Warby Parker's Virtual Try-On Is the Most Misunderstood AI Project in Mid-Market Retail

THE CRAFT

Warby Parker launched its virtual try-on feature in February 2019. Open the app, point your phone camera at your face, and the app overlays a 3D render of any frame in the catalog onto your face in real time. At launch, the tech press covered it as a gimmick that worked. It was the first major retail use of Apple’s TrueDepth facial-mapping API, the hardware that Apple had shipped with the iPhone X for Face ID and that suddenly had a commercial application that wasn’t security. Virtual try-on was, in early 2019, a cute thing that made you look like a dork in your own kitchen while deciding whether a pair of frames suited you.

Seven years later, it is one of the most operationally important AI projects in mid-market retail, and the consumer feature is the smallest part of the value. Every founder-CEO reading this needs to understand why, because the mistake that most of the retail operators who copied the feature have made is to copy only the visible part — the consumer convenience — without copying the part that actually generates the compounding return. I am going to walk you through what Warby Parker is actually doing with it, and what you should take from that for your own AI roadmap in 2026.

Warby Parker is a $771M (FY2024) publicly traded (NYSE: WRBY) direct-to-consumer and physical-retail eyewear company. It was founded in 2010 by four Wharton classmates: Neil Blumenthal, Dave Gilboa, Jeff Raider, and Andrew Hunt. Blumenthal and Gilboa are still co-CEOs. The company has gone from a dorm-room hypothesis (that prescription glasses were wildly overpriced because one company, Luxottica, owned the entire supply chain) to one of the most credible mid-market retail brands in America, with about 280 stores, a growing eye-exam and contact-lens business, and a platform that keeps quietly adding adjacent product lines. It is, for the purposes of this newsletter, almost exactly the operator you should be studying: founder-controlled, mid-market, traditional category (glasses have existed for 700 years), but operated with a level of technical sophistication that most of its peers have not matched.

And the virtual try-on is the cleanest example of the gap.

THE OPERATOR

What the consumer sees

Open the Warby Parker app. Tap “Try on at home” or use the in-app try-on button. The camera turns on, the app asks permission to use TrueDepth, and a real-time 3D render of any frame in the catalog appears on your face as you move your head. You can see how it sits on the bridge of your nose, whether the temples clear your ears properly, whether the lens shape works with your jawline. You can screenshot it, send it to your partner, and ask whether they think you look like a graduate student or a mid-level attorney. This is the feature. It is beautifully executed. It works on all modern iPhones and, through a more approximate model, on newer Android devices. In the consumer press it gets written up every year as a “convenience.” That is the correct but shallow read.

What the company sees

Every time a customer tries on a frame using virtual try-on, the app captures (with the customer’s permission, within the terms of use) an anonymized but extremely high-quality facial geometry profile — the shape and dimensions of the customer’s head in three dimensions. Not a photo. Not a face-ID recognition signature. The actual measurements of bridge width, temple distance, pupillary distance, nose angle, cheekbone projection. The app is designed so that the 3D render fits correctly, which means the model has to know the face’s geometry with high precision. That is a dataset.

Now multiply it by scale. Warby Parker has processed tens of millions of virtual try-on sessions over seven years across a customer base whose age, gender, and geographic distribution closely mirrors the U.S. optical population. No other company in the category has anything remotely close to this dataset. Luxottica does not have it — their stores are in malls and they measure faces with plastic rulers. The big retail optical chains do not have it — they are not running native apps with TrueDepth integrations at this scale. The DTC eyewear startups do not have it — they are newer, smaller, and mostly came up after virtual try-on had stopped being interesting to the tech press. Warby Parker owns the biggest proprietary dataset of American face geometry in consumer retail, and the tech press is still writing about the feature as if it were a convenience.

