L’Oréal Beauty Genius: personalised advice before the retailer hand-off

External-source research synthesis, not an OfferSignal client result. Based on the retained October 2026 research and source-specific dates.

A buyer may know the problem—skin tone, hair colour or a routine—but not which product fits. A useful assistant must connect personal intent and visual cues to current product information without turning a plausible recommendation into an unauthorised purchase. L’Oréal Beauty Genius is a concrete example: the public evidence supports an advisory service for L’Oréal Paris products, not an autonomous buying agent.

What L’Oréal launched

L’Oréal Paris launched Beauty Genius in the United States in 2024. The service offers diagnoses, virtual try-on, answers to questions and recommendations from more than 750 L’Oréal Paris skincare, makeup and hair-colour products. L’Oréal’s 2024 annual report says it recorded more than 50,000 conversations in its first two months. That is reported use, not evidence of accuracy, retention or sales lift. 1

CES 2024 should not be treated as the launch date: L’Oréal described the technology as demonstrated there. In June 2025, the company said a Meta partnership would make Beauty Genius available on WhatsApp “soon”; WWD reported in July that access was expected in early 2026. L’Oréal’s current undated corporate page, accessed at the research cutoff, instead says the web service and WhatsApp are live. WhatsApp’s exact launch date, countries, languages and eligibility remain undisclosed. France is the organisation’s case market, not evidence of a French consumer rollout. 4 5 6 9

The data and service boundary

The US web experience combines generative artificial intelligence, augmented reality, proprietary L’Oréal Paris knowledge, computer vision and colour science. It supports free-text beauty questions, skin, skin-tone and hair guidance, routines, education, recommendations and virtual try-on. Account holders can preserve conversation history, while customer-care requests are redirected to a service pathway. 2 3

L’Oréal’s corporate description adds selfie skin and hair scans, a recommendation engine trained on thousands of real consultations, a skin atlas with 150,000 dermatologist annotations, AR try-ons and connections to major US retailers. Recommendations are limited to L’Oréal Paris even when education discusses ingredients, myths and trends more broadly. Diginomica reported L’Oréal’s claim that the service integrated more than ten large language models and analysed hundreds of thousands of historical conversations; this is executive/vendor testimony, not an external technical audit. 4 6 7

The sources do not show access to a retail account, cart, live price or stock, coupons, payment credentials, order submission or negotiation. A retailer connection is a hand-off after advice. The consumer retains purchase choice and the retailer retains checkout and fulfilment authority. “Agentic” therefore does not establish transaction authority. 4 6

What the numbers do—and do not—show

WWD reported more than 480,000 conversations by 1 July 2025, while L’Oréal’s current corporate page says more than 1.2 million US consumer conversations. These are not one comparable funnel: definitions, users, repeat use, periods and geography are not supplied. WWD also quoted a L’Oréal executive saying engaged consumers had significantly higher conversion and average order value than non-engaged consumers, but gave no figures. That self-selected comparison is not causal evidence of incremental conversion, revenue or return on investment. 4 6

The FAQ reports virtual try-on training on more than 6,000 inclusive images and makeup-artist testing on more than 10,000 products in 50 countries. It reports skin algorithms trained on more than 16,000 inclusive images and more than 150,000 dermatologist-annotated images. No external test set, subgroup performance, error rate or clinical protocol is published. Microsoft separately describes latency under five seconds at global scale, without a percentile, load, measurement period or independent benchmark. 3 8

An illustrative product-discovery application

This is an illustrative application, not a reported deployment. A consumer-products retailer could use the pattern for one governed product family: capture the customer’s need and permitted image-derived attributes; map them to versioned product, shade, ingredient and compatibility records; show the evidence and territory; then offer a retailer link. Catalogue, price, stock and promotion should remain separate fields owned by authorised systems. If input is insufficient or sensitive, the assistant should ask, abstain or route to a human rather than imply a diagnosis.

Test unique consumers, recommendation exposure, correct-fit decisions, retailer click-through, checkout, repeat use, returns, complaints and satisfaction. Compare against a holdout or defined baseline. Keep recommendation, merchant quote, consumer acceptance and payment as separate permissioned stages. Beauty Genius demonstrates the value of structured fit-oriented discovery; it does not prove that a fluent assistant should buy on the customer’s behalf.

Sources and next reading

Explore OfferSignal · Focused Business & AI Workshop · All research

References