Skip to content
Interface language: English
Audit / Signal

A fresh page, an old story: checking dates in AI research

Separate publication, event and access dates before using AI research to make a decision.

Article language: English

Illustration: part of a stone column, with a thin line marking a crack.
In this article

A page published yesterday may describe an offer that ended last year. A page from three years ago may still contain the current instructions. Sorting by newest helps you find material, but it does not decide what belongs in a recommendation made today. The date beside a search result is not a stamp saying “still available”.

When researching with AI, separate the page's publication date, the date of the event it describes and the time you checked it yourself. If the information applies for a period, record that separately. Only the combination of these details and the source's content supports a conclusion about the state needed for a decision. A recent page update tells you much less on its own.

Four dates behind an apparently simple answer

In a fictional assignment, an agent looks for a meeting-room booking tool. The team needs guest accounts in the basic plan because some facilitators attend only individual workshops. The assistant finds a recent product review and writes: “Guests can be invited without buying an additional plan.” That sounds useful. Without dates and scope, however, it may describe terms the team can no longer obtain.

Imagine this sequence: the feature was announced in March, trial availability ended in June, and an October article recounts the product's development. Assume we consult it for a decision at the end of October. All those dates can be correct. The mistake would be to assign the article's date to the feature's availability and overlook the trial's end. This example is invented to illustrate the reasoning; it does not assess a real provider.

First record the precise claim: “A guest can make a booking under the basic plan available to a new account on the selection date.” Existing accounts raise a different question; a limited trial raises another. If the agent does not know those conditions, it should not replace them with the broad statement “the product supports guests”.

Give each date a meaning. “Published” refers to the page. “From June” may mark the start of a new condition. “Applies to version 4” ties the information to a release. “Checked today” describes your own action. The date of access does not make material newer. It records when particular content was read, subject to the limits of that reading.

If a source gives only a month, do not invent a day. If you read it on a particular day but did not retain the hour, an honest date is enough. Precision to the second can be useful when real and necessary. An invented minute merely makes uncertainty look more elegant. For events near midnight, record the time zone or offset so that different calendar days do not appear to contradict one another.

For information valid over a period, you need both boundaries or an explicit statement that one is missing. An announcement that “the trial starts in March” does not itself give an end date. Do not guess it. You can record the historical start and leave current availability to be confirmed. That is more useful than a confident statement about today's offer built on an old announcement.

What a date in a tool actually means

Google explains that a date in a result is an estimate of when the page was published or updated, based on several signals. It also distinguishes the date of a page from the dates of events described on it. A visible date cannot therefore automatically be assigned to every claim in the text.

In research, start with the page, not just the search-result snippet. Find the passage intended to support the claim and read its surroundings. Words such as “previously”, “until the end of the trial” or “for existing customers” may change the answer more than the date in the heading. If the material is retrospective, label it as an account of the past even when published only hours ago.

There is also a real example of separating events. The history of RFC 9110 in the IETF registry shows a publication entry on 6 June 2022 and metadata changes on 20 May 2026. A later action described as a metadata change is not, by itself, an announcement of a new edition of the standard. Read what kind of event took place.

In product documentation, a later date may accompany a new section, an editorial correction or a change to another platform's description. A timestamp alone will not tell you whether the condition you care about changed. If a change history is available, compare the relevant passage. If it is not, record that absence rather than inventing what was supposedly updated.

Keep the address of the version actually read. A link to “latest” documentation may later point to another release. If a recommendation is prepared after a break, revisit the decisive condition and compare it with the retained evidence. In the report, distinguish the original reading from a later check. The same page name does not establish identical content.

Technical markers have their own scope. RFC 9110 describes Last-Modified as the time when the server believes a representation of the resource was last changed. A representation is the form of material sent in a response. This marker does not explain which claim changed or when an offer began to apply. It is a technical signal, not an assessment of substantive currency.

You need not inspect network headers for every brief investigation. It is useful, however, to know what a report from a tool using them does not establish. If an agent calls a source “current”, ask what that means: the page date, an observed change to a condition, or confirmation in current documentation? Those answers support different levels of confidence.

Anthropic's search documentation describes `page_age` as information about when a site was updated. It lists citation information separately. Neither field is defined there as confirmation that a feature remains available in a particular product plan. That conclusion requires the source's content and a properly scoped question.

Aeneia conceptual photography

A report can retain signals provided by a tool, but should not call them your own reading of the page. If the agent received only a snippet or excerpt, record that scope. A blocked page or a page requiring access remains unread in the relevant part. A known title and working address do not fill that gap.

