A study of 5,278 shoppers finds fitting chatbot avatars lift interest slightly

Expectations differed sharply by product. In the wine scenario 41.3% of 1,767 participants described the typical salesperson as older, formally dressed and male, while 51.4% of 1,773 fashion participants described a young, informally dressed woman. When free to choose, only 10.2% of wine participants picked the wine-typical avatar, a 31-point gap between stated expectation and actual choice.

The effect sizes were small, and how much the avatar liking drove the result is the main finding. Perceived sympathy toward the avatar carried the effect of avatar assignment on intention; adding it to the model raised explained variance from an adjusted R2 of .025 to .150. Self-selection did not beat a well-matched default.

The study used a 3 × 2 between-subjects design. Participants first saw a scripted 'Virtual Shopping Assistant' named Alex in either a wine or a fashion setting. In the two assigned conditions they first said whether a typical salesperson was male or female, under 30 or over 50, and formally or informally dressed. The chatbot then appeared as a matching avatar or the exact opposite combination. Dialogue and product recommendations (up to three wines, or T-shirts and shoes) were held constant, so only appearance differed.

Eight AI-generated avatars covered every combination of gender expression, age appearance and clothing style. Behavioral intention was an average of two items: willingness to use such an assistant and willingness to recommend it. Sympathy was rated on a 0 to 100 slider for the avatar each person had actually seen.

Expectations differed by who was asked. Male participants held the male-wine-salesperson view more strongly than female participants (74% vs 62%), and female participants held the female-fashion view more strongly than male participants (93% vs 69%). Younger participants more often expected a young fashion salesperson (94% vs 77% of older participants). In free choice, male participants picked young female avatars more often than female participants did (59.0% vs 46.9%), while female participants picked older female avatars more often (31.3% vs 9.3%).

Sympathy ran in opposite directions at once. With sympathy held constant, violating avatars produced slightly higher intention (β = +.10) and self-selection slightly lower (β = −.12). Through sympathy, violating avatars cut intention (indirect β = −.21) and self-selection raised it (indirect β = +.15). The authors suggest context-appropriate defaults, sympathy-focused design and optional personalization.