Freshness and Provenance Signals

A retrieved result without a timestamp or source is a claim an agent cannot verify. Freshness and provenance are what let an agent decide whether to trust a result, cite it, or check it against something more current.

FreshnessCopy Anchor Link

Every retrievable record should carry:

• updated_at (or equivalent last-modified timestamp), in ISO 8601, on the record itself, so it reaches every response that returns the record.
• A documented staleness policy where relevant: how often the underlying data refreshes, and whether a retrieved result can be out of date by design (a cached price, a nightly-synced inventory count).
• A way to ask for current data explicitly when the endpoint's default response may be cached - a ?fresh=true parameter, a documented cache TTL, or a separate real-time endpoint for volatile facts like price or availability.

An agent using a stale price or availability figure to make a purchasing decision is a correctness failure. Treat volatile decision facts (price, stock, appointment slots) differently from stable reference facts (a product's dimensions, a policy's text).

ProvenanceCopy Anchor Link

Every retrieved result should be traceable back to where it came from:

• A stable identifier and canonical URL for the source document or record, rather than the retrieval system's internal ID.
• A citation-ready reference: section, page, or anchor when the source supports it, so an agent quoting the result can point to the exact origin.
• Source attribution when a result is derived or aggregated - a summary, a computed value, or a merged record should say what it was built from, not present itself as a primary source.

How this differs from page-level freshness signalsCopy Anchor Link

Content Structure covers article:modified_time and sitemap lastmod as page-level signals for crawlers and agents. Those help a crawler decide what to recrawl and let an agent judge a page's age. Freshness and provenance on a retrieval response are a different contract: they let the calling agent decide, at the moment it reads a result, whether that specific record is current enough and attributable enough to act on.