Product Recommendations, Start to Finish: Catalog, Algorithm, Design and a Preview for Any Shopper
A recommendation widget looks like a design problem — a row of products under a heading. It is mostly a data problem. What it can recommend depends on your catalog; what it learns from depends on the events your site sends; what each shopper sees depends on the algorithm, its fallbacks and what it is told to skip. The design comes last.
Personyze’s campaign editor walks it in that order: Catalog, Product events, Content, Target, then QA and the Simulator before anything goes live, and Performance after. This post follows one campaign through every step — “Recommend products where people decide, on product pages” on a demo store. The screens are real; the numbers are the demo account’s.
- Check the catalog
- Check the events
- Pick the algorithm and its fallbacks
- Design it, and preview it as a real shopper
- Test it against no widget
- Decide where it appears
- Try it on any product
- Read the results
1. Check the catalog
Recommendations choose from your product catalog: its pictures, names, prices and stock. Connect your store once — Shopify, Magento and the other store apps keep it current by themselves — or import a feed. The first step of every recommendation campaign checks the catalog against what this campaign needs from it.
Fix what the warning names before you design anything. A recommendation that shows a blank card or a sold-out product costs more trust than it earns clicks.
2. Check the events
Recommendations learn from what shoppers do with your products: which ones they view, what they put in the cart, what they buy. The Product events step shows which of those your site already sends, how many arrived lately and where each is read from — on most stores, the same events Google Analytics already receives.
3. Pick the algorithm and its fallbacks
Add the widget from the Content step: pick Product recommendations, then the template it starts from.
The widget opens on the algorithm: what it recommends, and in what order. Pick the page it will be on first — the same algorithm works from the product being viewed on a product page, from everything recently viewed on the home page and from the cart on the cart page.
On a product page, Others Who Viewed Also Viewed is the natural choice — products co-viewed by other visitors with the one on screen.
The settings around it are where a widget becomes yours:
- Fallback algorithms fill the slots when the main algorithm has too little to go on — a new product, a quiet day. The yellow note above them is about the main algorithm: it works from the product being viewed, so it needs a product on the page — which is why this campaign is shown on product pages.
- Skip items the visitor already… is viewing now, has viewed, has in the cart, has saved or has bought.
- Restrict the catalog — match the visitor’s gender or age, only discounted items, only items with an image, a link or a price, only the viewed item’s categories.
- Fine-tune the algorithm — gentle boosts for profit margin or new items, or a slot reserved for new items. The algorithm still leads.
How an engine weighs these is covered in how a personalization engine decides.
4. Design it, and preview it as a real shopper
The second step is the look: a template, a style accent, the layout, fonts and the order of a card’s parts, or Edit with AI to describe the change. Preview as on switches between phone, tablet and desktop.
Above the preview sits the most useful control in the editor: Preview as. By default the widget shows sample data. Pick a person instead — the list opens on real visitors, filtered by what they did: viewed, added to cart, purchased — and the widget renders the products that person would really get.
A visitor who is not on the site right now is previewed from the profile and history stored for them, which is exactly what their next visit brings back. Step through a few people with the arrows before you publish: someone with a long history, someone who only bought, and someone new — who should see the fallback, not an empty row.
5. Test it against no widget
A recommendation widget always gets clicks. The question is whether it adds orders that would not have happened anyway. The answer is a baseline: a share of visitors who see the page exactly as it was.
Add a second version — another algorithm, another layout — and Auto-pick can move traffic to the winner by itself. How to read the result is in how to read A/B test results.
6. Decide where it appears
The widget is built; the Target step decides which pages and which visitors it appears for. Everyone matches until you add a rule. This one needs a single rule: product pages.
A different widget for a different audience — first-time visitors and returning buyers, say — is a second campaign with its own rules, so each can be measured on its own. Describing an audience in plain words is covered in the AI targeting assistant.
7. Try it on any product
QA checks the campaign as a whole before launch — who it would reach, what it presents and where. The Simulator next to it runs the recommendation algorithm on its own, for one real visitor, for one product, or for a new visitor with no history at all.
It runs the algorithm as it is on screen, including changes you have not saved, so you can try a setting before committing to it. Keep the widget, vary the person or the product: what changes between runs is only who or what is being simulated.
8. Read the results
Once it is live, Performance shows the widget’s whole funnel: how often it was shown and actually seen, how many sessions engaged with an item, added one to the cart and bought, and the revenue attributed to it.
The checklist
- Catalog: no product without a picture, a name and a link; stock kept current.
- Events: product views arriving; add to cart and purchase set up.
- Algorithm chosen for the page; two fallbacks; skips set.
- Preview as three real people — long history, buyer, newcomer.
- A baseline share that sees no widget.
- One rule per campaign; a second audience is a second campaign.
- Simulator run on a few products, including a new one.
- Results read on orders and revenue against the baseline, not on clicks.
Questions people ask
What does a recommendation widget need before it can work?
A product catalog with a picture, a name, a link, a price and stock for each product, and the product events your site sends: at least product views, ideally add to cart and purchases too. The first two steps of the campaign check both.
What happens when the algorithm has nothing to recommend?
The fallback algorithms fill the slots, in order. A product-page algorithm such as Others Who Viewed Also Viewed works from the product being viewed, so put that widget on product pages.
How can I see what a particular shopper will get?
Use Preview as in the widget’s design step. Pick a person from the list, filtered by what they did, and the widget shows the products they would really get, with the algorithm or fallback behind each one.
How do I know the widget adds sales?
Keep a share of visitors on the site original as a baseline and compare orders and revenue, not clicks. The A/B settings on the Content step split the traffic.
Will it recommend products that are out of stock?
No. Out-of-stock products are never recommended, and you can also skip items the visitor is viewing, has in the cart or has already bought.
Book a demo with a personalization expert
30 minutes with a personalization expert. Bring your stack, your goals, your skepticism. We'll show you what changes when every visit feels like the only one.