Caramel was a mobile shopping marketplace and my first serious product as a full-time founder.
We connected Shopify stores to a shared shopping app and tried to distribute their products through bloggers. The stores kept their inventory and fulfilled the orders. We were supposed to bring the buyers.
The bloggers brought us none.
We changed the model. Instead of sending people back to individual stores, we imported the products into one marketplace and promoted it ourselves. In one week, we added around 200,000 products.
Then we spent roughly $1,500 on Google Ads.
We got one buyer.
We closed Caramel.
At the time, the conclusion felt rational. We did not know how to acquire customers profitably. SEO looked slow and difficult. The margins did not look large enough to justify years of work. Agencies and freelance marketers did not give us a credible alternative.
So I compressed all of that into one sentence: the unit economics do not work.
That sentence sounded precise. It was almost useless.
Why I call this my biggest mistake
In April 2020, Shopify launched Shop, a consumer app that combined product discovery, personalized recommendations, checkout, and delivery tracking across Shopify merchants.
That was remarkably close to the core bet behind Caramel: take products scattered across independent Shopify stores and turn them into one consumer shopping layer.
By the beginning of 2021, Shopify reported that Shop and Shop Pay had more than 19 million monthly active users. At one point, I personally saw Shop at number one in the Shopping category of the US App Store.
Shop was not identical to Caramel. Shopify had an enormous installed merchant base, existing customer relationships, Shop Pay, order tracking, and distribution we could not reproduce.
I also cannot prove that Caramel would have succeeded.
I call the shutdown my biggest first-time-founder mistake because we abandoned a viable category without learning whether our version could work. Shopify later validated the consumer layer at scale. My own experience with cheap.market later showed that similar marketplace economics could support a real business.
The market was not the conclusion we had reached. Our diagnosis was.
What did one buyer actually prove?
Caramel had a bad result. I am not disputing that.
What I no longer accept is the diagnosis we attached to it.
One buyer could have meant that the traffic was wrong. The products might have been unattractive. The prices might have been uncompetitive. The listing pages, cart, address form, checkout, delivery terms, or payment methods might have created friction. The average order value, repeat purchase rate, or margin might have been too low. Google Ads itself might simply have been the wrong channel.
We did not isolate these possibilities. We treated their combined result as a property of the entire business.
Even the impressive-looking number was weak evidence. Importing 200,000 products did not mean that we had 200,000 things people wanted to buy at good prices. A large catalog can contain a lot of irrelevant inventory.
We also did not spend enough to learn reliably from purchases. Today, if I wanted to test a paid channel in the US, I would expect to spend at least $10,000 on a serious marketing hypothesis. In less expensive regions, I would still expect something closer to $5,000.
That does not mean every small founder needs that budget. It means the research method has to match the amount of data available.
With little traffic, I would not wait for enough purchases to make a confident statement about the deepest part of the funnel. I would start with events that happen more often. Do visitors open product pages? Do they add anything to the cart? Where do they stop during checkout? Does the address form work? Do they understand delivery and payment?
If I could not buy enough traffic even for those events, I would run moderated usability tests and customer interviews. Ten people trying to complete a purchase while explaining what confuses them can reveal problems that a dashboard with one order cannot.
The useful budget question is not, "How much money should I spend on marketing?"
It is, "How much traffic do I need to observe the next decision-relevant event?"
The budget follows from that.
Experience changed the result before it changed the model
Years later, we built cheap.market.
The mechanics were substantially similar. Products from different sellers appeared in one consumer marketplace. We tried to solve discovery, transaction, support, logistics, and repeat purchase around them. The approximate margins were also similar.
I was no longer a first-time founder.
By then I had much more experience with acquisition, funnels, retention, localization, support automation, and marketplace operations. The team could inspect where users dropped, distinguish a traffic problem from a checkout problem, and improve the system around recurring customer questions.
This does not prove that Caramel would have succeeded. We still do not know whether its assortment was good, whether its prices were competitive, or whether another channel could have produced profitable demand.
It proves something narrower and more uncomfortable: we closed Caramel before answering those questions.
SEO is the clearest example. We dismissed it because it was slow and difficult. Both descriptions were correct. They did not make it a bad channel.
Search remains a major source of ecommerce traffic. Large stores and marketplaces invest in it precisely because the work compounds. It may take a long time to build product pages, category structure, content, links, and technical foundations. Once those pages rank, they can keep bringing high-intent visitors without paying for every click.
I am not claiming SEO would have saved Caramel. We never ran that experiment. I am saying that "slow" and "difficult" are costs, not a verdict.
The marketing advice around us did not improve the diagnosis. I spoke with agencies and freelancers, yet nobody made the small budget the central problem. They were willing to run the campaign without clarifying what the available traffic could possibly prove.
I did not know enough to challenge them.
A failed business needs a literal cause
"The unit economics do not work" combines several numbers that behave very differently.
Customer acquisition cost depends on the channel, audience, creative, and conversion path. Gross profit depends on price, product cost, payment fees, logistics, returns, and support. Lifetime value depends on repeat purchase and retention. A technical error in the address form can make the whole acquisition channel appear unprofitable.
Most of those causes can be changed.
Sometimes the decomposition reaches something much harder.
We experienced that with AI Boost, our AI photo editor. We believed AI generation would replace much of the old photo-editing market. The landscape looked ready to change, so we moved aggressively.
People did try the AI features. The problem was repeated use. Generating an unusual image was entertaining once. Far fewer people needed to keep doing it. We searched for audiences with a recurring reason to create and edit AI photos, but that audience was much smaller than the market we had expected.
That was closer to a fundamental constraint. A button, onboarding change, or cheaper ad would not create a large recurring need where one did not exist.
Caramel never reached such a conclusion. We did not discover that consumers fundamentally refused the model, that viable supply could not exist, or that the economics conflicted with an external reality we could not change. We stopped while several ordinary operating explanations were still alive.
I might use the idea again
cheap.market may return in another form. The old Caramel mechanism could return with it: Shopify stores contributing additional product offers to a larger marketplace.
That would not mean Caramel was secretly a success. It would mean the first attempt left behind useful infrastructure, unanswered hypotheses, and a lesson that took me years to formulate.
I no longer want to hear that a business model does not work without knowing exactly which part failed.
If the explanation ends with a channel, funnel step, price, retention curve, technical defect, or customer segment, there is still a specific experiment available.
If the explanation reaches a real property of the world that contradicts the intended business, shutting down may be the right decision.
We killed Caramel before we knew which one we had found.