The boring AR wins. The spectacular AR doesn’t.
A decade after Pokémon Go made augmented reality a household phrase, the AR that actually pays for itself in retail is the one nobody films a hype reel about: a phone camera, a lipstick swatch, a body-shape selector on a kurta page. Sourabh Paithane has spent seven-plus years inside Myntra watching which experiments stay live and which quietly get switched off. The pattern he describes is unfashionably specific — apparel try-on, beauty try-on, smartphone-first, ML models tuned on impressions and clicks rather than headsets and demos.
In sixty seconds.
The retail AR conversation has two channels and they are not the same conversation. Channel one is the headset, the metaverse store, the Apple Vision Pro demo — the kind of AR that earns press coverage and almost no purchase orders. Channel two is a phone camera pointed at a face or a body, a model running on-device or in a thin cloud round-trip, and a customer deciding whether the lipstick or the kurta is the one. Sourabh is firmly inside channel two and barely concedes channel one exists.
Myntra’s live AR is two features: an apparel try-on that lets the shopper preview a garment on a chosen body shape, and a beauty try-on that overlays a lipstick shade through the live camera. Neither is glamorous. Both are smartphone-only. Both are powered by a stack the company describes as in-house proprietary models trained on “humongous” volumes of impressions and clicks, with partnership models layered in where the build-vs-buy maths favours buy.
The framing he keeps returning to is uncharacteristically unflashy for a category that has spent a decade overselling itself. AR helps the customer “look at the product in the online setting from more nuanced ways.” That is it. Not immersion, not presence, not spatial computing. A better look at the product, on the phone that is already in the customer’s hand, on a journey that ends in a purchase that does not get returned.
Where to land in the conversation.
Each chapter opens the YouTube video at that timestamp in a new tab.
Six ideas to carry into your own work.
Mental models lifted from the conversation that travel beyond Myntra. Each one is the kind of thing you can quote in a product review on Tuesday.
Boring AR beats spectacular AR
The AR features that survive in retail are the ones that look like product photos with a little more agency — try a lipstick, see a body shape — not the ones that feel like a tech demo. Sourabh’s Myntra has shipped apparel try-on and beauty try-on; it has not shipped a metaverse store. The category’s graveyard is full of spectacle. Its survivors are utilitarian.
Return-rate prevention is the real ROI
Conversion lift is the metric AR gets sold on. Return-rate reduction is the metric AR gets paid on. Apparel returns sit at 30–40% in Indian D2C; even a single-digit improvement is months of margin. Sourabh does not say the word “returns” once until the very first answer, then keeps coming back to “the right purchase decision” — the polite phrasing for the same metric.
Smartphone is the headset that already shipped
The honest answer to the headset-vs-phone debate in India is that the phone has the volume, the camera, and the wallet behind it. Myntra is an Android-and-iOS product. AR investment that does not run on the device the customer already owns is investment in someone else’s roadmap. Sourabh says it plainly — “mintra has taken very big bold bets on smartphone as a center.”
ARKit and ARCore make capability a commodity
Since 2017, every modern Android and iPhone has shipped with a platform-level AR runtime — Google’s ARCore and Apple’s ARKit, plus the on-device ML primitives that came with Vision and Core ML. The cost of doing AR collapsed. What did not collapse is the cost of doing AR well, which is now an ML and content problem, not a platform problem. The moat moved upstream.
VR retail is a graveyard. Don’t visit alone.
The brand-store-in-the-Metaverse era of 2021–22 produced impressive Decentraland and Horizon Worlds press shots and almost no recurring revenue. Walmart shut its VR retail pilots quietly. Most of the “flagship” stores from luxury houses have empty visitor counts. Sourabh’s answer when asked about headsets is two sentences and a redirect to the phone. That is the appropriate amount of energy to spend on it.
The skill stack, not the hire
Sourabh lists five disciplines that have to ship together: product management, data science, UX design, frontend engineering, and analytics. AR retail features fail when any one of those is missing or behind. The temptation is to hire a “head of AR.” The pattern that works is treating AR as a feature of the existing PDP team, with the same cadence and the same metric ownership.
