The Ten Questions a Retailer Should Ask Before Buying "AI Skin Analysis"
They may all call it "AI skin analysis." They are not selling the same thing
Disclosure, same as in the map: I'm the founder of SKINBOT, one of the vendors this checklist applies to. These questions are fair to ask any of us, including me. That's the point of a checklist: it doesn't care who's answering.
The map came first: "AI skin analysis" is not one category, but four layers wearing the same label, visualization (Perfect Corp), measurement (Haut.AI), brand-aligned guidance (Revieve), and the neutral decision layer (SKINBOT). This is the checklist. Because the natural next question is the practical one: the map is useful, but I'm a retailer with three pitches in my inbox. What do I actually ask?
Fair. Here is the due diligence list I would bring to the pitch. Ten questions, grouped by what they reveal. None of them are trick questions. Every serious vendor should be able to answer the questions that apply to its layer.
Part 1: Ownership of the recommendation
If two products in my catalog fit the same skin profile, what decides which one your system recommends?
This is the whole game in one question. Listen for the mechanism: clinical logic, margin, brand partnership, popularity, randomness. Any answer is legitimate. A vague answer is not.
Can a brand pay you, directly or indirectly, to appear more often in recommendations?
Yes is a valid business model, it's called retail media. But you need to know whether you're buying judgment or inventory placement, because your customer will assume it's judgment.
Would the recommendation change if the seller's margin changed?
The Indifference Test, asked out loud. Watch how long the answer takes.
Part 2: The economics
What exactly do I pay for: the analysis, the recommendation, or the conversion?
Pricing reveals architecture. Per-scan pricing means you're buying measurement. Revenue share on recommended products means the system has a financial stake in what it recommends. Flat infrastructure pricing means the answer is decoupled from the outcome.
Who else pays you in this transaction?
If the vendor has revenue from brands and revenue from you, ask which side the recommendation logic answers to when the two conflict.
Part 3: The data
What happens to my customer's photo and answers after the session?
Retention policy is a compliance question and a trust question at once. Stateless means no personal skin profile or photo is retained after the session. Anything else means you need to know where, how long, and under which jurisdiction, before your legal team asks.
Whose catalog does the system recommend from: mine, yours, or a brand's?
A system that recommends from your full catalog serves your shelf. A system that recommends from a curated subset serves whoever curated it. Ask to see the subset logic.
Who is liable when the recommendation is wrong for a customer?
Pregnancy contraindications, ingredient conflicts, sensitive skin. Somebody owns that risk. Read the contract before assuming it's the vendor.
Part 4: The future
When my customer asks ChatGPT instead of visiting my site, where is your system in that conversation?
AI channels are becoming the front door of discovery. A vendor with no answer here is selling you the previous decade.
Show me a recommendation your system made against the commercial interest of a paying party.
The only question on this list that can't be answered with a slide. If they have an example, you'll learn what their architecture actually does under pressure. If they don't, you've learned that too.
How to read the answers
The goal is not to make vendors fail. The goal is to make vendors sort themselves. A visualization layer will answer honestly that recommendation logic isn't their product. A measurement layer will point downstream to your own stack. A brand-aligned layer will tell you which brands are behind the guidance, and that's a legitimate model when it's disclosed. A neutral layer should welcome question 10.
The only wrong outcome is a mismatch: buying one layer while believing you bought another. That mismatch is invisible in the demo and expensive in production, because the customer standing in your store inherits it.
Print the ten questions. Bring them to the next pitch. Every vendor on the market, my team and I included, should be able to sit across from this list without flinching.
FAQ
What should a retailer ask an AI skin analysis vendor?
Ten core questions covering recommendation ownership (what decides between two matching products, whether brands can pay for placement), economics (what the retailer pays for, who else pays the vendor), data (retention, catalog scope, liability), and future channels (presence inside AI assistants like ChatGPT). The full checklist is in this essay.
What is the Indifference Test in vendor due diligence?
Ask the vendor: would the recommendation change if the seller's margin changed? If yes, the system produces placements, not recommendations. It is the fastest single question for sorting AI beauty tech vendors by layer.
Why does the pricing model reveal an AI vendor's architecture?
Per-scan pricing indicates a measurement product, revenue share on recommended products indicates a system with a financial stake in its own answers, and flat infrastructure pricing indicates recommendation logic decoupled from commercial outcomes.