Crisis communication simulation: 24 messages tested

Crisis message testing has to accommodate disagreement about the incident as well as the proposed response. In our consumer goods crisis engagement, the board wanted to restore reputation and retain market share. Views differed about the crisis’s cause, who was responsible and what should happen next. The board needed evidence for choosing how to communicate, with customer credibility and the company’s licence to operate at stake.

Live research was impractical because of time pressure, sensitive strategic options, and difficulty reaching policymakers, regulators and commercial decision-makers, according to How Artificial Societies steered a client through a major crisis. Putting those options to human participants would expose material the team needed to test privately.

Audience differences complicate the choice. Wording that reassures an investor may trouble a regulator or customer. Our crisis engagement kept five societies distinct: Policymakers and Regulators, Media, Influencers, Consumers, and Commercial Customers. We would preserve those distinctions in the decision discussion: agreement on the wording inside the company does not establish how the groups receiving it will react.

Our crisis simulation began with 1,174 AI personas in five societies, according to the crisis case study. The five societies represented Policymakers and Regulators, Media, Influencers, Consumers, and Commercial Customers. We grounded the personas in anonymised real profiles and conversations, using social listening and deep research. We also mapped social influence networks to model opinion formation among connected people.

Before exposing those societies to a message, we surveyed their views of the company and the issue. The baseline covered sentiment and likely purchasing behaviour. Purchasing intentions remain statements about likely behaviour, so we use them to compare responses without treating them as sales observations.

We then tested messages, statements and creative activations targeted at the relevant audiences. The study’s total was 24 messages and creatives. Reactions after exposure were compared with the baseline to examine how the proposed communication changed opinion. This is the pre-exposure/post-exposure design used in our message testing: establish where the audience starts, introduce the narrative, then measure the change.

The experimental detail worth examining is the memory reset. We built up and erased the societies’ memories of the messages and their delivery around the exposure tests. Resetting removes contamination from earlier messages and controls order effects: a reaction to a draft should not depend on having already read another draft in the test. Human participants cannot forget an earlier statement on demand. With simulated audiences, we could examine alternatives without carrying that exposure into the next comparison.

The engagement generated 328,576 responses, according to the Artificial Societies crisis study. Read that as the volume of reactions produced by the tests, alongside the persona count. For your statement, the useful comparison is the change in each audience’s response and the reasons behind it; the total gives the scale of the work, not a verdict on which wording to issue.

Our crisis message testing pointed the client towards evidence-backed statements about its actions to mitigate and resolve the issue. The case study describes that as the route forward for the business. The finding concerned what the company could say about its response to the underlying problem, rather than a more appealing description of the crisis.

Delivery needed attention too. The simulation identified likely backlash against certain activations in the context of the issue and distinguished how and when messages should reach key audiences. We would give that finding as much weight as the choice of wording. A board selecting a statement also needs to examine the proposed activation carrying it; a positive reaction to the text alone leaves the delivery decision open.

The client used the findings and recommendations to decide its crisis strategy and communications. It then ran follow-on simulations of delivery choices and where activations might backfire among audience segments. For the board, that meant progressing from a comparison of 24 messages and creatives to tests of the proposed delivery, with possible backlash examined by segment. The study demonstrates that decision process; the reactions being compared are simulated reactions.

The published crisis communication simulation took under two weeks from start to finish, according to our crisis case study. That duration includes constructing the societies, establishing the baseline, testing messages with memory resets and delivering the work.

Our published engagements distinguish elapsed project time from the time spent fielding research on a society that already exists. The table keeps those clocks separate. Their durations describe the engagements concerned.

Source: Artificial Societies’ case studies linked in each row. Teneo’s engagement took place in late 2025.

For repeat decisions, society construction becomes a planning choice. In the earnings-call engagement, the societies remained ready for the following quarter. Those four societies covered Sell-Side Analysts, Buy-Side Investors, Corporate Ecosystem and Financial Media, according to the earnings-call study. The client could rerun scenarios and retest its script as the numbers firmed up. We would discuss a standing society when you expect to revisit an audience, and agree the scope and deadline of the next test separately. The earnings-call duration belongs to that engagement; it is not a crisis-response service commitment.

Reputation risk extends the horizon of crisis message testing. A society can examine a recovery strategy after an incident or reactions to a proposed acquisition or leadership change. It can also be surveyed again as strategy and external events change. The crisis client’s follow-on delivery tests show how work continues after the initial choice of response.

The transport-policy engagement shows how we stress-test a policy narrative. A global transportation leader needed to understand Washington, D.C. opinion leaders without exposing sensitive alternatives to them. Its Washington D.C. Opinion Leaders society represented 1,500 personas spanning more than 360 organisations and 800 job titles, according to the transport-policy case study. We tested two competing narratives and their proof points, then introduced damaging coverage and hypothetical scenarios. From 250,000 responses to more than 170 questions, the transport-policy study identified an existing credibility gap and showed where narratives withstood negative coverage or lost support.

That is a useful addition to periodic brand tracking. A tracking survey records perceptions at a point in time; a networked society tests how those perceptions would change under a proposed strategy. In the transport-policy work, even showing the drafts to the intended audience would have changed the debate being studied. Simulation let the team examine its choices before doing that.

Confidential strategy work is the other published case. In our Teneo engagement, the team tested a major US company’s technology narrative before launch.

Crisis message testing supports a choice between responses by examining changes in simulated audience opinion. Our crisis study connects the baseline, message exposure and delivery tests to a decision the client made.

Audience construction depends on observations of the people you need to represent, whether public or client-provided. In the crisis study, anonymised profiles and conversations provided that grounding. If your audience needs further research before it can be modelled, obtaining those observations belongs in the study plan. Our guide to hard-to-reach audiences addresses that research problem.

Where the decision requires evidence of actual behaviour, we would commission human research to confirm it. Our published blog describes the intention–behaviour gap that simulation inherits from survey research: stated intentions can differ from action. A favourable purchasing response in a crisis test supports a comparison of messages; establishing what customers subsequently buy requires evidence of purchases. Give simulation the job it can perform before publication: examining competing responses while you can still change the wording and delivery.