An AI Agent Read an Old Web Page: How to Spot the Gap

Check the dates of the evidence, not just the date the agent opened the page. A fresh fetch can return an old announcement. A new summary can repeat a superseded figure. The right source depends on whether you need the current position or an accurate account of an earlier one, and an agent that reports “retrieved today” has told you about the request, not about the fact.

When two primary pages disagree, record the precise claim and the version each supports before choosing. A visible date can help explain a conflict, but “newer page wins” is too crude for documents that describe different periods or conditions. This guide gives the agent a six-row worksheet for dating a claim, a fixed order for resolving conflicts, a worked example of a documentation change, the ways stale evidence still gets through, and a rule for when a dependent conclusion needs refreshing rather than the link under it.

Research agents are built to retrieve, and retrieval has a timestamp that is easy to log and easy to show. So the agent reports it: “as of today”, “retrieved this morning”. The phrase is true and almost meaningless. The page it retrieved may be an announcement from two years ago that still ranks well, a tutorial copied from an older version of the documentation, or a report whose data were gathered long before its publication date.

The failure is not that the agent fetched something old. Old evidence is often exactly right. The failure is that the agent let the fetch date stand in for the dates that matter, and the reader inherited a confidence the evidence never had. The remedy is to make the agent ask, for every claim that depends on timing, which of several dates it actually knows and which it is assuming.

Use this worksheet for an assertion where timing changes the answer. You do not need a timestamp for every sentence. Record unknown dates as unknown; a guessed date is worse than a visible gap, because a gap prompts a check and a guess prevents one.

The right-hand column is the misreading each date invites. They are the mistakes an agent makes by default when it has only one date and treats it as all of them. The worksheet works by forcing the agent to say which row each date it found actually belongs to.

RFC 9111’s HTTP caching standard uses freshness to govern reuse of stored responses. Its Expires definition does not imply that the underlying resource changes at that moment. Our editorial conclusion is separate: successful retrieval is not proof that a claim is still applicable. The protocol’s notion of “fresh” is about whether a cache may serve a copy without asking the origin; it says nothing about whether the origin’s content describes the present.

This matters because agents often surface cache-related signals as if they were evidence about facts. A response that arrived with no caching headers, or one the agent fetched directly from the origin, is just as capable of describing conditions from three years ago. The transport layer can tell you the copy is recent; only the document can tell you what period it describes, and only if someone reads it for that.

This article is about interpreting retrieved evidence. It does not diagnose a particular cache layer or prescribe headers. If the agent reads stale internal records, the data read-path guide addresses a different source of the problem.

First state the user’s time-specific question. “What was announced?” and “What is available now?” may legitimately require different answers from the same pair of sources. Then identify each page’s product, population, region or version scope. Apparent conflicts can disappear when those conditions are separated, and an agent that skips this step will spend its effort adjudicating a disagreement that does not exist.

Next look for an explicit correction, replacement notice, revision history or effective date in the source documents. Prefer the source responsible for the disputed fact within the relevant scope, while retaining the older document when explaining what changed. Do not let several copied summaries outvote a corrected primary; a dozen pages repeating the old figure are one source with twelve mirrors.

If two applicable primary pages still disagree, show the conflict and its practical consequence. For a decision that depends on the disputed detail, leave that detail unresolved or seek authoritative clarification. Avoid averaging incompatible values or silently selecting the more convenient one.

Consider a hypothetical documentation change: an announcement describes an upcoming limit, a later technical page documents the active limit, and a tutorial still repeats the announcement. A question about today’s implementation should use the applicable technical documentation and its version. A question about the announcement should preserve what the announcement actually said. Three pages, two correct answers, one wrong one, and which is which depends entirely on the question.

A study has another timing pattern. A recently released report may analyze observations from an earlier year. The report can be the newest available source while its data remain historical. State both dates when that distinction affects the recommendation: “a report published this year, on data gathered two years earlier” is a different piece of evidence from “this year’s data”, and an agent that writes the second when it means the first has invented currency.

If a secondary summary adds no newer evidence, describe it as a summary. The source-independence reference helps trace those relationships without confusing distribution with an update.

Even an agent that fills in the worksheet can be misled, and the three common routes are worth knowing because each looks like diligence from the outside.

The confident undated page: a vendor page with no visible date is read as current because nothing contradicts that reading; the worksheet row says “unknown”, and “unknown” gets silently rounded to “now”. The majority of mirrors: several search results agree, so the agent reports consensus, when all six copied the same superseded announcement and the one corrected primary sits on page two. The report-year shortcut: a study dated this year is cited as evidence about this year, and the methodology section that puts the data collection two years earlier was never opened.

The defense in each case is the same habit: when a date is missing, write “unknown” into the claim itself rather than into a footnote, and let the reader see the gap. An honest “as of an undated vendor page” is weaker than a dated claim and far stronger than a false one.

Choose refresh timing according to the decision. A fast-changing access condition deserves another check near the action; a historical definition may remain useful for much longer. There is no universal number of days after which every source becomes unreliable, and a policy that re-fetches every link monthly will confirm that the links resolve while the claims drift.

The unit of refresh is the conclusion that depends on the evidence. If a recommendation rests on a limit, re-check the limit before the recommendation is acted on, and re-read the recommendation in light of what you find. For edits to an existing article, follow the content fact-preservation guide, which traces a changed fact into the summaries and advice built on it. Keep the current uncertainty near the claim, rather than hiding it behind a general “last updated” label. Teams that want evidence dating built into their research agents rather than left to the prompt can work with our AI transformation practice.