--- title: "Sources, citations, and factuality" description: "Web research factuality distinguishes discovery results, sources, citations, grounded claims, and fact-checks, and defines verification of material claims against evidence." lang: en status: public-preview lastUpdated: 2026-08-31 url: https://dev.duale.ai/en/docs/web/sources-and-factuality --- ## AI-generated summary Defines source, citation, grounded claim, and fact-check terms, distinguishes the executor evidence package from the terminal SDK result, and gives an acceptance test for material claims. - A citation references a source but does not prove the source supports the nearby claim. - The executor returns a structured evidence package; the terminal SDK result follows your response model. - Accept a claim only when accessible evidence supports its full wording, scope, and dates. - Multiple URLs do not prove independent support; two publishers can repeat the same underlying report. Summaries were generated by AI. Generative AI is experimental. --- Here, **factuality** means whether each claim in an answer matches the relevant facts. Use sources to make an answer easier to inspect, not to claim certainty. A web research workflow has several separate steps: discover candidate pages, read selected pages, connect evidence to claims, and review the conclusions. Web access, grounding, and fact-checking are not the same function. | Term | Meaning in Web research | Do not infer | | ---------------- | ------------------------------------------------------------------------- | -------------------------------------------------------------------------- | | Discovery result | A page matched a search query | The page was read, supports the claim, or is authoritative | | Source | The research selected a page or document as evidence for a claim | The evidence is reliable, complete, current, or applicable | | Citation | An answer refers to a source, usually with a URL | The source supports the nearby claim or the answer represents it correctly | | Grounded claim | The research developed or checked a claim against fetched source material | The material or claim is true, current, complete, or independent | | Fact-check | A process compares a precise claim with evidence and handles conflicts | A guarantee that the conclusion or later decision is correct | A result can contain citations without grounded claims. More than one URL also does not prove independent support: two publishers can repeat the same underlying report. A fact-check can remain unresolved when evidence conflicts or is incomplete. ## Know what a typed result checks Inside the task, the dedicated Web research executor returns a structured evidence package to the coordinating agent. It includes claims, sources seen, claim-linked citations with excerpts, and gaps. Its response contract checks that a citation refers to a known source record that the executor marked as read. The terminal SDK result does not expose that package as a separate platform field. It follows your requested response model, so request the source fields your application must retain. These structural checks do not compare an excerpt with the page-read ledger, establish source independence, or prove the claim. Open and check every material source. ## Verify a material claim Use your own acceptance test for a material claim: 1. State the exact claim, date, and scope. 2. Open the final URL and locate the material that supports the whole claim. 3. Check the publisher, evidence origin, publication date, and effective date. 4. For corroboration, confirm that another source has independent evidence instead of repeating the first source. 5. Record conflicts and gaps, then apply the review that your decision requires. Accept the claim only when the accessible evidence supports its full wording and scope, its dates match the decision, and your required review is complete. Otherwise, qualify the claim, report the gap, omit it, or send it to a qualified reviewer. ## Account for freshness Search can use cached results and upstream indexes. A page can change after the task. The SDK does not return a public content snapshot or cache age that your application can use as evidence. Ask for publication and effective dates, then check them at the source and state the answer cutoff. ## Keep other proofs separate A [content mark](https://dev.duale.ai/en/docs/content-marking.md) can show that software produced a terminal text answer. It does not identify the web pages used or prove factuality. [Evidence and audit](https://dev.duale.ai/en/docs/security/evidence-and-audit.md) explains what a task trace proves and what evidence your application must keep. [Scope and limits](https://dev.duale.ai/en/docs/scope-and-limits.md) owns the review requirements for high-impact use. ## Related content - [Request Web research with source links](https://dev.duale.ai/en/docs/web/request-web-research.md) - [How Web search works: delegated research and evidence](https://dev.duale.ai/en/docs/web/how-web-search-works.md) - [Page access for the Web research executor](https://dev.duale.ai/en/docs/web/page-access.md) - [Troubleshoot Web search when results are missing or blocked](https://dev.duale.ai/en/docs/web/troubleshooting.md) - [Web search: delegated, source-led research for agent tasks](https://dev.duale.ai/en/docs/web.md) - [Technical documentation for developers](https://dev.duale.ai/en/docs.md) --- ## Sitemap See the full [Markdown sitemap](https://dev.duale.ai/sitemap.md) for all pages.