Simpler AI personas predicted ad clicks better in one campaign

A language model asked to act like a shopper can express interest in a product. Predicting an advertising click is a different task: people can like a product without noticing or clicking its display ad. Jiwoo Choi and colleagues investigated how instructions addressing that distinction affect AI-generated advertising forecasts.

Their study, posted to arXiv, compared ways of prompting a language model to simulate customers. In the campaign tested, basic demographic profiles combined with a qualitative reminder about ignoring advertisements produced the most accurate click estimates. Adding personality, occupation or explicit numerical click-rate guidance did not improve the results.

Testing simulated shoppers against campaign records

The researchers used Google Ads display-ad data for a jelly product, covering April 5–11, 2026. Five separately targeted demographic groups received 974,423 ad impressions—instances of an ad being displayed—and generated 3,578 clicks. The groups’ observed click-through rates ranged from 0.174% to 0.680%.

Each simulated customer received a profile containing age range, gender, parental status, income range and an interest setting shared by all groups: “Shoppers.” The model also received the advertisement’s image and text metadata, then returned a binary decision: click or do not click.

The team used a fixed version of GPT-5 nano at temperature 1.0, a setting governing randomness in generation. Each experimental condition involved 974,423 simulated decisions, matching the campaign’s impression count. Main comparisons were repeated in three independent runs.

The proportion of simulated clicks became a predicted click-through rate. Multiplying that rate by recorded impressions produced predicted click counts for each demographic group on each day. This means the experiment evaluated response estimates given known exposure volumes; it did not predict how many impressions a future campaign would receive.

Less detail, lower error

The best-performing prompt retained the basic profile, omitted time information and richer personal attributes, and reminded the model that most users ignore display ads. It also included the sentence: “Younger adults tend to show much stronger banner blindness than older adults.” Banner blindness refers here to users overlooking or ignoring advertisements.

Across the 35 group-day observations, this configuration had a mean absolute error of 24.9 clicks, with a standard deviation of 1.9 across runs. Mean absolute error measures the average size of prediction errors without allowing overestimates and underestimates to cancel. Its mean absolute percentage error was 26.1%; that is an error measure, not a claim of 73.9% predictive accuracy.

The lowest mean absolute error among the tested non-language-model methods was 51.1 clicks. That method simply assumed a 0.2% click-through rate for every group. The selected AI configuration reduced this error by 51.3%. Other comparisons included fixed rates of 0.5% and 1.0%, plus age-based rules adapted from research-paper recommendation data.

A richer AI configuration containing time, personality and occupation information, alongside explicit numerical click-rate guidance, had an error of 109.9 clicks. Replacing its numerical guidance with the relative qualitative cue reduced error to 37.7 while retaining the rich profile. Removing the extra profile details lowered it further to 24.9.

These comparisons support the authors’ conclusion that additional detail was not automatically useful in this setting. Their proposed explanations—such as rich profiles encouraging excessive inference about preferences—remain interpretations rather than demonstrated internal mechanisms.

Wording mattered substantially

Even two similar qualitative instructions produced different forecasts. With the basic profile, changing “Younger adults exhibit significantly more banner blindness than older adults” to the selected wording reduced mean absolute error from 35.2 to 24.9 clicks. Across the tested profile configurations, wording changes shifted mean error by approximately 8–30%.

Removing the generic reminder about ignored ads produced much larger failures. The demographics-only condition predicted 947,516 clicks, compared with 3,578 actually recorded. This demonstrates a severe mismatch between uncalibrated simulated choices and observed advertising behavior in this experiment.

What remains uncertain

The evidence comes from one product campaign, five demographic groups, seven days and one model. The configurations were compared on those same campaign records, not evaluated on separate campaigns after the prompt was fixed. Three repeated runs measure variation in the simulation, not reliability across advertising markets.

The selected prompt also supplies an age-related behavioral assumption consistent with the campaign’s broad response pattern. Its success does not establish that younger adults universally ignore ads more, or that the same instruction would help when audience responses follow a different pattern.

Errors remained within individual groups: the selected method underestimated one group’s click-through rate by about 27% and overestimated another’s by about 50%. The result is therefore a useful finding about prompt design within this campaign—not evidence that simulated customers can replace prospective advertising tests.