Seventeen things to actually walk away with.
Each one carries the timestamps where the moment lives, and a transferable note for work that isn’t Myntra.
The killer retail-AR app is return prevention, not engagement.
The most consistent piece of evidence from a decade of retail AR is that engagement metrics — time on page, share rate, novelty taps — do not pay for the build. Return-rate reduction does. Apparel returns in Indian e-commerce hover at 30–40%; eyewear sits in the same band before any try-on layer. A category that knocks even four points off that number is paying for itself the same quarter.
Sourabh phrases this through the customer side — “help them make the right purchase decision” — but the operational read is the same. The AR features Myntra has kept alive are the ones that reduce the post-purchase regret loop. The ones that were time-on-page boosters quietly went away.
Lenskart’s 3D Try-On is the Indian flagship — and the template.
The Indian retail-AR experience most worth studying isn’t in this conversation but sits adjacent to it: Lenskart’s 3D Try-On, built on a face-mesh capture from the front camera and a 3D library of frames. The category fit is near-perfect — eyewear has high return rates, the product sits on a face you already have a camera pointed at, and the SKU library is finite. Lenskart turned try-on into a default step in the path to purchase rather than a feature tab.
The transferable observation: AR works best in categories where the product attaches to a body part the phone can see (face, head, hands, feet), where SKUs are finite enough to model, and where the alternative is a physical store visit the customer was not going to make anyway. Anything that fits all three conditions is a category that AR can change. Anything that fails any one of them is a category where AR will be a marketing feature.
Sephora Virtual Artist set the playbook for beauty try-on a decade ago.
The beauty category got AR right earlier than apparel did, and the reason is the geometry. A face has stable landmarks — eyes, lips, cheekbones, jaw — that ML models can pick out at high accuracy on consumer hardware. Sephora’s Virtual Artist, launched in 2016 and powered by ModiFace, was the first mass-market AR try-on that worked well enough that customers returned for it rather than tolerating it. Estée Lauder bought ModiFace in 2018; L’Oréal bought it in the same year. The infrastructure layer for beauty AR has been corporate since.
Sourabh’s description of Myntra’s makeup try-on — switch on the live camera, try lipstick shades, pick — is the same primitive. The accuracy bar he names is the one ModiFace solved early: “the shade gets applied so that it doesn’t kind of spill over.” The unsexy battle in beauty AR is segmentation quality; everything else is downstream of it.
IKEA Place was the proof-of-concept the rest of retail learned from.
IKEA Place shipped in September 2017 alongside Apple’s ARKit launch and immediately became the example every retail AR pitch cited. The product idea was modest — point your phone at a room, drop a 3D model of a couch, walk around it. The execution was modest. The strategic lesson was not. IKEA had spent a decade building a CAD library of every SKU it sold; the AR feature was a re-packaging of an existing asset, not a fresh investment.
The downstream message is one the apparel and fashion world is still catching up to. Apparel does not have a CAD library of its kurtas. Building 3D garments is genuinely expensive — photogrammetry rigs, manual cleanup, fabric simulation. Myntra’s apparel try-on works around this by simulating onto body-shape avatars rather than meshing each garment in full 3D. The cost curve for full 3D apparel is falling — generative methods are now usable — but the gap between “AR works for furniture” and “AR works for apparel” is fundamentally a 3D-asset gap, not a runtime gap.
Apparel try-on is harder than beauty try-on by an order of magnitude.
The apparel try-on Sourabh describes is body-shape-aware rather than per-body-mesh — the customer picks a body shape close to theirs and previews the garment on it. That is the honest path through a problem that nobody in retail has solved end-to-end. Per-body apparel try-on requires depth-sensing or photogrammetry to capture body measurements, plus a fabric-simulation pipeline to drape the garment realistically, plus a 3D representation of the garment itself. Each layer is its own multi-quarter problem.
The interesting question is whether generative video — Sora-class models, Veo-class models, Wan-class models — collapses the cost of garment-on-body rendering by skipping the 3D layer entirely. Late-2024 demos suggest the gap to plausible video try-on is closing fast. The strategic uncertainty for retail PMs is not whether generative try-on becomes possible — it is whether it arrives in time to make the current 3D pipelines they are mid-way through building obsolete.
