What is the signal?
According to trade reporting, Maybelline New York is the first cosmetics brand to offer real-time virtual makeup try-on inside ChatGPT. The technology is ModiFace, L'Oréal's augmented reality platform, and users connect by searching for "Maybelline" in ChatGPT's Apps section. Recommendations are built around occasion, skin undertone and personal style, and products and shades can be previewed through the device's live camera or an uploaded photograph. Four categories are covered — lips, eyes, face and brows — and a Where to Buy feature connects consumers with local physical stores and online retailers.
The groundwork was laid a few months earlier. L'Oréal and OpenAI announced their collaboration at VivaTech on 17 June 2026, and L'Oréal's own press release stated that Maybelline New York would bring Makeup Virtual Try-On directly into ChatGPT, letting consumers try makeup looks in real time through AI-powered conversation. The same announcement said OpenAI's latest model would also power CreAItech, L'Oréal's in-house generative content platform. ModiFace, which L'Oréal acquired in 2018, already provides virtual try-on and skin diagnostics to more than 200 websites in over 70 countries.
Virtual try-on itself is not new. It has been attached to brand sites, retailer apps and social commerce for years. What we read differently here is where it now sits. Until now virtual try-on opened after the consumer had already chosen a brand or a store. Inside a general-purpose conversational tool, products are called up during the question that comes before that choice — "what would suit this occasion?" The front end of merchandising has moved one step earlier.
So what changes is not the marketing channel but the way a product has to be described. A store has testers, lighting and a sales associate; a chat window has none of them. All that recommendation and virtual rendering can draw on is the structured information attached to the product, and the colour the consumer sees is a colour drawn from that information. Product data becomes the shelf, and colour data becomes the swatch.
Four limits apply when reading this signal. First, outlets date the launch inconsistently, so a single release date cannot be fixed (coverage dated to late October sits alongside items in an early-November weekly review). Second, no performance data — users, conversion, sales effect — has been published. Third, the supported regions and languages are not confirmed in public material, so whether the same experience opens in Korea has to be checked separately. Fourth, the scope is one brand's colour cosmetics, and the development judgements below are our reading of the published feature description, not a forecast that this becomes the standard.
1. What gets merchandised in a chat window is product data, not product
If the recommendation criteria are occasion, undertone and style, only products that register against those criteria get called up. In a store an evocative shade name was fine because the customer checked the tester and an associate filled in the rest. A chat window has no such backup. If it is not written as data which undertone a product suits and what finish it has, that product never enters the candidate set.
That lengthens the list of items to settle during development. Until now, formulation freeze decided colour, format, container and quantity. Now the values that describe the colour join them: measured colour values, undertone class (cool, neutral, warm), finish (matte, satin, glossy), coverage level, recommended occasion. These are not adjectives marketing adds later; they come out of the lab as measurements.
Order matters. Fix the shade name and pack copy first and fit the data around them, and the same product is listed under different classifications on your own store, a retailer and a marketplace. Confirm the measurements and classes in the same meeting that freezes the formula, and that one set of data goes into every channel unchanged. The only work a new channel adds is translation.
When developing through an OEM or ODM, decide who produces and hands over this data. Different measurement conditions, instruments and lighting references produce different values for the same product. Writing into the specification which values are supplied, measured under which conditions, is cheaper than re-measuring the product just before launch.
Finally, this does not stop at colour cosmetics. Skincare also has to be described through attributes — texture, point of use, skin type, the routine step it sits in — to become a recommendation candidate. Colour opened first because there was something obvious to show through virtual try-on, but the requirement to have describable attributes is category-agnostic.
- Whether measured colour values, undertone class, finish and coverage were set in the same meeting as formulation freeze
- Whether the order was changed so attribute data is fixed before the shade name
- Whether measurement conditions (instrument, lighting, reference) are written into the development specification
- Whether skincare items are also described by texture, point of use and skin type
2. The shade on screen is a promise
Virtual try-on draws the product's colour data onto a live image of the face, and the consumer decides to buy from that screen. A gap between screen and product is then not a matter of taste but a mismatch of expectation, and mismatched expectation comes back as returns and low review scores. In a store there was the option of swatching on the back of a hand and walking away; for someone who bought the colour they saw on screen, that option has already passed.
