Recommendation engine

Allergens are a wall, budget is a weight: how a skincare recommendation engine should treat safety

A good skincare recommendation engine treats allergens as a hard filter and budget as a soft preference. How Claré builds a routine, not a leaderboard.

12 November 2026 · 7 min read · By the Claré AI team

Recommendation engines often score everything on one scale. Price, fit and safety are all turned into numbers and added up. For a movie or a playlist, that is fine. For skincare it is a mistake, because some things are preferences and some things are walls.

Safety is a wall, not a penalty

If a shopper declares an allergen, any product containing it is removed. It is not given a lower score. It is not moved down the list. It is gone. A penalty can be outweighed by a high match score, and an engine that lets a great fit override a declared allergen will eventually recommend something harmful.

The same applies to very sensitive skin. Shoppers who say their skin is very sensitive only see products the brand has marked as suitable for sensitive skin. These walls never relax. Background on how allergic reactions to products happen is covered in this overview of contact dermatitis, and it is a good reminder of why we do not treat allergens as a matter of taste.

Budget is a weight

Budget works differently. It is a soft ceiling, a preference rather than a rule. If fewer than three products survive the filters, the engine relaxes the ceiling by one tier, and any product picked that way is labelled above your budget. The shopper is never surprised, and never shown a quietly expensive product pretending to fit.

Why the difference matters to a brand

A wall protects your customer and your name. A weight protects your conversion. Mixing them up means you either lose sales by being too strict about price or lose trust by being too loose about safety. Separating them gives you both.

Scoring without tag-stuffing

Once products are filtered, the engine scores what is left. The concerns a shopper names, in priority order, get a scoring floor and a multiplier by rank, so what they care about matters even if the scan only barely saw it. Without further rules, though, a clever pattern emerges: a product tagged for every concern scores well for everyone because it matches everything a little. In an early version, a product tagged for all seven concerns scored far higher than a focused treatment for blemishes.

Diminishing returns on a product’s targets fixed it. After the first few matched concerns, each extra tag adds less. A focused product that really fits now beats a catch-all. If you enter your own catalog, this is a good reason to tag honestly: tag what a product truly targets, not everything it could possibly touch.

A routine, not a leaderboard

The last layer turns a ranked list into a routine. Shoppers do not need ten serums. They need a sensible sequence.

  • At most one product per routine step. Cleanser, face wash and soap count as the same step.
  • Step one always addresses a concern the shopper actually named.
  • Match percentages descend with rank, so the list reads honestly and never shows two products tied at the top.
  • Claré recommends four products, each chosen for a concern the scan detected, with a plain-language reason.

Pure functions, same answer every time

The recommendation engine is made of pure functions. The same input always gives the same routine. That makes it testable, which matters because a recommender that changes its mind between sessions is hard to trust and hard to debug. It also makes every pick explainable. When a customer asks why was this recommended, there is a reason that can be traced to a concern, a tag and a rule.

The explanations themselves are written by a language model, but only from fact sheets about each product, and each item has a fallback template if the model fails. The model can phrase the reason. It cannot invent the product.

How to prepare your catalog

Garbage in, garbage out applies hard here. Before you onboard, write down for each product what it targets, who it suits, which severities it is meant for and which allergens it contains. Today products are entered by hand in the brand dashboard, and catalog import from your store is on our roadmap but not built yet. We would rather tell you now.

What this looks like to the shopper

A shopper never sees the filters. They see a short list of four products that make sense for them, and a plain reason under each. What they do not see is the pile of products that were quietly removed because of a declared fragrance allergy, or because a product was not suitable for very sensitive skin. That invisibility is the point. A safe recommendation should feel normal, not nervous.

It also changes how you tag your catalog. If a product contains a common allergen, tag it. If you are unsure, check with whoever formulated it. A missing tag is the one mistake the engine cannot catch for you, because it can only filter on what it has been told. Treat your product data as part of your safety system, not just a marketing field.

To see how the engine handles your range, get early access for your brand, or read how the pipeline works first.

Frequently asked questions

What happens if a shopper declares an allergy?

Any product containing that allergen is removed from consideration entirely. It is not scored lower, it is excluded.

Is budget a hard limit?

No. Budget is a soft ceiling that relaxes by one tier only if fewer than three products survive, and a relaxed pick is labelled above your budget.

Cosmetic analysis only. Not medical advice.

Ready when you are

See Claré on your catalog.

Book a demo and we will walk through a scan, the checks and a routine built from a sample catalog, then talk about how it fits your brand.