The Fifty-Millisecond Verdict

An empirical test of whether the millisecond-scale first-impression judgement Lindgaard's group demonstrated in 2006 can be reconstructed for professional-services websites at scale, using frontier AI agents conditioned to read the way a UK buyer would.

When a UK owner-managed professional-services firm is read by a frontier AI agent instructed to behave as a calibrated buyer archetype, does the verdict reached in the first 50–500 milliseconds reliably predict the verdict reached after two minutes of exploration?

The Lindgaard, Tuch et al., and Willis-Todorov findings are foundational but were established with human participants on small samples. No published study has tested whether frontier AI agents — at the Park et al. (2024) Stanford fidelity threshold of 85% — produce statistically equivalent millisecond verdicts on professional-services websites. The study extends a twenty-year-old finding from human laboratory conditions into AI-mediated buyer-perception analysis at scale.

1. 01Capture the buyer-perceptible reading path for 50 UK owner-managed professional-services firms (10 each across solicitor / accountant / architect / IFA / consultancy) at their actual rendered state.
2. 02Construct three calibrated archetype agents from the Audience Intel Buyer Archetype Library (Layer B operational specifications): The Anxious First-Timer, The Time-Poor Pragmatist, The Considered Researcher.
3. 03Instruct each agent to render a first-impression verdict at three exposure durations — 100 ms, 1,000 ms, 120,000 ms — operationalised as the maximum tokens the agent may consume before issuing the verdict.
4. 04Score inter-duration correlation per agent per firm: do 100ms verdicts correlate with 120-second verdicts the way human studies suggest they should?
5. 05Test inter-agent agreement (do two archetypes converge on the same firm at the same duration?) and inter-firm rank stability (does the ordering of the 50 firms remain stable across re-runs?).

50 UK owner-managed professional-services firms, sampled by Companies House registration, £1m–£5m revenue band, across five verticals (10 per vertical). Sites captured at their unmodified public state.

We hypothesise that agent first-impression verdicts at 100 ms exposure will correlate with their 120-second verdicts at r > 0.7 — within the band the human-participant literature reports — and that inter-archetype agreement will be lower than within-archetype consistency, replicating the Layer B isolation discipline the methodology requires.

Hypothesis — not a finding

This study does not test whether the agents reproduce HUMAN first-impression verdicts. That is a separate, more methodologically demanding study (proposed as a follow-on). The Fifty-Millisecond Verdict tests internal consistency of the methodology — a necessary but not sufficient condition for its empirical claim.

Operationalising millisecond exposure for an AI agent is an inherent approximation — token-budget proxies are not literal millisecond perception. The study will publish the operationalisation transparently and invite methodological critique. No firm in the sample will be named publicly without consent.

A 4,000-word peer-reviewable paper plus a public dataset of the 50 firm-level results (anonymised). Companion blog post for circulation. Targeted for Q4 2026 release contingent on agent inference budgets and sample-firm consents.