Personalized video advertisements made with generative AI received more clicks than two other ad formats in a WhatsApp experiment involving more than 21,000 customers of an Indian online retailer.
The research compared AI-generated personalized video with personalized images and a generic video. The results show how customers responded to those particular messages. They do not establish that the videos increased completed sales or that the same approach will work for every brand.
Anuj Kapoor and Madhav Kumar published the study in Marketing Science in 2025. Their experiment involved a direct-to-consumer company selling environmentally oriented products, a generative AI provider and a WhatsApp marketing provider.
Three ways to make an advertisement
Customers were randomly assigned to receive one of three types of message. One group saw an AI-generated video personalized using purchase history. A second received a personalized image ad, while the third saw a generic video.
The personalized video used an avatar and a message adapted to the recipient. The human marketing team retained control of the message content, according to an MIT research brief about the study.
Random assignment helps compare groups that should be similar on average before the messages arrive. The comparison still combines several differences. The personalized AI video differs from the personalized image in format and production method, and it differs from the generic video in personalization as well as production.
Consequently, the experiment cannot attribute the entire difference to AI, personalization or video alone. Each comparison tests a particular combination of those features.
Clicks are not the same as purchases
The researchers used click-through rate, the share of recipients who clicked, as their principal engagement measure. The AI-personalized videos produced a rate 9.4 percentage points higher than the personalized images and 6.5 points higher than the generic videos, the MIT brief reports.
A percentage-point difference is the subtraction between two rates. For illustration, a rise from 10% to 15% is five percentage points, even though the relative increase is 50%. That example is arithmetic, not a result from the experiment.
The study’s click results do not by themselves establish the value of the orders generated. A campaign can attract attention without increasing profitable purchases, and an ad that draws fewer clicks can still reach customers who are more likely to buy.
That is why a business evaluating its own campaign should connect ad exposure to later outcomes where it can do so lawfully and reliably. Our earlier coverage of retail app advertising examined a similar gap between engagement and measured sales.
What marketers should test
Before comparing formats, a marketing team can decide which outcome matters most: a product-page visit, an inquiry, an order or the contribution from that order after advertising costs. It can then use a control group and keep offers, audience selection and measurement periods as comparable as possible.
The researchers also estimated the potential production costs of personalized video. Those calculations are estimates, not a measured guarantee of savings for another business. Production prices, review requirements and the volume of variations will affect a company’s own costs.
The MIT account identifies unanswered questions about novelty, privacy and quality control. The experiment involved one exposure in one setting, so it cannot establish how customers would respond after receiving such messages repeatedly.
Businesses using purchase history for personalization also need to consider the permissions they have, the sensitivity of the information and how customers can opt out. A message may draw attention while still damaging trust if its personalization feels intrusive.