Tecof • September 15, 2026
How to Use AR (Augmented Reality) in E-commerce
In Brief
AR, or augmented reality, is the technology that projects a product's three-dimensional model into a real room through the phone's camera, or onto the user's own face. Its job in e-commerce is to close a single question in the customer's head: "How will this look on me, and will it fit there?" When that question closes, add-to-cart rises; when it does not, the customer either does not buy or buys and returns. AR is not a showpiece but a measurable returns-and-conversion tool in specific categories; in the wrong category it is simply an expense that slows the page down. As of 2026 the question is no longer "should we do AR" but "on which 40 products, at what cost, and for what difference in returns".
Thursday night, 22.10. A home textiles retailer has the last 90 days of returns open in the admin panel. In rugs: 1,240 orders, 214 returns, that is 17.3 percent. The reason breakdown reads: 96 "too big for the room / wrong size", 61 "colour arrived different", 34 "not what I expected", 23 other. Every return costs 74 TL in return shipping on the negotiated Yurtiçi rate plus roughly 18 TL in repackaging and inbound handling. The 96 orders that said "too big" alone mean 8,832 TL of direct logistics cost in three months, and that is before counting the 11 days each returned item spends unsellable in transit and on the shelf.
The real cause is not that the product is bad. The product page states that the rug is 160x230 cm, and there is even a size chart. But the customer cannot convert the number 160x230 into the empty space in her own living room. The measurement is on the page; the meaning of the measurement is not. That is exactly the gap AR closes: turning a number into an image.
The Three Real Uses of AR in E-commerce
Marketing decks list dozens of AR use cases; three of them make money in practice. Those three mean different technologies, different costs and different categories. Having built one does not mean you have built another.
Placing it in the room
The user points the phone at the room, the true-scale 3D model of the product settles on the floor, and she walks around it, sets it beside the sofa, pushes it against the wall. Furniture, rugs, lighting, wall art, large appliances and garden products get the most out of this one. The critical point is scale: if the model's real centimetre equivalent is wrong, AR does more harm than good, because the customer believes what she saw, and if what arrives is different the return is guaranteed.
Trying it on the face or body
The camera detects the face and lays a lipstick or foundation shade onto the lips and skin, a glasses frame onto the nose and ears, an earring onto the earlobe. In cosmetics this speeds up shade selection: a user stuck between 24 lipstick shades narrows down to two or three after trying them. In eyewear the question is literally "does this suit my face", so AR touches the buying decision directly. Be careful with skin-tone matching: the phone camera's white balance and the ambient light shift a foundation shade by two steps without effort, so position AR in foundation as a "narrow it down" tool rather than an "exact match" one.
3D 360-degree inspection
Rotating and zooming the product inside the page, without placing it in a real environment. This should be most stores' first step because it is the cheapest and it works on desktop too. For watches, bags, shoes, small appliances and furniture detail, where the question is "what does the back look like, how is the stitching, how thick is it", it replaces a photo set. Better still, the same 3D model gets reused if you later move to room placement, so this step is not a wasted investment.
Where It Makes Money and Where It Loses It
AR's return depends on a single variable: how much of that category's returns come from "I bought the wrong thing because I could not see it". If that share is low, there is nothing for AR to fix.
Break returns down by reason
Before you start, split the last six months of returns by category and by reason. "Wrong size", "colour different", "not what I expected" are AR's targets. "Arrived faulty", "arrived late", "changed my mind" are lines AR cannot touch, and for those you need to look at the warehouse and logistics side. If AR's target reasons are not at least 40 percent of a category's total returns, take that category off the list.
Shoe sizing is a separate problem
In Turkey most shoe returns come from sizing, and what AR provides here is a visual try-on, not a dimensional one. Putting a virtual shoe on a foot answers "does it suit me", not "is it a 39 or a 40". The right tool for sizing is foot scanning plus a brand-level fit chart; buy AR as a sizing fix and your return rate will stay exactly where it was and the investment is written off.
