A retailer sent six onboarding requirements. Read twice, they were really one: take the manual effort out. Why onboarding needs reinventing, not improving.
Earlier this week, a colleague forwarded me a list from one of our customers – a large, well-run retailer you've almost certainly shopped with.
Six requirements for product onboarding. Neatly numbered. Attributes, titles, descriptions, search, gaps, accuracy. The kind of list that looks like six separate workstreams, six owners, and probably six vendors queuing up to pitch for each one.
I read it twice. And the second time, I realised it wasn't six requirements at all. It was one requirement, written six different ways. It screamed:
“Please take the manual effort out of this!”
I wanted to talk about it publicly, because requirements lists are honest in a way strategy decks never are. When one retailer writes something like this down, most of their peers are thinking it.
The crossword with no clues
The same customer shared a spreadsheet of real supplier data for a handful of dresses. The "mandatory, always provided" columns include a stock code, a description, a season and a colour. Sounds reasonable. Until you read it.
The season field says "Spec". Spec what? Special? Specification? Speckled? Nobody knows. The selling colour is "Lima Bean". Now, I'm not sure I could pick a lima bean out of a line-up, and I'm even less sure your customer is searching for one.
Then you scroll right, and there's a whole block of columns that essentially say: can someone fill these in? Neckline. Silhouette. Back style. Fastening. Sleeve type. Fit. Every one of them is the kind of thing a customer actually filters on. Every one of them blank.
It's like being handed a crossword with no clues. A few letters pencilled in here and there. Off you go. Oh – and actually, we don't want the crossword finished. We want a story.
We've talked before about why your product needs a CV. This is the raw material for that CV. It's barely a name badge.
But supplier data has always been poor. The story is what happens next.
A relay race nobody wins
Think of onboarding as a three-leg relay race.
The first leg is origination. The manufacturer sends what they've got, written in catalogue-speak that makes perfect sense to the factory and almost none to anyone else. Brands, to be fair, have a bit to answer for here. There's a certain "we made it, there it is, what more do you want?" attitude to product content.
The second leg is the wall. The data gets thrown over it to the merchandising team, and some very dedicated people try, by hook or by crook, to work out which blanks matter and how to fill them. I remember a manager at a luxury e-tailer fielding constant complaints from their content team. All they'd get was the buyer's notes, maybe a picture, the brand. Can we get this on the site, please?
The third leg is the customer. And this is where it falls apart, quietly. The product made it onto the site, so the box got ticked. But if it rarely surfaces in search, it might as well be sitting in a dark corner of the stockroom.
The problem is the finishing line. Most onboarding processes measure "onboarded". What matters is whether somebody found it, understood it and bought it. It's not enough to onboard a product. It must be onboarded in a discoverable way.
And the cost of getting that wrong isn't theoretical. Salsify's 2026 consumer research found that 45% of shoppers had returned an online purchase because the product information was incorrect or misleading. That's not a merchandising problem. That's a margin problem.
Why the usual fixes don't stick
So why don't retailers just fix it? Most of them are trying. They're just trying in ways that keep the same relay race intact.
The first instinct is AI. "Oh, the team's using AI now." Brilliant. That usually means someone asking a chatbot "what is this?" a couple of hundred times a day and pasting the answer into a field. That's not a new process. That's data entry with a nicer keyboard.
The second instinct is the PIM. Move the workflow into the product information system, because that's where the data lives. But the PIM is where data lives, not where it gets understood. Put poor data into a better PIM, and you've got lovely new walls with the same old interior.
The third instinct, historically, was to outsource it. But this was never admin. It's research, and it needs a methodology most teams simply don't have. What they have instead is senior merchandisers working as glorified manual researchers, when they should be building ranges and thinking about customers. And the quality depends on things nobody writes into a process document – how well that person knew the category, how busy they were, whether there was decent coffee in the kitchen.
The homework nobody revisits
And it gets worse, because nobody ever goes back.
Once a product is onboarded, it's homework that's been handed in. Nobody wants to see it again. But six months later, that product has a life – reviews, questions, customers styling it on Instagram. All the lived experience you didn't have on day one now exists, and none of it finds its way back to the product page.
If anything, every product should be re-onboarded three to six months after launch. Not as a clean-up exercise, but because that's the first moment you genuinely know enough about it.
So no, onboarding doesn't need improving. You can't optimise your way out of a process that was never designed for discovery in the first place. It needs reinventing.
Start with better questions
And the reinvention starts somewhere unexpected. Not with answers, but with questions.
Every product category needs its own questionnaire. The meaningful blanks for a dress are not the same as for a teaspoon, a duvet or a tub of protein powder. Designing them is taxonomy work, and it only works if it starts with how customers actually shop – which is exactly where the Four Forces of Product Discovery begin.
Take protein powder. The standard attributes are flavour, brand and size. Vanilla, chocolate, one kilo. But people buying it want to know whether it's vegan, low in sugar, kosher, or whether they can take it before bed. If your questionnaire doesn't ask those questions, no amount of effort will answer them.
You don't need more answers. You need better questions.
Once you have the questions, you can build a research method around them, in layers. The product text and images. How people search for products like it. The manufacturer's own site, which sometimes has more than they sent you. What can be inferred – a wallet with RFID protection has a security feature, whether or not anyone wrote that down. Then lived experience: reviews, social, video.
This is where AI genuinely earns its keep. Not staring at a blank page but working through a defined questionnaire. Is it floral? Yes. Is it a ditsy print? Yes. Multicoloured? Yes. AI is very good at this – provided someone has written the questions first. Most retailers haven't.
Then the people come back in, in a very different role. Not data entry, but validation. If the machine calls a dress "elegant" rather than "chic", nobody's getting sued. But "contains nuts" is not a claim that goes live because a model thought so. That's where provenance matters – where did this answer come from, and has it been checked twice? Where you can't verify something, agree the wording up front – "cannot be verified as nut-free" – so nobody reinvents it every time.
Titles and descriptions, the thing everyone starts with, become the easy part at the end – mostly assembly, once the answers exist. And "optimising titles for SEO" turns out to mean what it always meant: optimising them for customers, because they're the ones doing the searching.
Three places to start
So, if you run e-commerce, this is where I'd start.
First, write the questionnaire around intent, before you buy anything. Pick your most important category and start with how customers actually shop it – the questions they ask, the words they search with, the details that make them choose one product over another. Then look hard at your “nice to have” fields. I’d wager a lot of them are critical to have
Second, mine what you already know, then push for more. Your PIM, search logs, reviews and customer service conversations are already full of context nobody has connected back to the product. Start there. Then push your suppliers harder – though don’t expect them to hand you intent. They know how the product was made, not how your customers shop for it. And decide now which answers AI can own and which a person must verify, before a customer complains.
Third, move the finishing line. Stop measuring “onboarded”. Start measuring whether products are found, understood and bought. And put a date in the diary to re-onboard each product, because the richest context arrives after launch, once real customers have had their say.
None of this means ripping anything out. It means deciding what your onboarding is for – and giving it the context to get there.
Because right now, for most retailers, it's for getting products on the site. And that's the wrong finishing line.
Until Tuesday after next,
Warren