You can rent the model. You can't rent the memory.
Every company building in AI commerce right now is standing on the same few models. OpenAI, Anthropic, Google — a handful of foundation models that everyone, including us, rents by the token. That's the uncomfortable truth nobody in a pitch meeting wants to say out loud: the "AI" is a commodity. Your competitor can call the same model you do, tomorrow, with a credit card.
So if the intelligence is rentable, where's the moat? The honest answer is that it isn't in the model at all. It's in the thing the model can't supply on its own: memory.
The model is rented. The memory is owned.
A foundation model arrives knowing the whole internet and nothing about your customers. Ask it to recommend a product and it reasons from generic priors — what's popular, what's reviewed well, what shows up in its training data. It does not know that in this category, this item sells with that one. It does not know which week demand spikes, which region over-indexes, which basket a loyal buyer builds on their fourth order versus their first. That knowledge doesn't live in the model. It lives in years of real transactions — and you either have those or you don't.
A rented model gives every competitor the same brain. Owned purchase history gives one of them a memory none of the others can buy.
This is why "we use AI" is not a strategy. Everyone uses AI. The strategy is what you point the AI at — and the most valuable thing to point it at is a long, clean, correlated record of how real people actually bought.
Why food and CPG is where this matters most
Not all purchase data is equal. A one-time big-ticket buy teaches you almost nothing repeatable. But food and consumer packaged goods are the opposite: high-frequency, habitual, seasonal, and deeply correlated. The same household buys the same staples on a rhythm, adds the same companions to the cart, spikes the same way around the same holidays year after year. Run that for fifteen years and you don't have a spreadsheet — you have a map of how a cuisine actually gets cooked and bought across a country.
That map is what makes an AI storefront convert. When our engine recommends, it isn't guessing from the model's generic instincts. It's recommending the way a fifteen-year buyer in that category actually shops — co-purchase, seasonality, geography, cultural demand, all already learned. The model supplies the language. The data supplies the judgment.
"But won't AI-era data make this a commodity too?"
It's the right question, and the fair version of the skeptic's case: as agentic commerce scales, every platform will accumulate behavioral signals, so won't proprietary history stop mattering? Two reasons it doesn't.
First, history can't be backfilled. A competitor who starts collecting today has today's data. They cannot collect 2011's. Fifteen years of correlated behavior is not a thing you can buy, rent, or generate — it's a thing you either started recording a long time ago or you didn't. The clock is the moat.
Second, first-party beats aggregate for recommendation. A giant cross-merchant panel knows a little about everything. A deep first-party record knows everything about one thing. For the question that actually drives a sale — "what should this shopper, in this category, buy next" — depth wins. The brand that owns its own customers' history will out-recommend a platform that rents a slice of everyone's.
The takeaway for anyone building or buying in this space
Stop evaluating AI commerce tools by which model they wrap — they all wrap the same ones. Evaluate them by what proprietary, durable signal they bring that you can't get elsewhere. A tool with no data of its own is a thin layer over a rented brain; useful, replaceable, and priced like it. A platform built on owned, deep, correlated purchase history is something else entirely — because the day the models get cheaper, faster, and identical for everyone (and they will), the only thing left standing between competitors is the memory. We bet the company on owning ours.