Where AR does not work
In food, cleaning products, stationery, spare parts, books, digital goods and generally in low-priced items with already-low returns and fast decisions, AR is a cost line. The 3D model production cost for these products exceeds the product's margin. If you want to lift basket size in those categories, the classic ways of raising average order value return far more than AR does.
| Category | AR use | Typical return rate | Expected reduction | Priority |
|---|---|---|---|---|
| Furniture, rugs | Room placement | 12-20% | 2-5 points | High |
| Lighting, decor | Room placement | 8-14% | 1-3 points | High |
| Eyewear | Face try-on | 15-25% | 3-6 points | High |
| Lipstick, eye makeup | Face try-on | 6-10% | 1-2 points | Medium |
| Foundation, complexion | Face try-on | 10-16% | 0-2 points | Low |
| Shoes | Visual try-on | 20-30% | 0-1 point | Low |
| Watches, bags, accessories | 3D inspection | 8-12% | 1-2 points | Medium |
| Food, consumables, stationery | None | 1-3% | 0 points | Do not |
Technology Options: What You Actually Have to Build
There are two ways to build AR, and the difference between them is less about cost than about how many people you reach. Before deciding, let us pin down five terms, because they get muddled in supplier meetings and the wrong package gets bought.
- WebAR: a setup that calls the phone's own AR engine from a button in the browser, with no app download. The widest reach and the cheapest build.
- USDZ: Apple's 3D file format. Quick Look opens this file on iOS; if you want to show AR to an iPhone user you need this file for every product.
- glTF / GLB: the standard 3D format on the Android and web side. Scene Viewer and in-browser 3D viewers use it. GLB is glTF packed into a single file.
- model-viewer: the web component you add with one tag, which handles 3D rotation on desktop and AR on mobile by itself. It puts both formats behind one button.
- Scale accuracy: how many real centimetres one unit in the model corresponds to. This is the most critical value in room placement; get it wrong and AR raises returns rather than lowering them.
WebAR: from the browser, no app download
The user taps the button on the product page, the phone's own AR engine takes over, nothing is installed. Concretely: on iPhone, Safari opens your model's USDZ version through Quick Look; on Android, Chrome opens the same product's glTF/GLB version through Scene Viewer. So you have to keep two file formats per product. The standard component that puts both behind one button is model-viewer: you place a tag on the page and it handles 3D rotation on desktop and AR on mobile itself. For a store starting out, this is the correct beginning.
In-app AR
If you have your own mobile app, AR runs more smoothly, room scanning and light estimation are better, and measurement tools can be added. But your reach is limited to users who installed the app, and if most of your traffic does not come from there the number of people who will ever see AR stays very low. We covered the general logic of app investment in mobile app or mobile-friendly site; AR on its own is not a sufficient reason to commission an app.
The marketplace constraint
You cannot run your own 3D component on Trendyol or Hepsiburada product pages; the product description HTML is restricted and external scripts are not accepted. If you sell on marketplaces, the AR investment only pays back on your own site. That turns AR into a reason to pull traffic to your own channel: an experience that does not exist on the marketplace is a concrete difference in your own store's favour.
| Option | Build effort | Reach | Formats | When to choose it |
|---|---|---|---|---|
| model-viewer (WebAR) | 1-3 days | All mobile traffic | GLB + USDZ | Almost every store's first step |
| Packaged AR service | 1-2 weeks | All mobile traffic | Managed by the service | When you want model production outsourced too |
| In-app AR | 6-12 weeks | App users only | GLB/USDZ + SDK | When app traffic exceeds 30% |
| 3D viewer only | 1-2 days | Mobile + desktop | GLB | Products that need no camera |
Where the 3D Model Comes From: Four Routes and Their Costs
AR projects stall on the model, not the technology. Producing a model for every product in a 3,000-item catalog is impossible; the job is choosing which 40 products get one.
Photogrammetry
Shooting 60 to 120 photos of the product from every angle and converting them into a 3D model with software. It works well on richly textured, non-reflective products: sofas, bags, wooden furniture. It fails on glass, chrome, polished metal and transparent items. It is the cheapest method you can set up in your own studio, but it costs 40 to 90 minutes of human labour per product, which makes 40 products two weeks of one person's time.
3D scanner
Scanning with a desktop scanner or the depth sensor in newer phones. Faster than photogrammetry, down to 10-20 minutes per product, and dimensionally more accurate. Hardware runs between 60,000 and 250,000 TL for a mid-range scanner. If new products cycle through the catalog constantly, it pays for itself past roughly 200 products.
Agency or 3D modeller
The highest quality and the highest cost. You will see prices between 2,500 and 12,000 TL per product depending on complexity. It is the right choice for hero products, bestsellers and the 10 to 20 items that are the face of the brand. It is not for the whole catalog.
