Begin with authorized source material
Use customer research and reviews you are permitted to analyze. Record the source, date, product version and context. Remove unnecessary personal details from shared workbooks. A public review is not automatic permission to turn its author into a named testimonial in an advertisement.
Choose material relevant to the current offer. A complaint about an old delivery policy may not explain reactions to a new product demonstration. Keep that context attached rather than asking an AI summary to infer it.
Build two evidence columns
Download the evidence workbook.
| Real customer evidence | Synthetic exploration |
|---|---|
| Exact excerpt from the source | Exact generated answer |
| Date, context and source link | Question, persona assumptions and run context |
| What that customer reported | What the model hypothesized |
| Limits of the available sample | No population sampling claim |
Do not invent review excerpts for the sheet. The empty template is deliberate: it becomes evidence only when completed with real material. The synthetic side must remain labeled even when its language sounds like a customer interview.
Turn a theme into a question
If real reviews repeatedly discuss setup difficulty, ask what information the current ad gives about setup. That is a neutral question tied to a source theme. “Why does this ad fail to explain setup?” already assumes the conclusion.
Ask follow-ups about the exact phrase or scene that produced an interpretation. Keep the original answer and the follow-up together. Separate what the ad visibly says from what the answer infers.
Use agreement carefully
When a simulated objection resembles a real review theme, it may help you explore wording. It is not independent confirmation if the review was supplied in the prompt. The model may simply be repeating its input.
When the simulation disagrees, inspect the assumptions before announcing a new segment insight. The generated persona may have been given different knowledge, a different use case or a misleading summary.
Research by Bisbee et al. illustrates reliability limits in synthetic public-opinion responses. That study is not an ad-testing validation; it supports asking task-specific questions rather than assuming human-like fluency establishes representative evidence.
Write a message-change hypothesis
Use this structure: source theme → visible creative gap → proposed message → evidence needed. For example, in a fictional setup-related brief, the proposed change could show the first setup action instead of asserting that setup is easy.
The acceptance check is whether the demonstration is clear and truthful. Any claim that it improved conversion requires separate campaign evidence. Any claim that a population now understands it requires an appropriate human study.
Keep the workflow honest and useful
This article describes a research method, not an automatic integration that imports every review platform into SaliencyLab. Bring relevant context into your review and check what the study actually used.
Read synthetic users for ad testing for a question bank and the revision brief to hand a supported edit to the creative team. Book a demo to discuss your messaging question.