Dated narratives: why AI still cites the 2019 crisis and how to fix it

Almost every corporate assessment contains the same scene: the communications director reads a model's answer about their own company and finds, in the first line, a fact from five years ago. A closed lawsuit, a restructuring, a dismissed investigation, a withdrawn product. True, but over. The model tells it in the present tense.

How a dated narrative arises

It takes a single condition: an authoritative source that no more authoritative source has updated. The 2019 article from the news outlet still exists, is well indexed and well linked; the Wikipedia entry cites it; some discussions pick it up. Later, the company announced the closure of the lawsuit with a release on its own website, read by few and cited by nobody. For the model the balance of power is clear: ten third-party sources against one partisan source.

Then there is training memory: if the fact was present in many sources when the model was built, the model "knows" it, and without browsing repeats it even if today's web says otherwise.

Why a press release is not enough

Models weigh three things: the authority of the source, the consistency across sources and the date. A company release is a partisan, isolated source, often without a visible date in a readable form. To move a narrative you need a third-party source, consistent with others and more recent, or an update of the original source.

The five interventions that work

1. Update the primary source. If a Wikipedia entry exists, propose on the talk page, with sources, the update on the subsequent facts. On Wikidata, update dates and statuses directly. It is the intervention with the best cost/effect ratio.
2. Create a more recent, more authoritative source. An interview or an article in a trade publication that tells the current state, with dates. Browsing engines find it; outlets index it; Wikipedia can cite it.
3. Publish a "what has changed" page on the website: a timeline of the facts with dates and external sources. It is not as authoritative as a news outlet, but it is the page a browsing model reads to verify.
4. Reply where the narrative lives. If the fact is picked up in communities or reviews, a public, dated, documented reply changes what the model summarises.
5. Repeat the measurement. The 90-day re-analysis shows whether the fact still appears, in how many answers and with what context. In the cases we follow, the dated fact does not disappear: it appears in one answer in five instead of all, and is accompanied by the subsequent facts. That is the realistic result.

What does not work

Asking the model to "correct itself" in a chat has no effect beyond the conversation. Removing the original article is rarely possible and often counterproductive. Publishing lots of generic content on the website moves nothing, because it adds no third-party sources.

For investor relations and founders

Dated narratives weigh most where due diligence is fastest: an analyst who asks Gemini what risks a company presents receives a montage of what is available, with the closed lawsuit in the first line. The same applies to the question "who is [CEO]", where an old article or a namesake defines the person. The Investor and Founder assessments start from these questions.