Retrieval and grounding
Retrieval obtains potentially relevant information. Grounding uses external material to inform a generated answer. Neither term means every claim is correct or fully supported.
How the concept fits together
- Step 1Question
- Step 2Search and retrieve candidates
- Step 3Use selected source material
- Step 4Synthesize an answer
What it means for GEO
Publish verifiable details and inspect whether the final answer preserves your conditions and limitations. Record whether the answer used the public web, files or connected sources.
What not to infer
A retrieved page may never be cited. A citation can point to material that does not support the associated claim.
How it works
Retrieval finds candidate evidence; grounding uses external evidence in generation. Candidates may be selected, excluded or revisited through additional searches. The resulting answer can mix source-backed claims with unsupported conclusions, and a retrieved document does not necessarily become a displayed citation. Grounding should therefore be evaluated at the level of the claim and the material supporting it, not by counting links alone. The ALCE research benchmark treats correctness and citation quality as separate evaluation dimensions.
GEO example
Illustrative scenario, not observed assistant behavior or measured results.
Illustrative scenario: an assistant says an audit tool “checks every page and guarantees accessibility compliance,” citing a guide that explains sampled pages and automated checks. The topic is relevant, but the cited document does not establish either the exhaustive coverage or the guarantee. Split the response into claims and review each against the actual passage. Publish the sampling and automated-check limitations prominently so accurate evidence exists, without implying that a page update will compel an assistant to use it.
What you can do
- Write specific, verifiable claims and attach material limitations to the relevant statement.
- Keep original evidence, documentation and update dates easy for readers to inspect.
- When reviewing an answer, distinguish a relevant source from a source that actually establishes the claim.
What to measure, and what it cannot prove
Retain the answer text, source URL, relevant passage and review outcome. Evaluate unsupported claims, outdated evidence and missing qualifications separately. In your own retrieval system you can also inspect fetched candidates; a public citation list does not expose every document the assistant retrieved.
Evidence and authoritative sources
- OpenAI: Searching the web with ChatGPT
Describes source-linked search responses and warns that search results and citations can be wrong.
- Gao et al.: Enabling Large Language Models to Generate Text with Citations (EMNLP 2023)
Introduces a citation-evaluation benchmark separating correctness and citation quality. Its historical results are not a current ranking of assistants.
The sources establish the documented concepts and product behavior. The publishing checklists, scenarios and operational measurement rules are GeckoCheck educational interpretations, not provider-endorsed ranking instructions.
Source: OpenAI: Searching the web with ChatGPT · Gao et al.: Enabling Large Language Models to Generate Text with Citations (EMNLP 2023)
Period: Reviewed October 12, 2026. Living provider documentation; cited research published in 2023 and 2024 where applicable.
Sample: Conceptual explanation and illustrative scenario. No measured observations.
Limitation: GeckoCheck educational interpretation. Examples are fictional; formulas use declared operational definitions, not universal provider metrics. Source documentation does not guarantee citation or ranking outcomes.