Colour control therefore moves from the quality checklist to the conversion column. What has to be set at formulation stage is the standard colour, the limit samples, and the permitted colour difference between batches. Without a number, "close enough" means something different to every person involved. How tight that tolerance should be bears directly on pigment cost and production yield, so it is a number development and manufacturing have to agree, not marketing.
Difficulty varies by format. Opaque matte formats are relatively simple to render, but lip glosses where transparency and shine work together, tints whose colour develops over time, and base formats whose tone shifts with oxidation and wear on skin all drift further between the single screen moment and a few hours of real use. For items designed around virtual try-on, both "immediately after application" and "a few hours later" belong in the colour specification.
Raw materials need checking too. The same pigment moves between lots, and when base yellowing or colour change over shelf life is layered on top, early and later batches diverge. Putting pigment lot control and colour-change-over-time testing into the development schedule is faster than tracing the cause after launch, once customers report the colour changed.
One operational item follows. Keeping per-batch colour measurements means that when an enquiry arrives you can tell which batch drifted and by how much. Without records there is nothing to explain with, and no point to fix so it does not recur.
- Whether standard colour, limit samples and batch-to-batch colour tolerance are set as numbers
- Whether development, manufacturing and marketing agreed that tolerance together
- Whether both "immediately after application" and "a few hours later" are in the specification
- Whether pigment lot control, colour-change-over-time testing and per-batch measurement records are scheduled
3. A shade range has to be described as a system, not a count
If recommendation runs on undertone and occasion, eight shades with clean axes may be called up more often than twenty unclassified ones. That is our reading, but the reasoning is simple: matching happens on attributes, and a shade with empty attributes cannot answer any question. Which is why investment in describing shades comes before investment in adding them.
Expanding a shade range carries cost. More SKUs split the production quantity allocated to each one, and every shade needs its own pigment combination, limit samples and stability data. COSLAB consults on projects from 1,000 units upward, and what is actually feasible depends on the formula, container, raw materials, process and specification. Quantity planning has to reflect that adding one shade adds not only inventory but one more colour specification to control.
It is safer to design the first launch set as the minimum unit that covers the axes. Cross undertone (cool, neutral, warm) with depth, plus one or two finishes, build the smallest set with no gaps, and fill the empty cells later once real sales and enquiry data exist. Lay the range densely from the start and you cannot read which shade sold, or why.
Naming rules belong in the same decision. Evocative shade names are brand equity, but they are not matching criteria. The product name or description has to carry attribute words pointing to undertone, depth and finish for search and recommendation to catch it. Combining both is simple: keep the shade name and put the attribute words into the description field in a fixed format.
For export items, names and attribute words need a second look per market. The vocabulary used for undertone and the preferred depth range differ by market, and translated shade names sometimes read differently locally. Decide the first launch markets and fix the attribute words in that market's vocabulary, and the same data can go into every channel as it is.
- Whether the first shade set is the smallest gap-free set across undertone, depth and finish
- Whether the effect of more shades on per-SKU production quantity and colour specification workload is in the quantity plan
- Whether attribute words sit in the description field in a fixed format, separate from the shade name
- Whether undertone and depth wording was fixed in the launch market's vocabulary
4. Once Where to Buy is attached, sellability becomes a specification
After the recommendation and the virtual swatch, this feature adds Where to Buy. Reporting says it connects consumers with local physical stores and online retailers. That means the person who received a recommendation checks the purchase route on the same screen, and a product that cannot be bought there leaks demand even when it is recommended.
For a brand this structure ties exposure and sellable status into one line. Until now the order was to grow overseas exposure and attach distribution afterwards. If the recommendation surface also guides the purchase route, the product has to be sellable in that market at the moment of exposure. Registration, labelling, the local responsible party and stock location all move into the product specification.