The metaverse retail era was a marketing budget, not a product.
The window from late 2021 to mid-2023 saw most major brands — Nike, Gucci, Walmart, Coca-Cola, the Indian carrier Reliance — ship some form of branded virtual store inside Decentraland, The Sandbox, Roblox, or Meta’s Horizon Worlds. Almost every one of those experiences is now dead, archived, or quietly de-prioritised. Walmart shut down Walmart Land and Universe of Play in March 2024. Meta’s Horizon Retail Pilots wound down. The Decentraland concurrent-user counts collapsed into the low thousands.
The diagnosis is not that “the metaverse failed” in the abstract. The diagnosis is that there was never a sustainable commerce loop in those environments — the friction of putting on a headset to browse a Nike store was always higher than the friction of opening the Nike app on the phone. The category got tested honestly and the answer was no. Sourabh’s near-zero engagement with the headset question on this episode is a recognition of that test result.
Apple Vision Pro is interesting to engineers and irrelevant to apparel buyers.
The Vision Pro launched in February 2024 with a US$3,499 price tag and the most polished pass-through AR consumer hardware ever shipped. The retail demos have been impressive — Decathlon’s Vision Pro app, J.Crew’s personal shopper experience, Lowe’s 3D product visualisations. The unit sales have been a few hundred thousand globally. Apple cut Q3 2024 production after disappointing demand.
The product is genuinely important as a long-term platform bet — Apple is buying its way into spatial computing whether the first generation sells or not, and the engineering learnings get baked into the lighter Vision device that ships in 2026 or 2027. None of that helps the apparel PM in Bangalore in 2026. The fashion-on-Vision-Pro demos are interesting in the way the first-generation iPad apps were interesting; they will matter in five years and they do not move this quarter’s numbers.
Snapchat showed that AR shopping has a real venue — and it isn’t headsets.
Snap’s Camera Kit and AR Try-On lenses have moved more retail product than every headset-based commerce experience combined. The numbers Snap discloses are uneven, but the ones that have leaked — half a billion monthly users engaging with AR features, 250-million-plus AR shopping experiences served, partnerships with Dior, Gucci, Prada, Ralph Lauren — describe a quietly large business. The interface is the same camera every Snap user already has open.
The product lesson is the same as Sourabh’s: AR retail works when it lives inside an app the customer is already using for another reason. Snap had the camera open for messaging; the AR layer is one swipe away. Myntra has the camera open for trying on a kurta or a lipstick; the AR layer is the same camera. The retail experiences that need a new app or a new device install have a steeper hill than the ones that piggyback on existing behaviour.
WebAR is the friction-free distribution channel app AR cannot match.
The split inside AR is between app-based AR (Snap, Myntra, Lenskart, your bank’s card-scanning feature) and WebAR — AR experiences that run inside the mobile browser without an install. Niantic acquired 8th Wall, the leading WebAR platform, in 2022 specifically because the format had real product-market fit for campaign-driven AR. Brands run a Coca-Cola can scan, a movie promo, a Pop-Mart blind-box reveal in WebAR because the conversion from QR code to experience is one tap.
The trade-off is real. WebAR sacrifices the higher-quality tracking and the access to on-device ML primitives that native AR has. The Myntra-class deep features — body-aware try-on, ML-tuned lipstick segmentation — sit firmly in app territory. The campaign-class features — try the new shade on a poster, scan the box to see the product hero — sit in WebAR. Pretending they are the same channel produces bad strategy decisions on both.
Indian D2C is experimenting more loudly than the conversation lets on.
The cluster of Indian retail-AR experiments is broader than just Myntra and Lenskart. Nykaa shipped AR makeup try-on in 2020 on the back of ModiFace; Tata CLiQ Luxury added AR for watches and select jewellery; Tata 1mg layered camera-based prescription scanning that uses AR primitives even though nobody markets it as AR. Reliance Trends has piloted AR in select stores. Fab India ran kurta try-on experiments. The pattern is that the AR layer arrives quietly, attached to specific catalog cuts where it pays off, rather than as a flagship announcement.
