I Thought AI Shopping Was a Discovery Problem. I’m Starting to Think It’s a Product Data Problem First.
Before an AI can recommend a product, it has to understand what the product actually is. That sounds obvious. In real catalogs, it often isn’t.
When I first got interested in AI shopping, I thought about it the same way most people probably do.
A customer asks an AI:
“What running shoes should I buy?”
The AI looks across the market.
It compares products.
Then it recommends a few.
From the merchant’s side, the obvious question becomes:
How do I get my product into that recommendation?
That sounds like a discovery problem.
SEO, but for AI.
Visibility.
Ranking.
Distribution.
I still think those things matter.
But the more time I spend working on Mingche, the more I keep getting pulled one step backward.
Before a system can decide whether to recommend a product, it has to answer a more basic question:
What exactly is this product?
And sometimes the catalog does not make that nearly as clear as the storefront does.
A human shopper has a huge advantage.
We are very good at filling in gaps.
We see a black T-shirt.
The page says:
Everyday Tee
There are size buttons underneath it.
The price says $49.
One color is selected.
The Add to Cart button works.
We understand what is happening almost immediately.
We do not consciously think:
Which exact offer object does this price belong to?
Is the selected variant represented by this URL?
Does the structured data describe this exact combination of size and color?
Is the availability attached to the parent product or the current variant?
We just understand the page.
Machines have to build that understanding from data.
That is where things get messy.
A product can look perfectly normal to a customer while different systems describe it differently.
The storefront might say:
Black / Medium / $49
The product feed might describe:
Black / Large / $49
The structured data might still contain:
$59
And the URL may reopen the product on:
White / Medium
None of those problems are dramatic by themselves.
The website still works.
A customer can probably figure it out.
But now imagine another system trying to answer:
“Which exact product am I looking at?”
It has to reconcile all of that.
This is why I’m becoming less interested in the phrase:
AI optimization
At least as a starting point.
Because optimization implies that the underlying object is already clear.
First understand the product.
Then optimize how it is discovered.
If the underlying product information is ambiguous, adding another layer of AI optimization feels a little like improving the signage outside a warehouse before checking whether the boxes inside are labelled correctly.
Maybe that analogy is too dramatic.
But that is increasingly how I see it.
There is another thing that bothers me about AI shopping discussions.
We tend to talk about the visible moment.
The recommendation.
Get Rowmint | Product Data Lab’s stories in your inbox
Join Medium for free to get updates from this writer.
The conversational answer.
The “Buy” button.
That is the interesting part.
Nobody gets excited about:
GTINs,
variant relationships,
availability values,
currency codes,
or structured data.
I certainly didn’t.
And yet those boring pieces determine whether systems can reliably distinguish one offer from another.
A recommendation engine eventually has to deal with ordinary facts.
Which model?
Which size?
Which color?
What price?
Available where?
Still in stock?
Those questions are not futuristic.
They are catalog questions.
This has changed how I think about what Mingche should be.
I don’t want it to become a product that promises:
“We’ll make AI recommend you.”
That is an attractive promise.
It is also not something I think a product-data tool can honestly guarantee.
There are too many things outside our control.
What I am much more comfortable saying is:
Let’s make sure the machines looking at your catalog receive a clearer version of your products.
That feels narrower.
It probably is.
But it is also something we can actually work on.
The distinction matters because discovery has two sides.
There is:
Can the system find you?
And there is:
Once it finds you, can it confidently understand what you sell?
People often collapse those into one problem.
I don’t think they are the same.
A product can be visible and still be badly represented.
A system can crawl the page and still see inconsistent information.
A feed can be accepted and still contain a product record that is much less clear than it looks to a human.
Visibility is not understanding.
That sentence has been sitting in my head for a while.
It also makes me think the transition to AI shopping may be less exotic than it sounds.
Yes, the interfaces are new.
People can ask questions instead of typing keywords.
AI can compare products conversationally.
Recommendations can become much more personal.
That part is genuinely different.
But underneath it, merchants still have the same old responsibility:
describe what they sell accurately.
The interface may change.
The need for good product information does not.
If anything, it becomes more important when more of the interpretation is happening before the shopper ever reaches the merchant’s website.
Maybe AI shopping eventually creates an entirely new layer of product infrastructure.
Maybe agents consume new feeds.
Maybe standards change.
Maybe websites become less important.
I don’t know yet.
That is part of what makes this interesting.
But I keep coming back to one thing that feels much less speculative:
a machine cannot make a good decision from a product it has misunderstood.
That is why, at least for now, I’m less interested in trying to “hack AI recommendations.”
I’m more interested in making the product easier to understand in the first place.
It is not as exciting.
It might be the more useful place to start.
