Can synthetic respondents measure brand awareness?

Define the awareness measure first

Unaided awareness asks someone to name brands from a category without showing a brand list. Top-of-mind awareness records the first brand named. Aided awareness supplies a cue, such as a name or logo, and asks whether the person recognizes it. Ipsos's awareness glossary describes these distinctions.

A synthetic system's ability to describe a brand does not establish how many people in your market recognize it. The estimate you need might concern U.S. category buyers this month, residents served by a regional business, or people who have never purchased from you. State that population and time period before discussing the model.

Keep consideration, awareness, reported use, and verified purchase separate. Someone can recognize a name without considering it, and a generated account of buying something is not a transaction record. If a system uses observed purchases or previous human answers, ask which conclusions come from those records and which are inferred.

Check what the model was told

For an unaided question, inspect the information available before the answer: the research brief, profile, earlier questions, retrieved documents, and prior conversation. A brand name can enter through any of these routes.

Pew Research Center's questionnaire guidance explains that earlier questions can supply concepts that people later mention in an open-ended answer. Our proposed synthetic-study check extends that concern to the model's whole input context: document brand cues, not only the visible question.

A product description can support a useful reaction to that description. It changes the task, however. An answer after reading a brand's story should not be labeled an unaided measure of prior market awareness.

A fictional setup problem

Imagine testing a fictional detergent brand called ArcMint. The synthetic profile says, “You regularly buy ArcMint,” and the next question asks which detergent brands come to mind.

An ArcMint answer follows information supplied by the researcher. It does not independently show that real category buyers remember ArcMint. Removing that sentence creates a different setup to evaluate; it does not by itself establish an accurate estimate. This example is invented and describes no Lewsearch run or customer result.

Ask for an awareness-specific comparison

The checklist below is our proposed buyer evaluation, not an industry certification or a claim that every synthetic method passes it.

1. Use a relevant human reference. Match the category cue, audience, market, dates, question sequence, brand list, and recognition stimulus. Record the human recruitment method, sample sizes, weights, and limitations. Account for the human reference’s uncertainty, especially for small brands and audience cuts.
2. Protect the outcome. Keep evaluation awareness scores and respondent answers out of model setup and tuning. Identify any earlier brand studies used as inputs. Public results can carry unresolved training-data overlap, even when you withhold their links.
3. Score each measure separately. Compare top-of-mind recall, total unaided recall, and aided recognition against their corresponding human results. Freeze how misspellings, brand families, empty answers, and “none” are coded before scoring. Report each brand’s human and synthetic percentages, their denominators, and the percentage-point difference.
4. Include the brands that matter to the decision. Inspect errors for smaller or unfamiliar brands alongside category leaders, rather than accepting one pooled score. Where appropriate, include a clearly documented fictional-brand diagnostic for unsupported recognition. Treat it as a diagnostic, not a substitute for the human comparison.
5. Review relevant audiences. Inspect customers and noncustomers, markets, or other decision-critical groups where the human evidence supports those cuts. A customer-heavy reference does not establish awareness among prospects.
6. Agree on the allowed use. Set decision tolerances before unblinding. If the result misses them, continue with human measurement or narrow the synthetic task. Material changes to the model, inputs, or scoring procedure need a fresh evaluation.

For the underlying holdout procedure, use our blind-validation guide. An awareness comparison at one date also does not establish tracking accuracy; our longitudinal guide covers changes between waves.

Keep recollections attached to their source

A generated statement such as “I saw this brand last week” does not by itself establish that a person remembers the brand. If an observed human record supports the statement, identify that record and distinguish any generated reconstruction from the original human response. Even an ad-exposure log shows a recorded exposure, not necessarily that the person noticed or remembers it. Human recall answers are self-reports and can also be mistaken.

AAPOR's Code, revised June 2026, distinguishes generated cases from human research participants and calls for disclosure of AI-generated data. Apply the same clarity in the delivery: label model-based estimates, preserve the human reference, and identify any observed records used as inputs.

For a new brand without relevant human evidence, use synthetic exploration to develop hypotheses and questions for people. Commission human awareness measurement when the decision requires a current estimate of what the market actually remembers. Discuss the audience and evaluation before treating an awareness output as a planning number.