The strategic read is that India is doing the boring, productive part of retail AR while the Western press still writes about headsets. The Indian unit economics force this: the marginal customer here is on a sub-US$300 Android, has paid-data sensitivities, and is buying a kurta worth less than the cost of one Apple Vision Pro. The pressure to make AR work cheaply on a phone is precisely why the Indian version of retail AR is closer to the long-term shape of the category than the spectacle-heavy Western pilots.
In-store AR for associates is the under-told B2B pocket.
The retail-AR conversation almost always defaults to customer-facing experiences, and almost always undersells the back-of-house. Walmart and Target run associate-training simulations in Oculus headsets; Bossa Nova’s shelf-scanning robots use computer vision primitives that overlap with AR; Lowe’s associate apps surface planogram compliance on a tablet using marker-based AR. The Decathlon Vision Pro pilot is half-customer demo, half store-design tool for the regional team.
The reason the B2B pocket is healthier than the B2C one is straightforward: an associate is paid to wear the headset and the ROI lives in a measurable productivity number. A customer is not paid to wear the headset and the ROI is whatever conversion theory the marketing team is selling that quarter. The B2B pocket of retail AR is small in revenue terms but is the only part of the category with consistent positive unit economics.
The always-on AR glasses dream is real and is also fifteen years out.
The endgame the AR industry has been chasing since Google Glass in 2013 is a pair of glasses you wear all day that overlays digital content on the world. Meta’s Ray-Ban Meta partnership, launched in 2023 and updated through 2024 and 2025, has sold real units — a couple of million by mid-2025 — but the device is a camera-microphone-audio play with no display. Snap’s Spectacles 5 in 2024 added a display but cost developers US$1,200 a month. Apple’s rumoured lighter Vision device is years away. The shipping form factor that customers actually buy is sunglasses with a camera, not glasses with an overlay.
The lesson for retail is that the device generation that would make spatial commerce ambient — overlay the kurta on the customer’s body in their bedroom mirror, suggest the matching dupatta in their actual peripheral vision — is real but distant. Building a retail AR strategy on the assumption that this device arrives in 2027 is building on a forecast. Building one on the smartphone in pocket today is building on a deployed substrate. Sourabh’s explicit smartphone-first stance is the appropriate response to that gap.
Spatial computing’s slow march does not invalidate phone-AR investment.
The temptation when a Vision Pro or a Quest 3 ships is to assume the smartphone-AR investment is a stopgap. The historical pattern says otherwise. The mobile-web era did not retire the desktop web; it added a parallel surface that dominated specific use cases. The car did not retire the train. The Vision Pro will not retire the iPhone, and the spatial-commerce surface will not retire the camera-based smartphone AR Sourabh is building today.
The right mental model is that retail AR will be a many-surfaces product, and the surfaces will be additive. The phone-AR you build now is not throwaway work; it is the codebase that ports forward to the spatial device when it eventually arrives. The companies that build phone-AR muscle today are the ones that will be ready when the spatial wave actually breaks. The companies that wait for the spatial wave start from zero on both surfaces.
The training data for try-on is impressions and clicks, not lab captures.
The most quietly important sentence in this conversation is Sourabh’s description of his training data. The Myntra models are trained on “humongous amounts” of impressions and clicks — what customers tap, what they zoom on, what they bounce from — layered onto whatever baseline models the team starts with. The lab-captured ground-truth that academic ML papers use is not the relevant corpus. The behavioural corpus is.
This is the same insight YouTube’s recommendation system shipped a decade earlier and that every modern retail-ML team has rediscovered. Your data moat is the millions of micro-decisions your customers make every day inside the existing product. AR models that get tuned on that corpus are AR models that learn the actual customer’s taste. AR models that get tuned only on labelled data are models that learn the labellers’ assumptions about taste.
The build-versus-buy decision sits at the model layer, not the runtime.
