A hierarchy only knows up and down. Why a knowledge graph, not your PIM, decides whether the AI you buy this year actually works for your customers.
One question separates the real thing from an expensive guess.
Think about your last holiday.
Not the highlight reel. Just picture where you were standing on about the third day, once you'd finally stopped checking your email.
Now give it a second and watch what turns up. The memory of the restaurant you went back to twice because the first night was so good. The walk home afterwards. The weather, which was better than it had any right to be. The people you were with. That one drink you have quietly promised yourself you will never touch again.
Notice that you didn't go looking for any of that. You asked yourself for one thing, a place, and your brain handed back a dozen others you never requested. All of them connected. None of them filed under the heading you started with.
That isn't you being clever. That's simply what happens when things live near each other. Fire off one memory and the whole neighbourhood wakes up alongside it.
And that, more or less, is a knowledge graph. Not a database. Not a taxonomy. It’s a map of what lives near what, and why.
Now, I suspect that the moment I used the words knowledge graph, some part of you quietly filed this under things I pay other people to worry about. Schemas. Taxonomies. Plumbing. And most of the time you'd be entirely right to. There is a great deal of muck and bullets down there that an ecommerce director is quite correct to rise above.
This is the exception. Not because you'll ever have to build one, or even look at one, but because it turns out to be the deciding factor in whether everything else you buy this year actually works. And you can't push the right idea down to your team if you've never formed an opinion on it yourself.
Because the system you almost certainly have instead is a hierarchy. Your PIM connects the things you've already told it to connect. For example, Foundation sits under Makeup, Makeup sits under Cosmetics, everything in its box, all very tidy. And your PIM is excellent at that. It was built to organise, and it organises beautifully.
But a hierarchy only knows up and down. It has no idea about sideways.
Say a customer has hyperpigmentation. You don't sell hyperpigmentation. What you sell is a face mask with the right acid in it, a foundation that covers well, and a vitamin that helps with collagen. Three products, three different corners of the catalogue, three separate boxes, all quietly answering the same human problem. Your PIM will never put those three together, because nobody ever told it to. A knowledge graph does it unprompted, because it knows they live near each other.
Which brings me to the beach.
Someone wants a speaker for the beach. Nobody in your business has ever tagged a speaker “beach”, and quite right too. It isn't a category, it isn't an attribute, and it would look very odd on a spec sheet.
But beach lives next door to some genuinely useful ideas. Sand. Water. No plug socket for a mile in any direction. A long, lazy day rather than a quick hour. Open air, which swallows sound. Other people, which means volume.
So, the neighbourhood lights up, and out comes the speaker with the twelve-hour battery, the splash rating and enough output to fill open space. Nobody asked for any of those things. They arrived because they live near the beach.
Which is precisely what your brain did with the restaurant.
And now the part that matters.
We've all had the conversation with an AI where you tell it something, it cheerfully ignores you, and the second you correct it, back it comes with “you're absolutely right, good catch”, as though that settles the matter, and we can all move along.
That isn't intelligence. That's a machine having a guess, missing, and displaying lovely manners about it. And it happens for one reason. It had nothing solid underneath it. Ask anything to be helpful with nothing to stand on, and it will invent something helpful-shaped.
Which is why I've become rather boring on this subject.
Anyone can point a large language model at your catalogue and have attributes sitting on every product by Friday. It will look tremendous. The demo will be excellent. The spreadsheet will be full.
The question is not whether the attributes exist. The question is whether they were earned.
Earned means somebody worked out what customers in that category care about, from evidence. What they searched for. What they compared. What they asked. What they complained about afterwards. What the reviews kept circling back to. And then matched all of that against what your product genuinely does.
Invented means a model made a plausible guess at what might matter and wrote it down with great confidence.
The two produce documents that look identical. They perform nothing alike.
I've written before about handing your product copy to ChatGPT and hoping for the best, and about products that turn up with no CV at all, no history, nothing anybody has ever bothered to say about them. This is the same argument arriving from a different direction. You cannot enrich your way out of not knowing anything.
So, the next time somebody sits across from you and uses the word intelligence, you only really need one question.
“Is it grounded in knowledge?”
Then keep going, because this is where it gets interesting. Where did that knowledge come from? What's your source of truth? How do you know those are the right attributes for this category, rather than the obvious ones anybody could have guessed? Show me.
You'll find that question does a remarkable amount of work. The demos all look much the same for the first ninety seconds. What separates them is whether there's anything underneath.
And if there's no knowledge graph under it at all, my honest advice is not to go anywhere near it.
You've been running the finest one ever built since you were about three years old. It's the reason you can still taste the holiday, wince at the drink, and picture that speaker on the sand.
Ask the same of the things you're buying.
Until Tuesday after next,
Warren
Founder & CEO, FoundIt! - The intent-led commerce platform
About Warren Cowan & FoundIt!
Warren Cowan is Founder and CEO of FoundIt!, the intent-led commerce platform used by leading multi-category retailers including M&S, B&Q, Neiman Marcus, John Lewis, and Fenwick. He writes about the future of product discovery, the intent gap, and what it actually takes to make an e-commerce site work the way customers expect.