Grounding
Grounding is the practice of tying an AI model's output to retrieved source material rather than to its training memory, so each claim can be traced back to a document. In commerce, a grounded answer about a product is one where the price, spec, and availability came from a live source at answer time instead of from whatever the model absorbed months earlier.
Grounded versus ungrounded answers
An ungrounded answer is generated from the model's parameters. It reflects whatever the model saw in training, has no timestamp, and cannot be checked. An answer about a discontinued part number, generated ungrounded, will be confidently wrong about its availability and possibly about its specifications.
A grounded answer runs a retrieval step first, pulls in current documents, and writes from those. The distinction shows up visibly as citations, though a citation is evidence of grounding rather than a guarantee of accuracy — a model can still misread a source it correctly retrieved.
Why this matters more for products than for most content
Product facts change on someone else's schedule. Manufacturers supersede parts, revise pack quantities, reissue cut sheets with corrected values, and discontinue SKUs without notice. Price and stock change daily.
So product answers are only useful when grounded in something current. That is the reason the agentic commerce stack exists at all: a product feed, a UCP catalog endpoint, or an MCP server are all mechanisms for giving a model a fresh, authoritative source instead of letting it answer from memory. Google uses the same word for the same idea: its Google-Extended control governs both training and "grounding in Gemini Apps and Vertex AI."
Being the thing an answer is grounded in
Three conditions decide whether your record is usable as grounding material.
It has to be reachable. Server-rendered HTML, an accessible feed, or an endpoint. A spec that only exists after JavaScript executes is not available to most retrieval crawlers.
It has to be unambiguous. Two records for the same physical part with different values force whichever one retrieval happened to pick. A golden record is not a data-governance nicety here; it is what keeps a grounded answer from being grounded in the wrong row.
It has to be verifiable. Certifications naming their standard number and issuing body, spec values with units, linked source documents joined to the SKU. Claims a model can trace back are the ones it repeats.
Grounding does not mean correct
Worth stating plainly, because the vocabulary implies more safety than it delivers. Grounding constrains a model to sources; it does not verify those sources. If your catalog says a valve is rated to 600 WOG and the manufacturer revised it to 400, a perfectly grounded answer repeats your error with a citation attached.
Which puts the accuracy burden back where it started. In enrichment work this is the fill rate versus accuracy trade: a populated field with a wrong value is worse than a null, because a null gets skipped and a wrong value gets quoted.
Frequently asked questions
What does grounding mean in AI?
Constraining a model's answer to retrieved source material rather than letting it generate from training memory alone. Grounded answers can be traced to documents, which is what makes citations possible and what keeps answers current rather than frozen at the training cutoff.
Is grounding the same as RAG?
Closely related. RAG is the architecture, retrieval feeding generation, and grounding is the property that architecture is meant to produce. You can also ground a model through tool calls or a live API rather than a vector index, which is what commerce protocols do.
Does a citation prove an answer is grounded?
It indicates retrieval happened, but it does not prove the answer accurately reflects the source, nor that the source is correct. A model can cite a page and still misstate what the page says, and it will faithfully repeat an error that exists in your catalog.
How do we make our product data usable for grounding?
Make it reachable, unambiguous, and verifiable: server-rendered rather than client-injected, one authoritative record per physical part, spec values as typed fields with units, and certifications and documents joined to the SKU so claims can be checked.