Sourabh’s framing of Myntra’s stack is openly hybrid: in-house proprietary models for the parts that benefit from Myntra-specific training (catalog, customer behaviour, India-skewed body and face distributions) and external partnerships for the parts that are commoditised (face-tracking primitives, baseline segmentation, the rendering layer). The cleanly drawn line is where Myntra’s data does something a generic vendor’s data cannot.
The strategic generalisation is that build-versus-buy in AR retail is not a question about the runtime — ARKit and ARCore are free and equivalent — and it is not about the rendering — Unity, Three.js and the platform-native renderers are roughly interchangeable. It is about the model. Build the model where your data is a moat. Buy the model where commodity vendors have more data than you ever will.
3D asset creation cost is collapsing — but slower than the demos imply.
The cost of producing a usable 3D model of a real-world product has fallen by an order of magnitude in the last five years. Photogrammetry rigs that cost tens of thousands of dollars in 2018 are smartphone apps now (Polycam, Luma AI, RealityScan). Neural radiance fields (NeRFs) and Gaussian splatting collapsed the rendering side. Generative 3D — Meshy, Tripo, Stability 3D, Hyper3D, Adobe’s Substance models — is now producing usable mesh-and-texture outputs from a single photo prompt at consumer prices.
The honest gap is that “usable for a marketing screenshot” and “usable for a try-on at customer-visible fidelity” are different bars. The fashion-fabric problem in particular — how does this dupatta drape, how does this silk catch light, how does this denim crease — is still well outside the generative pipelines that work for furniture or rigid props. The cost curve is collapsing. The fidelity curve is collapsing too. The point at which they meet, for apparel specifically, is probably 2027–2028.
The Gen Z habit thesis is the only forward-looking claim Sourabh makes.
The closing line of the conversation is the most consequential one. Gen Z and millennial customers, Sourabh argues, will make AR try-on “a habit” rather than a novelty. He does not stake a year. He does not stake a percentage. He stakes a behavioural claim — that a generation that learned to shop with the camera (Instagram filters, Snap lenses, TikTok effects) treats camera-based product preview as a normal step rather than a special one.
This is the right shape of forecast. Retail AR’s adoption was never going to be a single discontinuity; it was always going to be a cohort effect. The generation that grew up with phone cameras as a primary expressive surface treats AR shopping the way the prior generation treated checking customer reviews. The companies that read this correctly invest at the boring layer — better try-on quality, better latency, broader catalog — rather than at the dramatic layer.
Lines worth keeping near your desk.
The jargon, unpacked.
Some of these will be obvious; some won’t. Skim, mark the unfamiliar, come back later.
Check what you actually retained.
Try to answer before you click. The point is to notice where the conversation is fuzzy in your memory, then return to the transcript.
Five questions worth sitting with.
No correct answers. Type into the boxes — your responses are saved locally and exportable along with your notes.
Sourabh ties his AR investment to “the right purchase decision” — politely, return-rate reduction. Which feature in your own product is currently sold on engagement when its real business case sits on a hard-currency metric you have not made visible?
Myntra’s AR is two specific features — apparel try-on and beauty try-on — rather than a horizontal “AR strategy.” Where in your roadmap are you trying to ship a category when you should be shipping one or two specific features?
The bet on smartphone-first AR is a bet on the deployed substrate over the forecast device. What forecast is your team currently betting on, and what is the deployed-substrate version of the same bet?
The five-discipline skill stack (PM, DS, design, engineering, analytics) is the unsexy answer to “why does Myntra ship AR features that stay live.” Which discipline is currently missing or behind in your equivalent build, and what is the cost of pretending the gap is not real?
Sourabh closes on a generational habit claim, not a technology forecast. Which cohort change is your product currently under-investing in because it pays back across five years instead of one?
Where to push back.
The strongest version of each disagreement, written to be persuasive — not to win.
“Apple Vision Pro changes this.”
The push: the historical pattern of new computing surfaces is that they add to the older one rather than retire it. Mobile did not retire the desktop web; it became a parallel surface. The smartphone-AR work Myntra is shipping today is the codebase that ports forward to a spatial device when one actually ships. Waiting for the headset wave is the slower path; the institutional muscle — ML pipelines, content production, design language — is what carries forward.
“AR in retail is overhyped and the engagement numbers prove it.”
