The most common AI product recommendations mistakes are feeding the engine poor catalog data, putting the same widget on every page, optimising for clicks instead of margin or revenue, and launching without a proper holdout test. Each one makes recommendations look busy while adding little. Fixing them usually matters more than switching vendors.
AI product recommendations are suggestions generated by models that learn from browsing, purchase and catalog data. They appear as “customers also bought”, “complete the look”, personalised home page rows, cart add-ons and email blocks. Shopify, Adobe Commerce, Salesforce Commerce Cloud and specialist tools such as Nosto, Klevu, Algolia Recommend and Dynamic Yield all offer versions of this.
Mistake 1: treating the catalog as an afterthought
Recommendation models learn from attributes. If half your products lack a category, colour or size field, the engine has less to work with and falls back on popularity. That is why many stores see the same ten bestsellers recommended everywhere.
Before tuning the model, audit the feed. Look for missing attributes, inconsistent naming (“Navy”, “navy blue”, “dark blue”), duplicate products and variants listed as separate items. A clean feed often improves results before any setting changes.
Mistake 2: why does one widget on every page fail?
Shoppers have different needs at different points. On a product page they compare. In the cart they look for a small add-on. After purchase they might need accessories or a refill. Using one “you may also like” block everywhere ignores all of that.
- Product page: similar items and alternatives at nearby prices
- Cart: low-cost complements that make sense with what is already there
- Post-purchase email: accessories, consumables, care products
- Home page: personalised rows for returning visitors, curated picks for new ones
- Out-of-stock page: close substitutes in stock
A store selling running shoes, for example, might show other shoes on the product page, socks and insoles in the cart, and a replacement reminder after about the distance most runners expect a pair to last.
Mistake 3: optimising for the wrong metric
Click-through rate is easy to raise. Show cheaper or more eye-catching products and clicks go up. Revenue, average order value and margin may not move, or may fall.
Set the objective inside the tool where it allows. Many platforms let you weight recommendations by revenue, margin or inventory. If yours does not, at least measure attributed revenue per session rather than clicks, and watch whether recommendations cannibalise products customers would have bought anyway.
Mistake 4: is a before-and-after comparison enough?
No. Comparing this month with last month mixes the recommendation change with seasonality, promotions and ad spend. The result looks like proof and proves very little.
Run a holdout test instead. Most serious platforms support A/B testing, where a share of visitors, often 10 to 20 percent, sees no recommendations or the old version. Run it long enough to cover at least one full weekly cycle and enough orders to read the difference with confidence. Small stores may need several weeks.
Which mistakes appear after launch?
Launch is not the end of the work. Three problems usually surface in the first months, once real traffic hits the engine and the merchandising team stops watching it every day.
Mistake 5: ignoring cold start
New visitors and new products have no history. A model that relies only on behaviour shows them generic popular items, which is fine on day one and a lost chance after that.
The fix is a mix: content-based rules using product attributes for new products, context such as traffic source or campaign for new visitors, and behaviour-based models once data exists. Check how your tool handles this; some do it by default, some need configuration.
Mistake 6: letting the engine override business rules
Models do not know that a product is about to be discontinued, that a brand partner forbids being shown next to a competitor, or that a category has thin margins. Without rules, recommendations can promote items you do not want to sell.
Rules worth adding in most stores:
- Exclude out-of-stock and low-stock items
- Exclude items with returns above your threshold
- Keep recommendations within a sensible price band of the viewed item
- Block sensitive pairings, such as adult products next to children’s items
- Pin seasonal or launch items for a fixed period
Mistake 7: forgetting privacy and consent
Personalised recommendations use behavioural data. Under GDPR and Turkey’s KVKK, tracking that is not strictly necessary often needs consent, and many shoppers decline. If your consent setup blocks the recommendation script entirely, a large share of visitors may see nothing.
Check what the tool does without consent. A good setup falls back to non-personal recommendations based on the current page, which still adds value and stays within the rules. Test it yourself: open the store in a private window, decline cookies and see what each placement shows. If the blocks are empty, the fix is usually a configuration change in the consent tool or the recommendation script, not a new vendor.
How Alien Road applies AI product recommendations
In a typical engagement, discovery starts with a catalog audit and a review of current placements, objectives and tracking. We look at which widgets exist, what each one is optimised for and whether any test has ever been run.
Setup then follows the order above: clean the feed, assign one purpose per placement, add business rules and set up a holdout. Measurement uses attributed revenue per session and average order value against the holdout group, reported monthly with notes on catalog changes and promotions. When a vendor change makes sense, we say so, but most of the time the existing tool has room left.
Where to start this week
Export your product feed and count missing values in category, colour, size and price fields. That single number tells you how much the engine is guessing. Sort the gaps by revenue, so the products that sell most get fixed first.
Then open your store on a phone and walk through product page, cart and checkout. Note every recommendation block, what it shows and whether it makes sense for that moment. Most teams find at least one placement they would remove on the spot.