If you’re shopping for a cold drink online, picture these options: An aerial view of a High Noon hard seltzer can over ice. A White Claw on a ledge by a pool. A sideways can of Truly on a wood table.
A Boston startup says it can explain which picture will appeal most to millennial women — the High Noon image — and it can do the same for middle-aged men, college students, or first-time homebuyers, as well as for products from motorcycles to sneakers to toothbrushes.
As more consumer decisions happen online, sales have become dependent on the way products are photographed, as opposed to how they appear on a store shelf. Different variations of product images, no matter how big or small, can attract — or repel — potential buyers.
“When you think about the digital economy ... it runs on images,” said Boston tech entrepreneur Jehan Hamedi, who has worked in artificial intelligence and consumer sentiment for the past decade. “It is powered by trillions of commercial images that showcase products, showcase brands.“
Hamedi figured retailers would be interested in software that could evaluate online images in a way that allowed them to anticipate which ones would drive the most sales among different audiences.
“To predict what someone wants to see, I realized you needed to come up with a way to see through somebody’s eyes at scale,” he said.
Over several years, Hamedi developed an artificial intelligence engine that he claims can do just that. It “knows” which photos will appeal to specific groups of people and, even more importantly, why.
Hamedi launched a company, called Vizit, around the technology in late 2021 and raised $10 million from investors last fall. Today, big brands including L’Oreal, Keurig, and the maker of M&M’s and Orbit gum use the software to make marketing decisions.
“[He] taught computers how to see like human beings,” said former PepsiCo executive Luke Mansfield, who met Hamedi when he went through the MassChallenge startup program and now uses the technology at Harley-Davidson.
Brands typically learn about their customers through studies or focus groups with real people. Though tried and true, these efforts can be expensive and time-consuming, and sometimes they happen so late in a design process that it’s tough to implement feedback.
Hamedi saw an opportunity for technology, specifically artificial intelligence, to replace that. But first, he had to prove that different groups of consumers would have different reactions to what they saw online. He started with color.
”We created an algorithm that could say, ‘All right, blues are better than pinks for this particular audience,’” he said.
He slowly expanded the system to dozens of other characteristics that make up an image. After years of refining and testing the algorithm, Vizit’s AI showed the ability to predict the visual preferences of specific groups of people.
Vizit’s AI models have been trained by a variety of sources, including large social networks, retail sites, and market research studies. Vizit also uses data on mouse-clicking, which tends to have a tight correlation to where the human eye looks on a screen.
A few years ago, PepsiCo wanted to see if the AI actually worked and put Hamedi to the test during a meeting. The corporation had invested a lot of time and money to figure out which Lay’s potato chip variety appealed most to millennial women, and its executives wanted to know whether Hamedi’s AI system — then an early version of Vizit — could predict the right one.
In less than 10 seconds, it did.
“They almost fell out of their chairs and were like, ‘How did you do that?’” Hamedi recalled. “We basically predicted what they had spent millions of dollars researching.”
Other marketing executives said Vizit’s software delivers similar results to their traditional consumer research methods. The draw is that it delivers results within seconds and doesn’t cost as much.
With Vizit, brands identify the audience they are trying to sell to — Gen Z iced coffee drinkers, say, or parents with young children. Then they upload the images of their product that would appear online, and Vizit rates their effectiveness on a scale from 1 to 100.
Vizit also offers information on how a company’s competitors might score, as well as suggestions on how to improve content, pointing out the “positive” and “negative” aspects of a photo.
“Something as simple as, like, you’re trying to sell a grill, and when you show the image ... the fire pit is just distracting,” Hamedi said. “Tone it down a little bit, and you’ll sell more grills.”
Vizit had to develop hundreds of different AI systems, called “audience lenses,” to represent different groups of consumers. That’s because “people prefer different things across cultures, across demographics, life stages,” he said.
Roman Vorobiev, who heads global design for food conglomerate Mars, said he’s using Vizit because his company, which owns Ben’s Original rice, Pedigree, and Snickers, was stuck in an “archaic way” of design testing that didn’t consider online shoppers.
“We have great intelligence around how to build great TV ads... good intelligence about how to build packaging for on-shelf,” he said. “We were completely ignoring user experience online.”
Not everyone is rushing to add AI to the creative process. But many believe it has opened up a new frontier that could help marketers have more objective conversations. For instance, Vizit helped Mars realize that some of its flagship designs that worked well in stores didn’t translate to e-commerce, Vorobiev said.
One Mars pet food brand was agonizing over which picture of a steak filet to feature on its packaging. Vizit’s software indicated that no matter which steak it featured on the package, it would still score low with online shoppers. The AI suggested depicting a cartoon cow instead.
“We hadn’t even thought about that,” Vorobiev said. “We were in a rabbit hole improving this image of steak.”
Petr Kaplunovich, a longtime Boston-area product designer, said he’s intrigued by Vizit because it takes a different approach than the popular generative AI tools, such as DALL-E, which creates images, or ChatGPT, which produces text responses.
“Vizit is attacking it from the other side,” he said. “Now that you have ideas generated, let’s try to use predictive analysis to evaluate whether it is good.”
Kaplunovich does worry that if marketers use the same AI systems, they might all end up with images that look the same. And AI might have a hard time evaluating out-of-the-box, risky designs that have never been tried before.
Still, he’s optimistic that the AI boom will have a positive impact on his field.
“People will be able to do what’s considered great marketing today in five minutes, but then that won’t be considered great anymore,” he said. “The ‘good enough’ bar is going to be raised significantly.”
Anissa Gardizy can be reached at firstname.lastname@example.org. Follow her on Twitter @anissagardizy8 and on Instagram @anissagardizy.journalism.