The push: the skeptic is reading the wrong metric. The category that ships in retail AR — beauty try-on, eyewear try-on, furniture visualisation — sits inside specific catalog cuts where it pays off on return rates and average order values, not across the whole catalog. Lenskart, Sephora, IKEA and the Indian D2C cluster have all shipped persistent AR features that pay back. Calling the category overhyped because the horizontal hype overshot the actual product is true and irrelevant.
“VR retail will return when glasses go mass-market.”
The push: the timing problem is the thesis problem. A device that does not exist yet, at a price that does not exist yet, with battery and optics that do not exist yet, is a hardware roadmap with at least three concurrent breakthroughs. Even if each is plausible, the combined probability before 2030 is small, and the commerce loop on top still has to be built. The right posture is to keep an eye on the device side and to ship now on the substrate that already works. The patient bull is right in the limit and irrelevant for the planning cycle.
“Generative 3D collapses the asset-creation cost and opens apparel AR at scale.”
The push: cost is falling fast; fidelity for apparel-specific physics is falling slower. Fabric simulation, drape, light response on silk and chiffon, the way a dupatta moves — these are not solved by general-purpose 3D generators in 2026. The right read is that generative methods make the rigid-product categories (furniture, electronics, eyewear) cheaper now and will reach apparel quality in 2027–28. The optimist’s timeline is plausible for the next category; not for this one. Yet.
Three angles on Monday morning.
If you don’t work at Myntra, here’s what to take.
If you’re a retailer
- Pick one specific catalog cut where AR can pay off — eyewear, footwear, makeup, accessories, furniture — rather than launching an “AR strategy.” The features that survive are narrow.
- Tie the AR business case to return-rate reduction, not conversion lift. Make the controlled experiment a returns-cohort comparison, not a session-time chart.
- Ship smartphone-first. Treat headsets as a long-horizon platform bet that does not need a 2026 budget line.
- Audit which Indian D2C peers have shipped AR in your category. Lenskart, Nykaa, Tata CLiQ Luxury, Myntra, Tira and Reliance Trends have all run experiments worth studying before you build.
- If you license a third-party AR provider, license at the layer where your data is not a moat. Build your own at the layer where it is.
If you’re a product lead
- Run the five-discipline check before you commit a roadmap line: PM, data science, UX, engineering, analytics. If any discipline is missing or under-staffed, the AR feature will ship late or ship broken.
- Resist the demo-vs-feature confusion. A press-friendly AR demo and a customer-retention AR feature are different products with different success metrics. Most retail AR programmes ship demos and call them features.
- Watch the generative 3D cost curve quarterly. The category whose 3D-asset cost has just collapsed is the category where AR ships next; the category where the curve has not crossed your fidelity threshold is one to wait on.
- For apparel specifically, study body-shape-aware avatars before you commit to per-body 3D meshes. The cheaper path covers most of the customer value.
- Treat WebAR and app-AR as separate products. Campaign AR runs in the browser; conversion-loop AR runs in the app. Confusing them costs you on both sides.
If you’re an engineer
- Get fluent in ARKit and ARCore at the level where you can ship a feature without a vendor layer. The capability is platform-native; the muscle should be in-house.
- Build your behavioural training pipeline before you build the rendering pipeline. Sourabh’s humongous-impressions-and-clicks model is the moat; the renderer is commodity.
- Benchmark face-segmentation and body-keypoint quality on Indian skin tones, hair textures, and body shapes specifically. The platform models were trained on distributions that do not match your customers.
- Treat latency as a first-class metric. AR features that work in lab demos and feel sluggish on mid-tier Android phones are AR features customers turn off after the first try.
- Stay close to the photogrammetry and generative-3D tooling. The cost curve is moving fast; an engineer who knows what Luma, Meshy and Polycam can do this quarter is the engineer whose feature ships next quarter.
A decade of retail AR, briefly.
The arc the conversation sits inside, lined up.
The whole conversation, searchable.
Click a timestamp to open YouTube at that moment. Click any line to highlight it (yellow). Highlights and notes save in this browser only.
00:00 in the page to seek.