Product Images Get You Chosen in eCommerce with Jehan Hamedi of Vizit
Watch the Episode
This episode is sponsored by the B2B Commerce Trends and Opportunities Report, presented by Shopware and written and researched by Content Cucumber. Grab your free copy at contentcucumber.com/b2b-commerce-trends-report.
Jehan Hamedi looks at pictures for a living, which is how he described his job to me, and the pictures he looks at are the product images that brands like the ones in your kitchen cabinet put in front of shoppers. He founded Vizit, an AI company in Boston that scores product imagery for how well it converts, and he has spent about 15 years building AI measurement systems for enterprise brands and retailers, starting in text analytics before he moved to visual content. We recorded in June, he rated my Bruce Lee joke an eleven, and then we spent the better part of an hour on one question, which is how you know whether an image is any good.
The Last Unmeasured Thing on the Digital Shelf
Jehan is clear that the computer vision field solved the descriptive problem a long time ago. Detection and classification can tell you a kitchen is a kitchen and a person is standing on the left. What he wanted was something past that.
“All of those purchases have exactly one thing in common. And it’s that you saw something, it sparked something, and you wanted it. And that moment is very hard to explain and even harder to measure,” he says.
He was talking about the coffee maker on your counter, your shoes, the giant jar of pickles from Costco, and the point he kept coming back to is that the moment a shopper first sees a product is the one piece of the funnel nobody has instrumented.
“Imagery is the last unmeasured component of the digital shelf. We have data, e-commerce managers have data on everything except that moment, the first impression that a shopper has with a product, that is instant. It’s irreversible,” he says.
He puts that first impression at under 13 milliseconds, and he applies the same logic to video, where a 60-second spot lives or dies on whether the first frame gets clicked.
Simulating the Beholder
I pushed on the obvious objection, which is that good is subjective. I train models on a client’s voice for a living and I know how many definitions of good there are, so I asked him how Vizit builds a baseline.
“The right definition of good is not whether something is compliant or whether something is on brand or off-brand. It’s all about the shopper. Beauty is in the eye of the beholder,” he says.
His answer is that Vizit was founded on the belief that an AI system can simulate the beholder. The company trains its own models in-house on what he calls a visual conversion graph, decades of behavioral data, product data, sales data, social data, and search data for particular categories, so the model learns how shoppers in a category choose, and can then look at an image it has not seen before, one you have not put online yet, and predict how likely it is to convert that audience. Vizit calls the output a Vizit Score, and the score sits against the competitive set in the category rather than against a style guide.
Crickets from the Brand Manager
The problem he is solving is easier to see from the brand side. Call up the person who owns a popular deodorant brand and ask a simple question.
“If you called the brand manager of a popular deodorant brand, and you said, what are your best images of that brand? Let’s just say Old Spice, for example, you would get crickets because they don’t know,” he says.
Brands do not have a way to measure images, so they are driving paid search, paid social, and retail media traffic to content they cannot grade. Jehan’s argument is that the fix is smaller than a rebrand or a packaging overhaul.
“Oftentimes, what I have found is that the fastest path to incremental revenue starts with one better image,” he says.
The line that gives this episode its title came out of the same stretch of the conversation, and it is the one I keep repeating.
“Text kind of gets you found, but images get you chosen,” he says.
Cat Food in Japan
I asked whether he sees patterns in what works, and his answer was that context is king and the context includes the country.
“The best way to photograph cat food for a US shopper on Amazon is going to be markedly different than one in India, than one in the UK, than one in Japan,” he says.
He wants brands to stop treating the image as a creative deliverable that ships and gets forgotten, and start treating it the way they treat a price or a promotion.
“The biggest mistake that I think brands make is that they treat their imagery and track their imagery as a creative deliverable as opposed to a performance asset,” he says.
His illustration is to right-click any image on any website. You get the file size, the capture date, maybe the RGB values, and that metadata is descriptive, and it tells you nothing about whether the image sells the product, and he wants that measurement to travel with the asset for its whole life.
Context Is What Makes the Generator Useful
I have used the image tools inside Shopify and Google campaigns and some of them are poor, a 16 by 9 cropped to a square with no resize, so I asked him whether there is a better way to make the new image once you know the old one scores low.
“The single most important thing you can give an AI to do a job better or more effectively than you could do on your own is context,” he says.
Context is the gap he sees between a general purpose generator and one aimed at conversion. Vizit’s tool for rebuilding a low scoring image is Vizit Spark, which takes the score, the plain English recommendations about why an image underperforms, and the category context, and feeds the whole packet into generation, so the weak image gets rebuilt rather than described, and it does so without the five-finger problem, the weird crop, or the product that legal will not let you ship because the generator changed it.
“Reliably produce higher scoring, higher converting images in just a couple seconds,” is how he describes the output.
He also told me the generalist tools have their place, and that the applications that win are the ones sitting on proprietary context, customer data plus public data plus licensed data, that a general model does not have. And then there was the cookie.
“There is a statistically right number of chocolate chips to show in a chocolate chip cookie to make it look more appetizing. I did that research,” he says.
A Plug for Humans
Before the shameless plug I put in a plug for people, because a machine grading images is telling you what a machine predicts a human wants, and a human clicking on it confirms the prediction, and none of that produces anything new. AI looks sideways, at what has already been seen and scraped, and I still want people inventing the thing a model would not reach for.
Jehan met me there. His company is a company of humans who spend their days talking to customers, and the AI automates the parts that take a long time or need more context than a person can hold. He also does not buy the version of agentic commerce where the agent does the shopping unsupervised.
“This idea that agents are going to just take over and do all of the shopping for us, without any supervision, I don’t subscribe to that,” he says.
He cites a figure that 95 percent of consumers who use AI on a shopping journey take additional steps before they buy, and Walmart research on a three times higher conversion rate when a shopper clicks through from ChatGPT to a product page, because people want to see the thing.
“People don’t buy products, they buy outcomes,” he says.
Scored Before It Ships
His prediction for where this goes is the one I want to check back on in a couple of years.
“I believe that in the next couple years, if not even sooner, every product image will be scored before it’s shipped. Just like the same way if you were a brand, you would never put out a financial report without auditing it first,” he says.
His challenge to anyone listening was the same question he opened with, pointed at your own team.
“Ask your brand manager, ask your e-commerce manager, how good is our content? How good are our images? Are they better than our competitors? And if they can’t give you an answer, you’re doing something wrong,” he says.
Jehan Hamedi is the founder of Vizit. You can find him on LinkedIn at linkedin.com/in/jehanhamedi and the company at vizit.com.