Why the AI Personalisation Project Should Be Last, Not First
Sometime in the next year the AI personalisation project will land on your roadmap. A recommendation engine, an AI concierge, a homepage that rearranges itself for every visitor. The vendor demo will be impressive. It always is, because the demo runs on the vendor's sample catalogue, where every product has a clean title, a category, a full set of attributes and a purchase history behind it. Then the same engine gets pointed at a real store and recommends a third phone case to someone who just bought two, because that's all it can see. The problem was never the model. The model is the last step, and most stores that rush to it have skipped the steps that decide whether it works.
A recommendation engine can only read what you've written down
Any engine deciding what to show a visitor works off the fields in your catalogue: titles, categories, attributes, tags, descriptions. We've written before about titles that describe the pack instead of the product ("Your Product Titles Don't Say What the Product Is") and about the unglamorous taxonomy work sitting behind AI visibility ("Shopify Categories and Category Metafields: The Unglamorous Work Behind AI Visibility"). Both pieces were framed around search and external AI assistants, but the same fields drive any engine you install on your own store. If your catalogue doesn't record anywhere that a product is formulated for sensitive skin, no model can recommend it to the customer who told you they have sensitive skin. The information isn't hidden from the engine. It doesn't exist.
A quick test we use: could a new staff member group your catalogue into sensible ranges using only the exported product data, with no photos and no tribal knowledge? If a person can't, a model can't.
"What looks like an AI capability problem is nearly always a data entry problem wearing a better outfit."
It also needs to know something about the customer
The other half of every recommendation is the person, and here most stores are thinner than they think. The typical customer record holds an email, a postcode and one or two orders. Purchase history at that depth gives you "people also bought", which your theme already does for free.
The context that makes personalisation feel personal, the pet's breed, the child's age, the vehicle, the skin type, has to be captured deliberately, and the cheapest moment to capture it is the first purchase, when the customer can see exactly why you're asking. We covered where to ask on Shopify and what to do with the answers in "Capture the Data at the First Purchase, Not the Third". A year of that capture running quietly in the background is worth more to a future recommendation engine than any amount of model sophistication laid over empty records.
The sequence, and why it pays twice
So the order matters more than the tool. Fix the naming first, so every product says what it is, who it's for and what makes it different. Then the categories and attributes, working from best sellers down rather than trying to boil the whole catalogue at once. Then set up the zero-party capture and let it run while you get on with other things. Then, on top of records that actually contain something, the engine.
By that point, something interesting happens: a decent share of what you wanted the engine for turns out to be rules you can write yourself.
"If a customer bought puppy food and told you the breed, showing the adult formula for that breed at the right month isn't artificial intelligence. It's a merge field."
The engine still earns its place for the patterns you'd never spot manually, but it's doing the hard residue of the job rather than compensating for missing data, and that's the version of the project that works.
There's a second payoff for the same work. The fields an on-site engine reads are the same fields external AI assistants and shopping agents read when deciding whether your products make a shortlist at all, which we covered in "Can an AI Agent Actually Buy From Your Store?". Clean titles, filled categories and honest attributes improve the store you already have, this quarter, whether or not the personalisation project ever ships. That's the real sequencing argument. The foundation work is not a detour on the way to the AI project. It is the AI project.
Measure it yourself
Two checks, roughly an hour each.
Export your top 20 products and count how many have a category assigned and the attributes a buyer would filter on actually filled in, sizes, materials, life stage, whatever your category runs on.
Then try writing three personalisation rules by hand in the form "customers who bought X and told us Y should see Z". If you can't fill in the Y from any field you hold today, pause the engine budget and put it into capture instead.
When those rules become easy to write, you're ready, and the project that would have disappointed last year becomes a small one.
If you want help working out which side of that line you're on, get in touch with the team at beCommerce.
Is your store making any of these mistakes?
We audit Shopify stores for exactly these kinds of conversion-killing patterns — the ones that have become normalised until you see them through a customer's eyes.
Book your free audit →