From an announcement to a condition for today's decision

For the fictional booking tool, look for the current plan description and guest-account documentation. The announcement remains useful for establishing the history but cannot replace current terms. Also look for the end of the trial and changes in availability. A query combining the feature name with “withdrawal” may answer a different question from another review of the product's benefits.

Do not give priority solely to the newest text. Official documentation for the current version may be older than a new partner article. The source's role, its scope and its fit to the question matter. If two materials give different terms, use the guide to comparing conflicting sources (in Polish): first establish whether they concern the same users and the same time.

Recognise an announcement too. “The feature will arrive in November” does not mean it is available on an account in late October. An announcement of launch may concern a gradual rollout or a specified group. If the condition determines the choice of tool, confirmation must match the intended use, and sometimes a controlled trial in the plan the company intends to choose is needed.

With Aeneia Audit, it may be possible to discuss reviewing several decisive AI-research claims, from the question and query through the passage actually read to the final recommendation. A report should show where currency was confirmed and where a condition remains to be checked before acting.

Your own record need not be elaborate. Beside a claim, record the exact source passage or its location, the meaning of its dates, the time of reading and your conclusion. Separate the author's statement from your interpretation. “The documentation lists guest accounts for the team plan” is different from “the team plan will be a good choice for our studio”. The second statement also needs your own criteria.

In the booking example, one claim card might contain the old announcement and the result “historical trial confirmed”. A second concerns the current plan and records the absence of an unambiguous statement about new accounts. A third describes the team's requirement: a facilitator without a company account must be able to book a room and receive confirmation. This division lets you commission one decisive check rather than asking for another general comparison of tools.

If the current page mentions “guests”, read the definition of that role. It may mean a person invited to a particular event, not someone who can book a room independently. Matching words and dates still do not confirm the desired action. Preserve the distinction in the record and describe any trial through the facilitator's action and the expected result.

An agent should also distinguish lack of confirmation from confirmation of absence. Inaccessible documentation does not prove that a feature was removed. An explicit announcement ending a trial says more, but may still not resolve the offer for another group. State the evidence precisely: information was found about the end of a particular programme; no basis for current availability was found; or the specified condition was confirmed. Each statement has a different scope.

A scientific paper or report raises a further question about the identity of the material. A matching date will not repair a title, author and identifier combined by a model from different papers. The guide to checking bibliographies and DOIs (in Polish) examines that layer. Only then can you ask whether the paper found matches the period and scope of the analysis.

Stop the claim, not the whole investigation

If current guest-account terms cannot be confirmed, the agent can still prepare a comparison of the feature's history. It must simply label the result in line with the evidence. It need not reject all the research or fill the missing field with the most likely answer. “Availability for a new account was not confirmed” accurately describes the evidence.

The next step depends on the importance of the gap. If the feature is essential, the choice remains open until it is confirmed. If it is an extra, you can compare an option without it and explain the consequence. Agree that boundary before the agent starts. Otherwise the model may turn an uncertain detail into an apparently certain benefit to complete a comparison table.

Keep the old finding when a later change appears. Today's documentation does not prove what was available on the date of an earlier decision. To explain the past, use retained material and its actual date. The article on a company search system based on documents (in Polish) addresses managing internal versions; here we are checking what could be established on the open web.

If research is regular, tie the next reading to the kind of information. A change in product plans, an announcement of withdrawal or an approaching purchase date may justify updating a key claim. There is no need to reread an entire library merely because another day has passed. Nor may an old check be treated as confirmation today. The guide to market monitoring (in Polish) also calls for clear rules for interpreting a new signal.

The final review can be checked without judging the style of the whole report. Give another person one claim, its source and the recorded dates. They should be able to identify what the information concerns, the time for which it is confirmed and what still needs to be established. A longer explanation from a model cannot replace missing evidence. The article on AI explanations and independent evidence (in Polish) explores that distinction.

In our example, the decision is not to choose a plan on the basis of an October account of a March trial. First confirm whether a guest can book in the intended plan. The fresh page helped identify the question. The answer required more than a fresh date.

Aeneia / Aeneia editorial

Editorial article. Source links are included in the text.

About Aeneia

Brands we have worked with

  • L’Oréal
  • Garnier
  • Netflix
  • Carrefour
  • Warner Bros. Discovery
  • Biedronka
  • HBO Max
  • Wolt
  • Żabka
  • COSRX
  • Embassy of China in Poland
  • Mixa
Aeneia / Initium

From knowledge to a decision

Tell us what you want to put in order. We will agree on where to start.

Contact Aeneia

We partner with and use solutions from:

  • Google
  • OpenAI
  • Microsoft
  • Stripe
  • Anthropic
  • Linear