What Focus Groups Are Telling You (It's Not Whether People Will Buy) (askrally.com)

Companies have been turning to synthetic personas and “always‑on” AI focus groups to scale user research, only to find predictions don’t map to real purchasing behavior. The piece argues this failure isn’t that LLMs are bad — it’s that we ask them to do the impossible: predict human actions. Drawing on Richard Sutton’s distinction, LLMs learn the discourse layer from text (what people say and justify), not embodied experience, so they reproduce how people will talk about choices rather than reliably forecasting which option they’ll pick at the point of sale. The speech‑action gap (social desirability/hypothetical bias) is therefore expected, not a bug. That gap is also the product’s richest signal. Calibrated personas trained on interview transcripts and enriched with “memories” (e.g., Base Memories or manual_memories) reveal the narratives, trade‑offs, constraints and objections that actually shape decisions in conversations — the social scripts that spread ideas and influence aggregate behavior. Practically, the author urges teams to stop asking “will they buy?” and start asking “how will they talk about X?” — use personas to stress‑test messaging, surface hidden barriers (e.g., affordability framing, identity costs), and design interventions (mid‑tier products, reframed copy) that change the conversational terrain rather than chase impossible individual predictions.