Here is what that dataset enables, in rough order of how visible each application is to an outside observer:

1. Better fit recommendations. When a new customer opens the app for the first time and scans their face, the system can instantly return “the twelve frames most likely to fit you well, ranked by goodness-of-fit against a model trained on millions of prior fit outcomes.” This is the visible application. It also produces higher conversion, lower return rates, and higher customer satisfaction — measurable numbers that show up in the quarterly report.
2. Product design feedback. When the design team releases a new frame, Warby Parker can simulate how well it will fit the actual American face distribution before they order the first production run. Not hypothetically. Not with a ten-person focus group. Across millions of faces. If the new frame turns out to fit only a narrow range of bridge widths, they know that before the factory retools. This is a multi-million-dollar benefit per year on inventory risk alone.
3. Inventory allocation by store. Different metros have different face-geometry distributions — partly because of the ethnic composition of the customer base, partly because of subtler factors. The company can allocate frame inventory to individual stores based on the expected face distribution of that store’s customer base, not a national average. That is a straight improvement in sell-through rates at the store level.
4. Category expansion. The face-geometry dataset is not limited to glasses. It is equally useful for contact lenses (a major category for Warby Parker, with its own fit and sizing problems), for prescription sunglasses, for at-home refraction tools, and, increasingly, for non-glasses categories where face shape matters. Every adjacent category the company enters starts with an unfair advantage that its competitors cannot easily replicate.
5. The long game: AI-assisted optometry. The exam-room side of eyewear — the part where a licensed optometrist gives you a prescription — is the category’s biggest structural cost and its biggest structural moat. Warby Parker is investing heavily in tools that combine the face-geometry data with other visual and optical inputs to assist (not replace) the optometrist in producing a faster, more accurate prescription. This is where the dataset compounds into a structural advantage over the next decade.

Notice the pattern. The dataset was a byproduct of the consumer feature. The consumer feature was built to sell glasses. The dataset, once it existed, turned out to be the more valuable asset — and the feature keeps producing more of it every day, at essentially zero marginal cost, forever.

The Craft of AI read

The lesson is simple and completely general, and I think it is the most important lesson in this issue.

The value of an AI-enabled consumer feature is not the feature. It is the data the feature produces, and the compounding use of that data across every other decision the company makes.

Most founder-CEOs evaluating an AI-enabled consumer feature in 2026 are going to ask the wrong question. They are going to ask: “Will this feature drive conversion?” or “Will this feature improve customer satisfaction?” Those are the right questions for a traditional software feature. They are the wrong questions for an AI-enabled one. The right question is a two-parter:

1. What proprietary data will this feature generate as a byproduct of its use?
2. Can that data be used to improve decisions in parts of the business the feature does not directly touch?

If the answer to both is yes, the feature is a data asset disguised as a software feature, and it should be evaluated like an asset — which usually means prioritizing it much higher than the traditional-software evaluation would. If the answer to either is no, the feature is just a software feature, and it should be evaluated like one — which usually means lowering its priority relative to the hype around it.

This is how the Warby Parker team has been thinking about technology investments for at least the last five years, and it is why their AI spend produces a higher return per dollar than most of the retail operators in their category. It is not that they are smarter about AI. They are smarter about data as a byproduct of the customer experience, and the AI is the tool that operates on the data once the data exists. The feature was always the data collector. The AI was always going to come afterward.

Three specific decisions for founder-CEOs to steal:

Steal #1: Audit every customer touchpoint for byproduct data opportunities. Every time a customer interacts with your company — a form they fill out, a photo they upload, a configurator they use, a tool they try — there is a latent dataset that could be collected if you designed the interaction around capture rather than convenience alone. That dataset, compounded across time and customers, is the asset your AI roadmap will eventually need. Find it now.

Steal #2: Treat consumer AI features as asset-building projects, not feature projects. When a vendor pitches you a “customer-facing AI feature,” rewrite the pitch as an asset: “what proprietary dataset will this project accumulate over three years that no competitor will have?” If there is no dataset, the feature is a cost, not an investment. Treat it accordingly.

