Retrieval-augmented generation
Also called: RAG
Retrieval-augmented generation (RAG) is an AI architecture where a model looks up a specific source before answering and grounds its response in what it actually found, instead of generating an answer from patterns alone.
Retrieval-augmented generation is a way of building an AI system so it looks something up before it answers, instead of relying only on patterns baked in during training. The model fetches a specific document or page, reads what's actually there, and grounds its response in that content. Ask the same model the same question with no retrieval step and it does something different: it pattern-completes, generating the most statistically plausible answer for a question like that one, whether or not it's true.
Why this matters for GTM research agents specifically
Ask a research agent for a company's latest funding round with no grounding and it will usually hand one back: a dollar figure, a round letter, sometimes an investor name, stated with the same flat confidence as a founding date. None of it has to be true for the agent to say it that way, because without a retrieval step, the model is filling a gap with the most plausible-sounding completion, not reporting a fact it found. A properly grounded call looks different: the agent fetches a specific page, finds a specific sentence, and reports what that sentence says.
The two modes read identically on the page. Same confident tone, same punctuation. Nothing in the raw output tells a reader which one they're looking at, which is what makes the failure mode dangerous in research used for outbound personalization. Grounding has to be enforced in the prompt itself, not assumed just because the agent has search or browsing tools available. Having the capability to retrieve doesn't mean every claim in the output actually used it.
In practice
The practical version of RAG discipline in a research prompt: name which source types count as evidence, require the agent to produce the literal sentence supporting any claim before it counts, and require a second, independent source check before a claim is treated as confirmed. A claim without a retrievable quote behind it should come back labeled unverifiable, not filled in with a hedge like 'likely' or 'probably around.'
What people get wrong
Teams assume that because a model can browse the web, it's automatically doing retrieval-augmented generation on every claim it makes. It isn't. An agent with browsing access can still pattern-complete a fact it never actually looked up, and blend that fabricated claim into the same sentence as a properly retrieved one. Retrieval capability and retrieval discipline are two different things, and only the second one is enforced by a prompt, not by the tool the agent has access to.
