How to Run Crisis Communications When AI Is Repeating the Story

A negative story is breaking. Your media list is handled, the statement is drafted, and the team is watching coverage. But there's a channel most crisis plans still skip: people are asking AI about it.

Journalists background-check with ChatGPT before they call you. Customers ask Perplexity whether the story is true. Investors ask Gemini what it means. All of the major assistants, ChatGPT, Gemini, Perplexity, Claude, and Google's AI Overviews, handle current-events questions the same way: they run live web searches, pick a handful of top sources, and summarize them into an answer with citations. If the only sources are the negative story, then the negative story is the answer.

That's uncomfortable, but it's also workable. The mechanism is retrieval-augmented generation, and unlike the model's training data, it updates as fast as the web does. You can't edit the answer, but you can find out what it says, trace which sources feed it, and change that mix. This guide walks through the playbook step by step. It's written for comms teams generally; the callouts show how each step works if you run it in Spyglasses.

Step	What you're doing	What tells you it's working
1. Capture both narratives	Defining what to measure	Claims and counter-messages written down
2. List the questions	Building the monitoring set	Questions cover story, record, and buying intent
3. Daily monitoring	Getting a daily read	Data flowing from day one
4. Log the timeline	Making movement explainable	Every statement and story dated
5. Watch the decay	Judging the trend	Negative surfacing falls day over day
6. Trace the sources	Finding the leverage	Ranked list of sources to fix
7. Track displacement	Measuring the response	AI starts citing your side
8. Stand down	Closing the loop	Surfacing back at baseline, retrospective done

Capture both narratives in writing

Before you measure anything, write down what you're measuring. Two lists:

1. The crisis claims. Each harmful statement, phrased the way it actually spreads. "The recall was hidden from regulators," not "negative recall coverage."
2. The counter-narrative. The specific messages you need to land. "The company self-reported within 24 hours," and so on.

Keep each one short and concrete. You're going to check AI answers against these lists daily, and vague entries produce vague answers. This also forces a useful internal conversation early: agreeing on exactly which claims you're fighting is harder, and more clarifying, than most teams expect.

List the questions people will ask AI

Nobody types your press release headline into ChatGPT. They ask questions. Draft the set a concerned person would actually ask:

• About the story: "What happened with [brand] and the recall?" "Is the [brand] story true?"
• About the record: "Has [brand] had safety problems before?" "Is [brand] trustworthy?"
• About the decision: "Should I still buy from [brand]?" "Is it safe to use [brand]?"

That last group matters most and gets forgotten first. Story questions spike and fade with the news cycle. Decision questions are where the crisis quietly costs you revenue for months, because they're asked by people at the moment of purchase.

Aim for eight to twelve questions across all three groups. Fewer and you'll miss angles; many more and you'll drown in noise.

Put monitoring on a daily cadence

Checking AI answers once, by hand, tells you what one platform said one time. AI answers are nondeterministic; the same question can produce different answers in different sessions. What you need is a daily, repeatable read across platforms, starting the day the story breaks.

Day one matters because it's your baseline. If you don't know how often the claim surfaced on Tuesday, you can't say whether Friday's statement helped. Run every question, on every platform you care about, every day, and record whether each answer repeats a crisis claim, carries your counter-message, and which sources it cites.

Do the math on the manual version: ten questions across four platforms is 40 answers to collect and classify every day, 280 a week, before anyone analyzes anything. Automate it or assign it, but don't sample it.

Log every statement and story as it happens

Crisis timelines get reconstructed badly after the fact. Keep the record as you go: the initial story, your first statement, the CEO interview, each major follow-up, every correction you win. Date and one line each.

This isn't housekeeping. It's what turns your monitoring data into an argument. "Negative surfacing dropped nine points" is a chart. "Negative surfacing dropped nine points starting the day after the CEO statement" is proof your response worked, and it's the sentence that goes in front of the board.

Watch the narrative decay, not the daily noise

Here's the shape of a normal crisis in AI answers: a spike in the first few days as coverage lands and gets indexed, then a decay as the news cycle moves on and other sources re-enter the mix. Your job isn't to prevent the spike. By the time you're reading this, it happened. Your job is to steepen the decay.

So judge two trend lines, not any single day's answers:

1. Negative surfacing: what share of answers repeat at least one crisis claim. Expect it high early. Watch the slope.
2. Counter-message pull-through: what share carry your side. This should climb as statements and corrections get picked up.

A bad day inside a good week doesn't need a war-room meeting. A flat week does, because a narrative that isn't decaying has sources actively sustaining it. In the example below, negative surfacing peaks at 37% of tracked answers on day 5 and falls to 16% by day 14; that slope, not any single day, is the number to report. If it flattens for a week, that's your cue for the next step.

Trace the sources feeding AI's answers

AI answers cite their sources, and that's your leverage. Collect the citations behind every answer that repeats a crisis claim. You'll usually find a short list, five to ten sources, doing most of the damage: the original story, a few syndications, sometimes an old forum thread or a stale wiki page that predates the crisis entirely.

Then rank them, because they are not equal. A correction at a source AI leans on changes answers. A correction at an outlet AI can't even crawl changes nothing, no matter how prestigious the masthead. Many major publications block AI crawlers entirely, which cuts both ways in a crisis: their negative story may not reach AI answers at all, and neither will their correction.

Source situation	Priority	The move
High AI influence, factual error	Highest	Request correction today, with documentation
High AI influence, accurate but one-sided	High	Pitch the follow-up; offer the spokesperson
Stale page (old wiki, forum thread)	Medium	Update or respond where possible; publish the authoritative page it should cite instead
Low AI influence or blocked to AI crawlers	Low	Handle for traditional reasons; it won't move AI answers

Track whether your response is displacing the story

Your response generates its own artifacts: statements, an FAQ page, corrective stories, follow-up coverage with your framing. Treat them as a portfolio and track them the way you tracked the negative sources. Are they getting indexed? Are they being cited? Are they showing up in answers where the original story used to be alone?

This is the metric that closes the loop. Negative surfacing falling is good; your sources being cited in its place is better, because displacement is durable. The next time the topic comes up, the retrieval mix includes your side by default. It also tells you what to write more of; if the press release built for citation is getting picked up and the FAQ page isn't, that's next week's content plan.

Stand down and write the retrospective

When negative surfacing is back near baseline and your counter-message holds steady, stand down. Two things before the team moves on:

1. Keep a tripwire. Leave a small set of the decision-stage questions running on a schedule. Crisis narratives resurface; anniversaries, lawsuits, and copycat stories all re-trigger the same claims. You want to know in a day, not a quarter.
2. Write the retrospective while the data is fresh. How far did the narrative spread, per platform? Which statement moved the trend? Which outlets mattered, and which prestige placements turned out to be invisible to AI? That last finding usually changes the media strategy going forward, well beyond the crisis.

The channel is new; the discipline isn't

Comms teams already know how to run this playbook against traditional media: define the message, monitor coverage, work the sources, measure the shift. AI answers are the same discipline pointed at a new channel, with one advantage the old channel never offered: the answers cite their sources. You can see exactly which pages are feeding the narrative, and that turns crisis response from spray-and-pray into a ranked worklist.

If you want the monitoring side handled, a free AI Visibility Report shows how AI describes your brand today, before you need the answer under pressure.