Personalized Product Recommendations for Beauty Retailers: Ranking Across Many Brands

A single brand recommending its own foundation has forty or fifty shades to sort. A retailer has thousands, spread across dozens of brands, each with its own shade naming logic, its own claim vocabulary and its own level of data hygiene. That is why personalized product recommendations for beauty retailers are a different engineering problem from personalization inside one brand's site — and why most retailers end up defaulting to bestseller lists that ignore who is actually looking at the screen.

The way out is not a bigger recommendation model. It is a shared customer profile that every brand in the catalog can be matched against, plus a catalog normalization pass that makes matching possible at all.

Why the retailer problem is harder

Inside one brand, the shade ladder is internally consistent. "220 Natural Beige" sits in a known position relative to "210" and "230", and the product team knows the undertone logic behind it. A retailer importing feeds from many suppliers gets none of that for free. Three brands can use the word "beige" for three visibly different shades. One brand publishes undertone, another does not. Some feeds carry finish and coverage, others carry only a marketing paragraph.

So the retailer has two gaps to close: it does not know enough about the shopper, and it does not know enough about the products in a comparable format. Personalization fails on either gap alone.

How personalized product recommendations for beauty retailers actually get built

Start with a profile that is brand-neutral

The shopper side is the easier half to standardize. A selfie-based analysis returns attributes that do not belong to any brand: skin tone and undertone, skin type and visible condition, hair attributes, each with a confidence value. MaIA Scan delivers exactly that as JSON from an API call, trained on Brazilian skin, which matters for a retailer whose assortment and audience are local even when the brands are global.

These fields describe what is visible in a photo. They are not a diagnosis and should never be presented as one — the copy around the result belongs in the language of "visible characteristics", not health.

Because the profile is brand-neutral, it can be stored once on the customer record and reused against every brand in the assortment. The shopper does not repeat a quiz per brand, and the retailer does not maintain parallel personalization logic per supplier.

Normalize the catalog before you rank anything

This is the part retailers underestimate. Before a shortlist can exist, every relevant SKU needs structured attributes in one vocabulary: for complexion products, an approximate tone position and an undertone family; for skincare, the skin types and visible conditions it addresses, plus the step in a routine; for haircare, texture, scalp condition and treatment history.

Some of this arrives in supplier feeds. Some has to be derived from product copy, swatch images or merchandising knowledge. The honest approach is to start with the categories where the attribute mapping is clearest — foundation, concealer, base skincare — and expand once the matching rules are proving themselves.

Use confidence to decide how hard to commit

A ranking engine that treats every profile field as equally certain will embarrass itself. If undertone came back with low confidence because the lighting was poor, do not lead with a shade-critical recommendation. Widen the shortlist, show two neighbouring options, or ask one clarifying question.

Confidence is also a useful retail signal in the other direction: high-confidence profiles are the ones worth building reactivation campaigns around, because the attributes will still be reliable weeks later.

Where the recommendations appear

Category and listing pages

The highest-leverage placement for a retailer is the category page, where the shopper is facing the full assortment. A "matched to your profile" rail at the top, or a filter that reorders the grid, changes the experience from browsing a warehouse to being served.

Conversational channels

Recommendation quality shows up most clearly in conversation, where the shopper can push back. MaIA Beauty Advisor reads the beauty profile and recommends products from the retailer's own catalog on WhatsApp, product pages and chat. For a multi-brand retailer this is also a merchandising surface: the advisor can respect exclusivity deals, stock position and margin bands while still answering the shopper's actual question.

Confirmation before checkout

A shortlist narrows the choice; a visual check closes it. Letting the shopper see a recommended lipstick or foundation rendered on their own photo turns an abstract match into something they can judge. MaIA Try-On covers the photo-based case through the API, with palettes calibrated to each brand's real products — which for a retailer means one calibration project per brand partner, not one guess per SKU.

What to measure

Resist the single-metric trap. Three signals tell you whether the system is working:

• Shortlist engagement — do shoppers click the matched items more than the default merchandising order?
• Shade and type returns — for complexion and skincare, are wrong-match returns falling in the cohorts that completed a profile?
• Repeat relevance — does the second visit, using the stored profile, convert better than the first?

Any conversion figures you have seen published belong to the brands that published them. Results vary with each catalog, traffic mix and channel, so your baseline is your own.

A realistic build order

Normalize one category. Capture profiles in one placement. Rank, measure, then expand to the next category and the next channel. Retailers that try to personalize the whole assortment at once usually ship nothing.

Ready to test it against your own catalog? Create an account and get an API key, or talk to the B4A team about a retailer-scale rollout.