AI Personas or Causal Experiments: Where Research Agencies Should Use Each

Edited by Avi Yashchin, CEO and Co-Founder · August 7, 2026

An agency can turn a simulated choice study into a client recommendation by keeping the exact claim attached to the comparison. The useful deliverable is an auditable sentence: which alternatives were compared, whose responses were modeled, what changed, and which action the evidence supports. Qualitative findings can also support client decisions. The method must fit the claim, rather than the engagement stage.

A modeled finding that carried into a live campaign

Subconscious helped River City Federal Credit Union identify a weak slogan before campaign spending. The study used 140 simulated respondents. A staff-proposed slogan ranked weakest in the modeled comparison and subsequently performed worst in the live Facebook and Instagram test. That agreement gives an agency a concrete message-selection finding to explain, rather than a collection of plausible persona reactions. River City case study

The same public case reports campaign Result Rate rising from 0.52 to 0.74, a reported 42% increase against the prior campaign. Result Rate measures link clicks relative to impressions. The campaign comparison does not isolate a causal lift from the slogan or simulation, and the metric is not loan applications. The agency should preserve the useful message finding without converting the campaign comparison into a stronger claim.

Audit the sentence, not the tool label

An agency's evidence review should distinguish a modeled comparison, an observed campaign result and a causal explanation. The River City record supports the following client-facing sentences and actions.

The worked output is a corrected client slide with those three claims separated. The agency research lead accepts the evidence wording. The client campaign owner accepts the message and spending action. If the client brief concerns the same campaign and versions, the existing record can inform that action. A different audience, offer or creative version needs an explicit check of coverage before the agency transfers the recommendation.

This audit changes the handoff: retain the supported weakest-slogan finding, remove an unsupported attribution of campaign lift, and keep the actual decision owner visible. The public case does not publish the slogan wording, so the agency must use the client's retained versions rather than inventing example quotations.

Qualitative work can be a final client deliverable

A qualitative recommendation does not become invalid because a client uses the recommendation to allocate budget. Interviews can inform language, positioning, unmet needs and how customers describe a problem. QRCA describes qualitative research as useful for exploring perceptions, language and strategy; statistical validation and prediction require a different evidence basis. QRCA's qualitative research scope

A source-linked interview passage can support a statement about a participant's reported experience. The agency should identify the participant source, context and basis for the interpretation. A few interview passages cannot establish how common a view is across a market without a suitable sampling and estimation basis.

An AI persona conversation can generate candidate language, critique a draft or expose a missing question. Label generated responses as simulated. A plausible explanation is not customer testimony, and a generated reason does not establish the psychological mechanism behind a choice. None of those useful activities requires a universal interview-then-survey sequence.

Use a controlled comparison for a claim about changing the response

A controlled study is useful when the client asks whether a defined change affects a defined response. Subconscious runs controlled experiments on simulated buyers and reports estimated effects with confidence intervals. The resulting endpoint must remain visible: an effect on modeled choice is different from an effect on actual purchases. Subconscious study method

Before accepting such a recommendation, the research lead should inspect the actual alternatives, eligibility, assignment and response measure. For a headline effect, both arms need the same complete offer, visible facts and outside alternative. Randomly assigning modeled respondent units to different headlines supports the defined contrast; merely changing presentation order does not establish the effect of replacing one headline with another. Report the estimator, actual unit counts and uncertainty that accounts for repeated responses from the same unit.

Agree the action rule before reading the result. The client owner should specify a useful response gain relative to actual creative, distribution and delivery costs. A result too uncertain to meet that rule remains unresolved. The modeled comparison alone supplies neither campaign reach nor a sales or profit forecast.

Retain the evidence needed to defend the recommendation

The final agency record should link each exact client sentence to the source version, audience or participant basis, task, measured endpoint and relevant uncertainty. Add the recommendation, accepting owner and the conditions under which the recommendation applies. Existing suitable studies can complete that record; an audit does not automatically require new fieldwork.

When two sources disagree, compare the questions and conditions before choosing which source governs the action. A recent interview and an older experiment may concern different offers or populations. A study's validation on another task does not resolve that mismatch. Record the specific missing contrast or fact instead of automatically overruling one source with the other.

Book a decision review with the exact sentence the client will publish, the alternatives behind the recommendation and the evidence already held. Subconscious can help define the comparison and recommendation needed for that client decision. The agency retains responsibility for accurate interpretation and the client retains authority over the action.