Virtual Beauty Advisor KPIs: What to Measure in the First 90 Days

Most beauty teams launch a virtual beauty advisor with a single number in mind: conversion. Then the first month ends, the conversion chart moves a little, and nobody can explain why. The problem is rarely the advisor — it is that the virtual beauty advisor KPIs were never defined before launch, so there is no way to tell a reach problem from a recommendation problem from a catalog problem.

This article lays out a measurement plan in four layers, in the order a session actually happens: who reaches the advisor, whether the profile is captured, whether the recommendation is accepted, and what happens commercially afterwards. Instrument them in that order and every later debate has an answer.

Why a virtual beauty advisor needs its own scorecard

An advisor is not a banner or a landing page. It is a conversation with several steps, and each step can leak. A customer can open the advisor and abandon before sending a selfie. They can send a selfie and get a profile the system is not confident about. They can get a confident profile and a shortlist that contains nothing in stock in their shade. Each of those is a different fix — more entry points, better photo guidance, better catalog data — and a single conversion number hides all three.

There is a second reason. An AI beauty advisor reads a beauty profile and recommends your own products across WhatsApp, product pages and chat. Those channels have different traffic, different intent and different baselines, so a blended average tells you very little. Measure per channel from day one.

Layer 1: reach — does anyone actually get there

Entry point impressions and open rate

Count, per entry point, how many people saw the advisor was available and how many opened it. A product page button, a WhatsApp campaign and a homepage module will behave nothing alike. Low open rate is a placement and copy problem, not an AI problem, and it is the cheapest thing on this list to fix.

Session starts by channel and device

Split by channel and by mobile versus desktop. Advisors that depend on a selfie behave differently on a laptop webcam than on a phone. If mobile starts dominate, your guidance, your framing instructions and your error messages should be written for a phone first.

Layer 2: profile capture — the quality of the input

Photo submission rate and retake rate

Of the people who opened the advisor, how many sent a photo? Of those, how many had to retake it? A high retake rate usually points at lighting and framing guidance rather than the model. Our skin and hair analysis API returns a confidence value per field precisely so you can act on this: low confidence on undertone, for example, is a signal to ask for a second photo or to widen the shortlist rather than to pretend certainty.

Confidence distribution per field

Track the distribution, not the average. If undertone confidence is consistently lower than skin tone confidence in one channel, that channel has a capture problem — a cropped frame, a heavy filter, a dark room. Fixing capture raises every downstream metric at once.

Fallback rate

How often does the advisor fall back to a question instead of an analysis result? Some fallback is healthy. A fallback rate that climbs week over week is a signal worth investigating before it reaches the revenue reports.

Layer 3: recommendation quality

Shortlist acceptance and click-through

How many customers clicked at least one recommended product, and in which position? If clicks concentrate on position one, the ranking is working. If they scatter evenly, the ranking is close to random and your catalog attributes are probably thin.

Coverage of the profile by your catalog

This is the metric teams forget. For each beauty profile the advisor produces, could the catalog answer it with products that are mapped, in stock and in the right shade? A profile with no good answer is a catalog gap, and it is far more common than a model error. Log these cases — they are also a merchandising input for the next range extension.

Refinement and re-ask rate

How often does the customer ask for something else — lighter, cheaper, fragrance-free? Refinements are not failures; they are intent data. A rising refinement rate on one attribute tells you where the first recommendation is systematically off.

Layer 4: commercial outcome

Only now does conversion make sense. Measure it as a comparison, not in isolation: sessions that used the advisor versus comparable sessions that did not, in the same channel and the same period. Add average order value, items per order, and — with a longer window — repeat purchase rate and the return or exchange rate on shade-sensitive categories. Shade complaints are a slow metric, but they are the one that proves the recommendation was right rather than merely persuasive.

Be careful with benchmarks. Published figures from other companies describe their catalog, their traffic and their channel mix, not yours. Use your own pre-launch baseline as the comparison and treat outside numbers as context only.

A simple 90-day plan

Days 1–30: instrument. Ship with event logging on every step above, split by channel. Do not optimise yet; collect a clean baseline.

Days 31–60: fix the funnel. Attack the biggest leak, usually photo submission or catalog coverage. Change one thing at a time so the chart stays readable.

Days 61–90: tune the ranking and report. With capture stable, work on recommendation quality and present commercial results alongside the funnel so leadership sees cause and effect, not just the final number.

Two mistakes to avoid: averaging across channels, and reporting a conversion lift without a comparable control group.

If you want to see the underlying beauty profile before designing your own scorecard, create an account and get an API key — or talk to our team about an advisor project built on your catalog.