Synthetic respondent research
We put 488 synthetic personas through ten different moods and asked them the same questions. Mood alone moved the answer by up to 77 points — more than any demographic factor. Here is what we found, and why it matters.
A fully within-subject design: every persona answered every question in every mood, so any difference is caused by mood — not by who was asked.
"Would you recommend your phone?" — a costless social endorsement.
"Stay with your provider, or switch?" — a self-interested, committal decision.
Happy, excited, optimistic, nostalgic, hungry, tired, anxious, sad, frustrated, angry.
What each persona says they would do, and what they are modelled to actually do.
Q1: share of personas who would recommend their phone, by mood. Stated intent (SAY) saturates at the top; modelled behaviour (DO) spreads across the full range.
A persona that recommends its phone 94% of the time when happy does so just 17% of the time when angry. Same persona, same question. Bigger than any age or gender gap we measured.
Three patterns hold across the study.
Only angry and frustrated collapse the recommendation (~17–21%). Sad and anxious personas still recommend more often than not. The reasoning logs confirm it: angry replies are blunt, dismissive, won't endorse; sad replies are muted, lukewarm (lower enthusiasm, same decision); anxious replies hedge about steering someone wrong. Negativity is not refusal — only hostility reverses the choice.
SAY saturates at the top — happy, excited and optimistic are indistinguishable at ~94%. DO spreads across the full range (83% → 17%) and reveals graded differences SAY hides. If you only ask what people say, you overestimate follow-through in every positive and neutral mood.
When personas are tired, sad, hungry or anxious, they say they would recommend far more than they would actually do. At the extremes — very happy or very angry — say and do converge. Intent is least trustworthy exactly when someone is low-energy or down.
| Mood | Say–do gap (Q1) |
|---|---|
| Sad | 35 pts |
| Tired | 33 pts |
| Anxious | 28 pts |
| Hungry | 26 pts |
| Happy / Excited | ~11 pts |
| Angry / Frustrated | ~0–4 pts |
Q1 is a costless favour, so a "no" when angry could just mean "I don't feel like helping." We re-ran the full sweep on a self-interested, committal decision — stay or switch provider — and measured the churn (switch) rate.
Hostile moods drive the highest churn (angry 65%, frustrated 48%) — the same withdrawal-under-hostility that collapsed recommendation in Q1, on an unrelated decision.
Sad & anxious personas — which dampened Q1 — are the most loyal here (13–17% churn): switching feels like risk and hassle. The same mood pushes one answer down and the other up.
High-arousal positive moods (excited 30%, optimistic 25%) drive more switching than calm-positive happy (20%). "Good mood" is not uniformly good for retention.
On Q2, SAY and DO are near-identical — because switching is a real, costly decision. The large gap in Q1 is a property of costless opinions, not personas.
Why this matters: a measurement artefact would push every mood the same direction. The double dissociation proves mood is modelling distinct appraisals — hostility → defect, anxiety → risk-averse, sadness → inertia — exactly as affective science predicts.
Aggregates hide the mechanism. Two individual personas put through all 10 moods on Q1 — one swings wildly, the other never moves. The difference is not personality type; it is conviction.
The same hedge appears in every mood — only the verdict changes.
A concrete, evidence-anchored belief — mood only changes his tone.
The lesson: mood moves the undecided, not the convinced. Conviction anchored to concrete evidence is mood-proof; ambivalence is mood-driven. The wider a respondent's mood swing, the softer their underlying preference — so a mood sweep doubles as a conviction detector, separating real signal from emotional state.
Mood is not noise to be averaged out — it is a controllable variable that exposes the ceiling and floor of any audience. Ship Mood as a first-class dimension: run any panel across a mood spectrum the same way we run it across demographics.
Run a message against an angry / frustrated panel to find worst-case response. Angry personas churn at 65% vs ~13% when calm — retention teams can pressure-test save-offer scripts against the at-risk moment they actually face.
Run it against a happy / excited panel to see the best realistic case — and how much of the gap is mood-addressable versus structural.
The spread between best-mood and worst-mood response becomes a single mood-sensitivity score for any message — robust ideas hold across moods, fragile ones only work on a good day.
Because DO diverges from SAY most in low-energy moods, quantify where stated intent can't be trusted — invaluable for forecasting real behaviour from survey intent.