Model generation with AI
Tools that build a 3D model from a handful of photos became usable across 2025 and 2026. They bring cost down to a few dollars per product and time down to minutes. They have two limits: scale accuracy is not guaranteed, so you have to correct the model's centimetre values by hand for room placement; and texture quality is usually not good enough for a hero product. The right use is generating quick models for long-tail items to support 3D inspection.
| Method | Cost per product | Time per product | Scale accuracy | Suitable products |
|---|---|---|---|---|
| Photogrammetry (in-house) | 300-800 TL labour | 40-90 min | Medium | Fabric, wood, matte surfaces |
| 3D scanner | 150-400 TL labour | 10-20 min | High | Mid-size rigid products |
| Agency / modeller | 2,500-12,000 TL | 2-5 days | High | Hero products and bestsellers |
| AI generation | 50-250 TL | 5-15 min | Low, corrected by hand | Long tail, 3D inspection |
Page Speed: AR's Invisible Bill
AR's most frequently missed cost is file size. If the 3D model is not wired into the product page correctly, the page slows down for the 95 percent of users who never open AR at all, and you lose more than you gain.
A file size budget
On mobile, the target for a product's GLB file is under 3 MB, with a ceiling of 5 MB. The raw model that arrives from an agency is often 40 to 90 MB; compression, capping texture resolution at 2048 pixels and cleaning up unnecessary polygons are mandatory. Hold the same target on the USDZ side. This is learned once and repeated on every model, so write the size limit into your model delivery spec.
Lazy loading is not optional
The 3D model must not download while the page loads; it should download when the user taps the button. On first load the page should carry only a static poster image. That single rule brings AR's effect on Core Web Vitals close to zero. A properly built AR should not change the page's LCP; if it does, the build is wrong.
The mobile data reality
A significant share of Turkish mobile traffic browses on cellular data. A 4 MB model takes 3 to 6 seconds to arrive on 4G, and if you show the user nothing during that time she taps the button a second time or leaves. Put in a loading state with a percentage indicator; this small detail moves the AR completion rate considerably.
Measurement: How You Prove AR Paid Off
An AR investment usually closes with "that looks nice", and a year later nobody can say what it earned. Four numbers prevent that.
AR view rate
Of the people who viewed the product page, how many tapped the AR button? In a healthy store it runs between 3 and 12 percent on mobile. Below 1 percent the problem is not AR but the button's placement and its label: buried under the gallery, nobody sees it. Replacing "View in AR" with wording that says what will happen, such as "See it in your room" or "Try it on your face", lifts taps noticeably. Keep a second number beside it: of those who tapped, for how many did the model actually load. That completion rate should sit above 75 percent; if it falls below 50 percent the problem is file size or device compatibility, not interest.
The conversion gap among AR viewers
Users who open AR typically convert 1.5 to 3 times better than those who do not. Do not report that number as it stands; a user who opens AR is already a more engaged user, which is selection bias. To see the real effect, turn the AR button off for half your traffic and measure it with an A/B test. Split by user, not by session.
The returns gap
This is where the money is. Compare the 60-day return rate of orders bought with AR opened against orders bought without. Do the measurement by writing a flag onto the order line rather than trying to reconstruct it in analysis later. In categories like furniture and eyewear a 2 to 5 point gap is ordinary, and that gap alone can cover the entire cost of AR.
Misleading metrics
Time spent in AR and the number of model rotations produce interesting reports but decide nothing. If time on page rises while conversion stays flat, AR is providing entertainment, not sales. To see the whole picture on the conversion side, look at the other levers for raising conversion rate as well; AR is only one of them, and in most stores the most expensive one.
A 30-Day AR Pilot and the Payback Calculation
Starting AR with the whole catalog is the most expensive mistake available. The timeline below lets you decide on data and caps the total risk at 40 products.
Days 1-7: pick the category and the 40 products
Pull the last six months of returns by reason, and find the category where AR's target reasons have the highest share. List the 40 products in that category that contribute most to revenue. No models are produced this week; you only write down the current return rate, category revenue and product page traffic. Those numbers become your comparison baseline later.
Days 8-14: finish the technical build on 10 products
Produce GLB and USDZ models for 10 products, get them under 3 MB, and add the model-viewer component to the product page with lazy loading. Test on at least three different iPhone and Android devices. This week's goal is one product opening fast and at the correct scale, not 40 products.
