one.
1. Query interpretation and rewriting
Assistants rarely search your buyer’s literal words. ChatGPT Search, for example, may rewrite queries, and can use approximate location and conversation memory. Two users asking the “same” question can trigger different retrievals.
Implication: you cannot optimize for one magic phrase. You optimize for a question space — which is why FactRelay freezes a prompt set spanning the decision questions that matter, and treats the natural panel and the citation-diagnostic panel as separate experiments.
2. Retrieval
The system pulls candidate sources: your pages, competitors’ pages, comparison posts, documentation, forums, directories. Retrieval favors what is crawlable, extractable, and topically matched. If your data-residency facts live only in a PDF or a JavaScript-rendered widget, the 2024 blog post that misstates them may be the best source available.
Implication: technical readiness is answer accuracy. This stage is where crawl readiness work pays off.
3. Synthesis
The model composes an answer from retrieved material plus its trained knowledge. Three characteristic failure modes hit brands here:
- Stale facts win when old sources outnumber or outrank your current page;
- Opinions generalize — one 2023 forum thread becomes “users report that…”;
- Boundaries blur — “SOC 2 certified” loses its scope, plans, and dates.
Implication: publish facts in synthesis-resistant form: atomic claims with scope and dates (“EU data residency (Frankfurt), Business plans, since 2026-03”), stated boundaries, and consistent wording across your properties. This is the job of the fact registry.
4. Citation
Some surfaces attach sources. Citations tell you which pages the system considered load-bearing — and they are auditable. A citation can be current, outdated, or valid-but-misused (a real opinion presented as a general fact).
Implication: the citation layer is where diagnosis happens. FactRelay archives every response with its citations as an immutable snapshot, then traces each load-bearing claim to its source and its validity (evidence standard).
The honest caveat
This pipeline varies by platform and changes without notice. That is precisely why conclusions must carry their measurement surface and date, and why one-time audits decay. See the measurement standard for how we keep observations comparable over time.