The crisis is over, the newspapers have moved on — but anyone searching for your brand on ChatGPT or Perplexity still finds the worst version of the story. It's not slowness: what was written in the first 48 hours, often confused and incomplete, can remain the AI's official answer for months or years because the models keep drawing on those same sources. There is a way to circulate the correct facts and progressively displace that narrative — and the sooner you start, the less damage you accumulate.
The crisis is over, you handled the communication well, the press has already moved on. But ChatGPT still cites it as the top result on your brand. Because the corpus doesn’t forget.
Let me explain why this happens and, above all, what to do to prevent the first version of a crisis — often incomplete, written in a hurry, maybe even before you had the facts in hand — from becoming the “official” version that AI engines will repeat for months whenever someone searches for your name.
Why a well-managed crisis can linger in the AI corpus for a long time
In the classic media cycle, the outlets talk about it for 48 hours, then the topic recedes, then it dies out. That’s the world of the traditional search engine, where freshness overtakes older articles.
AI engines work differently. ChatGPT, Claude, Perplexity and Gemini don’t “forget” when a topic leaves the front page: what they processed during training stays in the model’s weights, and what they pull via RAG depends on indexes that also contain content months old.
Practical result: if the first published version about you during a crisis was wrong or partial, it stands a good chance of being the answer the AI engine gives when someone searches for your brand. Even six months later.
What the research says about emergency messaging
There is a recent study — Gonella et al., 2025 — that builds a dataset of around 18,000 crisis situations to train models to generate useful messages during real emergencies: floods, tornadoes, earthquakes, acts of violence. The authors also evaluate the models on out-of-distribution scenarios, that is, on crises never seen during training, and they test an automatic post-editor to correct the output before publication.
The paper’s context is civil emergency, not corporate communication. But from that work you can deduce a principle that applies one-to-one to brands as well: the messages published during a crisis become training and retrieval material for the models. If you, the brand, don’t publish your own structured and complete official version, the corpus will build its narrative on other content — reviews, forums, news articles written in the first two hours, when even the journalists didn’t have the full picture.
From the paper you can also draw a second relevant deduction. Gonella et al. use the FEMA guidelines (the US federal agency for emergency management) as a source of instructions because the model needs a structured, authoritative and verifiable reference to anchor its answers to. Translated to your brand: even the best AI engines respond well about you only if they find a structured and authoritative reference content to start from. The FEMA guidelines in their case. Your official statement, dated and signed, in yours.
The operational consequence is that crisis communication is no longer just an exercise in human reputation management. It is also an exercise in feeding the AI corpus with the correct version of the facts, published where the retrieval systems go to fetch.
It’s the worst mistake: you remove your version of the facts and leave only everyone else’s.
How it connects to the rest of the work on AI visibility
Everything I’ve told you in the previous articles — from content tokenization to the way AI recognizes an author as an entity through author entity recognition all the way to the weight of implicit citations in implicit reference weight — counts triple during a crisis. When the volume of citations on your brand explodes in 48 hours, the corpus recalibrates: if your newsroom is silent, it builds the version of the facts by reading whoever is talking about you.
Write in inverted pyramid: answer first, then facts, then context, never the other way around
The case of an aesthetic clinic chain in Bergamo
Let me tell you a concrete case, changing the name for obvious reasons. A chain of aesthetic medicine clinics with its main office in Bergamo and three branches in northern Italy had a reputational crisis: a former patient publicly denounced an unsatisfactory treatment outcome on social media, the story ended up in local outlets and then in two national outlets for about four days.
The clinic handled the human side well: it responded to the patient, it mediated, and the matter closed in civil court without any conviction. The press stopped talking about it after a week.
Six months later we stepped in. The owner asked me one precise thing: “when I type the clinic’s name into ChatGPT, it still talks about that story as if it were today”. We checked: on ChatGPT and Perplexity, the query “opinions on [clinic name] Bergamo” returned, in the first three sentences, a reference to the affair, citing as its source the news articles from the first 48 hours. None of the AI answers cited the clinic’s official statement, published on their site four days after the affair broke.
What we did in 60 days:
- Published on the site’s newsroom a timeline page reconstructing the facts with dates, documents, outcomes, and the medical director’s signature
- Reworded the existing statement into inverted pyramid format (who, what, outcome, clinical context).
- Added a complete Organization schema and a Person schema for the medical director, with a link to their Medical Board profile
- Asked two local outlets that had covered the affair for an editorial update on the outcome
Six months after the intervention, on a sample of 20 queries re-run on ChatGPT, Perplexity, Claude and Gemini, the reference to the crisis still appeared in 6 answers out of 20 — but in 5 of these the cited source was now the clinic’s official statement, no longer the articles from the first 48 hours. An indicative test on a single case, not a controlled study: the pattern, however, is consistent with what I see on the other clients to whom we apply the same scheme.
The mistakes I see most often in crisis communication
When an Italian company manages a crisis without thinking about AI, it tends to make one or more of these mistakes.
Silence in the first hours on its own newsroom. The official statement goes out by email to the agencies but isn’t published on the brand’s site as an indexable page. Result: the only version the AI corpus finds is the one written by the journalists.
Generic statement with no facts. Phrases like “the company categorically denies and reserves the right to act in the appropriate venues” give the AI engine no useful chunk to cite. The AI then cites whoever provided details, namely the news articles.
No update after the resolution. The crisis closes, perhaps with an outcome favorable to the brand, but this information isn’t published on the same page as the initial statement. The corpus stays frozen on the day-1 version.
Deleting the statement once the crisis has passed. It’s the worst mistake: you remove your version of the facts and leave only everyone else’s. The broken link doesn’t eliminate the information from the model’s training, it simply removes your voice from retrieval.
What to do operationally when a crisis hits the brand
The principle drawn from the work of Gonella et al., 2025 on the civil emergencies dataset is simple and transfers to the brand: a language model produces reliable answers about a crisis situation only if it has access to a structured and dated reference content. In the paper that function is performed by the FEMA guidelines used as an anchor to generate the answers; in your context the same function is performed by your official statement, if you publish it well and keep it online.
Here is the list of actions that works in practice:
- Publish within 24 hours a dedicated page on your newsroom with the established facts, the actions taken, the name and role of whoever signs it
- Write in inverted pyramid: answer first, then facts, then context, never the other way around
- Add an Organization schema and a Person schema for the spokesperson, verifiable with Google’s Rich Results Test
- Update the same page as the situation evolves, don’t create new pages that fragment the signal
- Keep the page published even after the resolution, adding a “final outcome” section with a date
An entry-level check: open ChatGPT and Perplexity, search for your brand with the query “what happened to [brand] in [crisis year]”. If the answer doesn’t cite your official statement among the first two sources, you have a corpus-feeding problem. The full analysis requires professional tools and months of work, but this first five-minute test tells you whether you’re in bad shape or whether you have room to maneuver.
The common thread: your version of the facts must be the version the AI retrieves
Coming back to the thread of this series on visibility in AI answers, the lesson of AI-aware crisis communication is consistent with everything else. AI engines don’t read the web the way we do: they build a representation of your brand starting from the most structured, citable and authoritative content they find. During a crisis this rule is amplified, and the cost of having ignored it is measured in months of a wrong narrative that you don’t control.
In the next articles in the series I’ll explain how to build a newsroom that works as a “canonical source” for AI engines, how to use press releases in an AI-friendly format, and how to measure the weight of your citations relative to the competitors in your sector.