Days 15-21: produce the remaining 30 models and start the test
Complete the remaining models with your chosen production method, and verify every one of them against the real dimensions in your product data. Open the AR button to 50 percent of traffic and write the AR flag onto the order line. Do not keep producing models before measurement has started.
Days 22-30: read the numbers and decide
Read the AR view rate and the conversion gap, and accept from the outset that the returns gap will take 60 days. If the view rate is under 1 percent, fix the button before you expand model production. If there is no conversion gap and the returns data shows no meaningful signal by day 30, closing the pilot is a legitimate decision.
| Line item | 40-product pilot | 400-product rollout |
|---|---|---|
| 3D model production | 20,000-80,000 TL | 200,000-800,000 TL |
| Technical build | 15,000-40,000 TL | No additional cost |
| Annual service / maintenance | 0-30,000 TL | 30,000-120,000 TL |
| First-year total | 35,000-150,000 TL | 230,000-920,000 TL |
Calculate payback like this. Say your pilot category does 4,000 orders a year at a 17 percent return rate, and your all-in cost per return is 92 TL. Today that is 680 returns and 62,560 TL of direct return cost a year. If AR cuts the return rate by 3 points to 14 percent, 120 returns disappear and 11,040 TL is saved directly. On top of that comes the shelf time of stock that no longer goes unsold, plus the extra revenue from the conversion lift. On that arithmetic a 35,000 TL pilot pays back in two years; a 150,000 TL pilot does not. Which is to say, your decision is made by looking at your own returns cost, not at AR.
Here is the job for tomorrow morning: open the last six months of returns, put them into a single table broken down by category and reason, and for each category write the share that "wrong size, colour different, not what I expected" makes up of total returns. If there is exactly one category where that share passes 40 percent, that is the address of your AR pilot; if there is none, spend this year's budget elsewhere. On a platform where 3D and AR components arrive ready on the product page the technical build line drops close to zero and the decision reduces entirely to model production cost.
Frequently Asked Questions
Does the customer have to install an app to use AR?
No. With WebAR, Safari on iPhone and Chrome on Android open the model using the phone's own AR engine. In-app AR is smoother but only reaches users who installed the app, so in reach terms the web route is almost always more profitable.
How many products need AR before it makes a difference?
The right 30 to 50 products make more difference than the wrong 500. Pick the items that contribute most revenue and have the highest returns. Rather than increasing the product count, fix the view rate and the model quality first.
Do I need a 3D model for my entire catalog?
No, and most stores do not. Typically 5 to 10 percent of the catalog carries half the revenue; make the investment there. For long-tail items, simple AI-generated 3D models are enough for 360-degree inspection.
Will AR wreck my page speed?
Not if it is built correctly. The model should download only when the button is tapped, and the page should carry a static poster image initially. A build that downloads a 3D file on page load damages LCP badly and costs you more than AR gains.
Can I run AR on my Trendyol and Hepsiburada product pages?
No. Marketplace product description fields do not accept external scripts or custom components. AR only pays back on your own site, which in turn gives you a concrete difference the marketplace cannot match when pulling traffic to your own channel.
Can I show foundation shades accurately with AR?
Do not promise an exact match. The phone camera's white balance and ambient light shift skin tone easily. Position AR in foundation as a narrowing tool that takes the customer from 20 shades to 3, not as a "this is your shade" tool; set the promise correctly and you avoid the disappointment too.
Does AR solve the shoe sizing problem?
It does not. Virtual try-on answers "does it suit me", not "which size". For sizing returns, foot measurement, brand-level fit charts and size memory from previous orders are far more effective.
How large should my 3D model files be?
For mobile, the GLB target is under 3 MB with a 5 MB ceiling. Raw models from an agency can be 40 to 90 MB; capping texture resolution at 2048 pixels and reducing polygon count is mandatory. Write the size limit into your model delivery spec, because fixing it afterwards is expensive.
How do I prove AR's real return?
The only valid method is an A/B test: turn the AR button off for half your traffic, split by user, and compare conversion and return rates. "AR viewers convert 3 times better" is selection bias and cannot be used for an investment decision.
When should I stop the pilot?
Stop when the view rate is still under 1 percent after fixing the button, the A/B test shows no conversion gap, and the 60-day returns gap is under 1 point. AR genuinely does not pay off on some catalogs, and accepting that is cheaper than rolling production out to 400 products.