Caitlin M. Davis
10

One beauty profile behind every touchpoint

Product strategy · PersonalizationEstée Lauder Companies
UX ResearchExperience StrategyProduct & Design LeadershipAI & Insight Systems
Methods
Capability inventory Surface mapping Source ranking State-based specification
A specification rather than a study: what personalizes, in which account state, and which source wins a conflict.
Inputs
Existing platform personalization surfaces Virtual try-on Diagnostic quiz Saved filters Account and review data
Four data sources ranked so conflicts resolved the same way on every surface.
The problem

The commerce platform was personalized in a dozen places and connected in none of them. Virtual try-on knew your shade. The diagnostic quiz knew your concerns. Your account knew what you had bought. A shopper could scan her face, get a shade match, walk to a category page, and filter through a panel that had never heard of any of it. Every feature was personal. The experience was not.

My role

Owned the research and wrote the site-wide strategy: which surfaces personalize, what each shows in each account state, and which data source wins when two disagree. Written for every brand on the platform, so it had to hold without knowing which brand was reading it.

Approach
  • Inventoried what the platform could already personalize against what it actually did
  • Named every surface where a saved attribute should change what a shopper sees
  • Ranked the data sources so conflicts resolved the same way everywhere
  • Specified each surface in three account states rather than assuming a logged-in user
What already existed
  • Offers, editorial and inline content
  • Pop-ups, modals and cart overlay banners
  • Product partials, category and product pages
  • Navigation and menus
Surfaces named
10
From the navigation opt-in through filtering, finders, editorial, reviews, offers, chat and sampling
Specified in v1
4
Opt-in, filtering, finders and editorial carried full wireframes and requirements
Account states
3
Anonymous, signed in, signed in with a face scan on file
Sources ranked
4
Face scan, diagnostic quiz, saved filters, account and review details
The hypothesis: personalization should be controlled by the shopper, and several sources of personal data living in separate features works against both the user and the business. The strategy was the argument for connecting them.
/ The precedence model
10 / Personalization strategy

Which source wins when
two of them disagree

This is the piece the platform did not have. Four features were each collecting attributes about the same shopper, and nothing said what happens when the quiz says one thing and the face scan says another. Without a stated order, every team building a personalized surface would invent their own, and the answers would drift apart.

Data storage priority model showing four sources ranked in order: virtual try-on facial recognition, diagnostic quiz, category page filter saving, and product page or review details, with a written rule for how each resolves against the others
The orderFour sources, one stated precedence. The rule that carried the most weight was the last line of the finder spec: a face scan always replaces a diagnostic match, and the shopper should never have to reconcile them by hand.Open full size
Face scan
If a shopper has used virtual try-on, that becomes the primary source for shade and skin tone. It outranks everything below it, including attributes she saved earlier by hand.
Highest
Diagnostic quiz
If she opted into saving her quiz answers, those lead, unless and until she scans. A scan overwrites them rather than sitting beside them as a second opinion.
Overridden by a scan
Saved filters
Attributes saved from a category page combine with whatever the tools already know rather than replacing them. Anyone without an account can still save, which is what turns a filter panel into an account prompt.
Combines
Account and review details
Saved account details populate automatically, but if she edits them while writing a review, the edit wins from that point on. The most recent deliberate statement is treated as the truest one.
Most recent wins
The principle underneath all four rows is the same. An attribute the shopper gave on purpose outranks one the system inferred, and the newest deliberate statement outranks the older one.
/ Opting in
10 / Personalization strategy

A door on every page,
and no account required

Personalization lived inside standalone tools, so a shopper had to already want it before she could find it. The strategy moved the entry point into the navigation, the homepage, and search, then set the rules that keep an offer like that from becoming an interruption: it does not re-trigger once dismissed, it stays reachable in the nav, and it can be used once without an account.

Navigation opt-in wireframes: a persistent beauty preferences row inside the mobile menu, and the overlay it opens offering a choice between scanning your face and answering quiz questions
In the navigationOne persistent row, and a choice of two routes into the same profile. The requirement that mattered most is the quiet one: preferences persist whether or not she is signed in, and she can edit them at any time.Open full size
Search wireframes: a search bar with voice exposed by default, and the full-screen overlay it opens showing face scan, image search and previous searches
In searchVoice exposed by default for accessibility rather than hidden behind a tap. The overlay shows previous searches inside a session and trending ones to a new shopper, so the field is useful before anything has been typed.Open full size
Not requiring an account is what made the rest work. Attaching personalization to sign-up would limit it to shoppers who had already committed, which is the opposite of who it helps.
/ Three account states
10 / Personalization strategy

The same panel, specified
for who is actually looking at it

Most personalization specs describe the signed-in case and leave the rest to whoever builds it. The category filter panel was written three times instead: anonymous, signed in, and signed in with a face scan on file. Each state gets the same category attributes and differs only in how much it already knows.

Category page filter panel for an anonymous shopper, offering facial recognition, formula, skin tone and skin type attributes, plus an option to save beauty filters that prompts account sign-up
AnonymousEverything works. The only difference is that saving prompts sign-up, and it prompts from inside the panel rather than sending her away to an account page and losing the filters she just set.Open full size
Two signed-in versions of the same filter panel side by side, one without a face scan and one with, where the scanned version shows a resolved shade match and undertone at the top and pre-selects the matching attributes below
Signed in, with and without a scanThe scanned version opens with the match already resolved and a re-scan available. Saved attributes apply but do not silently default outside the same session, so a shopper is never quietly filtered by a decision she made months ago.Open full size
Filters carry through to the product page and survive the trip back, which sounds like an engineering detail and is actually the whole promise. Personalization that resets on navigation is not personalization.
/ Two ways in, one answer
10 / Personalization strategy

The quiz and the scan stop
being competing products

Diagnostic quizzes and virtual try-on had been built as separate tools with separate results pages, leaving shoppers holding two answers and no way to tell which one the site believed. The spec folds them into one experience with two entrances, and pushes the result outward so shade matches surface on product pages instead of dying on a results screen.

Finder wireframes: an entry screen offering a face scan or quiz questions, and the results page showing a resolved skin tone and skin type match with product matches labeled by the attribute that produced them
One entrance, either routeResults are labeled with the attribute that produced them, so a shopper can see why she is being shown a product. Whichever route she takes, the newest result updates the profile rather than starting a second one.Open full size
Product page with the shade match applied and marked as a match, alongside the edit details overlay for re-scanning or adjusting formula, skin tone and undertone
Carried onto the product pageThe shade is defaulted and the interface says it is a match rather than leaving her to guess. Editing opens in place, and updating the match refreshes the shades on the page underneath.Open full size
Brands kept design flexibility inside these overlays. What they did not get was flexibility on the data model, because the point of ranking the sources was that a shopper gets the same answer across the portfolio.
/ What it produced
10 / Personalization strategy

A specification teams could
build against without me

The document was not a vision deck. It was a working spec: user goal, requirements, and wireframes for each surface, plus one data model the surfaces all answered to.

01 One data model

Four sources ranked in a stated order, so two features could never leave a shopper with two conflicting answers about her own skin. Every surface downstream inherits it rather than negotiating it.

02 Every surface in every state

Requirements written for anonymous, signed in, and signed in with a scan, on every surface that was specified. The anonymous case was written first on purpose, because it is the one that usually gets dropped.

03 An honest scope line

Ten surfaces named, four specified. Reviews, offers, chat, sampling and alternate imagery were scoped and left open, so the next person could see exactly where the work stopped and what it stopped short of.

Personalization fails as a feature and works as a system. The hard part was never the face scan. It was deciding what to believe.