Colour cosmetics add one more layer. Colourants have market-specific permitted lists, and the usable scope is set separately for areas such as around the eyes. The colour seen on screen may be made with a colourant that cannot be used in that market. Fixing the launch market combination before the formula, and choosing colourants inside the overlap of those markets' permitted ranges, creates room for one formula to cover several markets.
The reverse order — adding markets later — becomes a job of remaking the colour. Swap a colourant and the colour changes; change the colour and the measured values, limit samples and the colour data used for virtual try-on all have to be re-aligned. The moment the colour data you have built diverges by market, management cost multiplies by the number of SKUs.
In practice, split the first launch markets into two bundles: markets where distribution is attached now, and markets where it will be attached within twelve months. Design the formula inside the intersection of the two bundles' permitted colourant ranges, and when the second bundle opens you change the label rather than the formula.
- Whether the product is sellable in that market at the moment of exposure (registration, label, responsible party, stock)
- Whether the launch market combination was decided before formulation freeze
- Whether colourants were chosen inside the overlap of the target markets' permitted ranges
- Whether the design lets a market be added without reopening the formula
5. For now this is one brand's colour cosmetics app
Not overstating the scope matters. What has been published is one brand's four colour categories, no performance data has been released, and there is no announcement that this becomes the platform's standard way of merchandising. Supported regions and languages are not confirmed in public material either.
So the place to invest now is not a specific integration but the assets any surface uses. Measured colour values, attribute classes, colour tolerance and per-batch measurement records are used by your own store, retailer apps, marketplace listings and internal quality control alike. This work is not a loss if no integration opens.
What to avoid is the opposite: widening shade ranges or SKU counts on the assumption of AI exposure. Add SKUs without demand data and inventory and colour specification workload rise first, and that cost stays whether or not the exposure arrives.
There are three things to watch: whether supported regions and languages widen, whether the same approach spreads to other brands and categories (particularly skincare using skin diagnostics), and whether the attributes used as matching criteria are disclosed. The third is the most practical. Knowing which attributes drive matching tells you which fields in your product data to fill first.
Until then the task is simple. Define the next product's colour as numbers, agree the range within which production will hold those numbers, and keep the records. When surfaces multiply, the prepared side copies its data across; the unprepared side has to measure the product again.
- Whether investment went into measured values, attribute classes, colour tolerance and records before any integration
- Whether shades or SKUs were widened in advance on the assumption of AI exposure
- Whether region and category expansion and disclosure of matching attributes are on the monitoring list
- Whether the data built now is in a format your own store, retailers and marketplaces can use as is
The first question for the brief
Delete the line that reads "respond to AI channels" and write two instead. The first is defining the colour as numbers: whether measured values, undertone class, finish, coverage and recommended occasion were confirmed in the same meeting that froze the formula. The second is the range within which production holds those numbers: whether standard colour, limit samples, batch-to-batch colour tolerance and the measurement record format were agreed jointly by development and manufacturing.
Next write the shade and market combination. Check that the first launch shades are the smallest set covering the gaps across undertone and depth axes, and that those colours were built inside the colourant ranges permitted in both the first and second launch markets. Get those two right and new surfaces do not force the formula back open.
The last line is ownership. Unless it is settled who produces this data, under which measurement conditions, and where it is stored, the data gets made in a rush just before launch and differs by surface. The most useful single line to add to the brief today is who measures your product's colour, under what conditions, and in what format it is recorded.
Sources
- L'Oréal and OpenAI join forces for Transformation in Beauty with AI — L'Oréal, 2026-06-17
- Maybelline Launches AI Makeup Virtual Try-On in ChatGPT — BeautyMatter
- Maybelline Launches Virtual Make-Up Try-On in ChatGPT — Global Cosmetics News
For more context, see the product development guide and MOQ 1,000 guide.