Steal #3: Put the data strategy one level above the AI strategy on the org chart. The classic mistake is to hire a head of AI who reports up to the CTO. That wires the organization for features. The better wiring is a head of customer data strategy, reporting up to the COO or the CEO directly, with a mandate that crosses every function. The AI roadmap becomes a subcomponent of the data strategy, not the other way around. This is a harder reorganization than most boards are willing to authorize, but it is the one that produces the Warby Parker pattern.

Things to consider

1. What is your company’s virtual try-on? Not literally. The functional analogy: what is the one customer interaction that, if you redesigned it slightly to capture data that you currently throw away, would produce a proprietary dataset worth more than the interaction itself? Every founder-CEO in a traditional category has one of these hiding in plain sight.
2. Who owns the byproduct data at your company? Not “who owns customer data” — that is usually a marketing-team answer about email list management. I mean the byproduct data — the stuff that is generated as a side effect of how customers use your product or service. If the answer is “nobody owns it,” then nobody is going to build anything on top of it, because it does not exist in any organized form. That is the gap.
3. Which of your AI-feature evaluations in the last year were asset-building projects in disguise, and which were pure cost centers? Go back through the last three AI proposals your team brought you. Rewrite each as an asset proposal. The ones that still make sense after the rewrite are the ones worth funding. The ones that do not are the ones you should have passed on.
4. What data does your most important competitor not have that you could plausibly own exclusively in three years? This is the question Warby Parker was able to answer with “face geometry” in 2019. Every other optical retailer now lives with the consequences. What is the equivalent question in your category, and who is going to answer it first? If it is not you, it should be.
5. Are your consumer-facing product decisions being routed through your data team, or around them? In a company that has correctly understood the Warby Parker pattern, product and data are the same conversation. In a company that has not, product ships features and data reports on what happened. That structural difference is worth more in 2026 than almost any specific AI tool you could buy.

THE WORKBENCH

Here’s the tactical takeaway for your Monday.

Spend 30 minutes this week on one exercise I’ll call the “byproduct audit.” Sit down with the two or three people in your company who know the most about how customers actually use your product or service day-to-day. Ask them a single question:

“What does a customer give us in the normal course of using us that we currently do nothing with?”

Write down every answer. Do not filter. Some of the answers will be obvious and already exploited (names, emails, order history). Most will not. You are looking for the ones where the collective reaction in the room is “huh, yeah, we could probably use that somehow” followed by a pause. The pause is the signal. The pause means nobody has ever tried to turn the input into an asset. That is your byproduct gap, and it is the single most valuable conversation you can have with your team in Q2 2026.

Once you have the list, rank the items by two dimensions: how proprietary is it? (could any competitor collect the same thing easily?) and how cross-functionally useful is it? (does it improve decisions in more than one part of the business?). The top-right quadrant — high proprietary, high cross-functional — is where your next real AI project lives. Not the project your vendor pitched you. The project your own company is already half-building and does not know it.

This exercise takes 30 minutes. It will produce more value than the next AI vendor demo you sit through. I promise.

THE QUESTION

Here is what I honestly want to know.

What is the piece of data your company accidentally collects that you have always known was interesting but never figured out what to do with?

Every company has at least one. For most, there are several. They are usually hiding inside a workflow that was built for another reason — a form, a photo upload, a configurator, a preference center, a returns reason dropdown, a survey, a call transcript. You know the one. You have probably made a joke about it in a meeting. You may have gestured toward it in a planning session and then moved on, because turning it into an asset felt too hard or too far from the core work.

Hit reply, or send a note to grant@thecraftofai.com and tell me what yours is. No commitment, no shared project, no pitch. I am collecting these because I want to test a theory: that the biggest unused AI opportunity at most $100M–$500M founder-led companies is sitting inside data the company is already collecting and already paying to store, and has simply never been organized around. The Warby Parker pattern is the proof of concept. The question is how many of you already have the data, and just need someone to name it.

I read every reply.

— Grant grant@thecraftofai.com