The short answer: three signals rule most vendors out before budget ever comes up

Prepare three things before you talk to anyone. The source model at real-world scale, with clean topology and parts named the way your catalog names them. The textures at their original resolution, with a note on where each one came from and who owns it. And a single page describing what has to be interactive: what spins, what opens, what changes color, what has to match a SKU. Then use that package as your screening tool. Three signals rule most vendors out on the first call: they ask for your budget before they ask what format your model is in, they cannot name a target file size or load time for the finished page, and they quote a fixed price without having opened your files.

The numbers behind those signals are public. The median web page weighed 2,652 KB on desktop and 2,311 KB on mobile in October 2024, with images accounting for 1,054 KB of the desktop figure (HTTP Archive Web Almanac 2024, Page Weight chapter). A product assembly exported straight out of CAD routinely exceeds that on its own, before a single line of your site's own code loads. Compression closes the gap: on the Khronos samples, Draco geometry compression took the Buggy from 7.6 MB to 0.82 MB and cut a 738 MB city tileset to 179 MB, with load time falling from 18.9 seconds to 10.5 (Cesium engineering blog, April 2018).

That is the conversation a competent web 3D vendor starts with. A maintainer of Google's model-viewer states the working rule in the project's own discussion thread: with proper compression it is rarely necessary to use more than a few megabytes for high-quality mobile rendering, and anything over about 20 MB should be treated as a problem (google/model-viewer discussion #2716). A vendor who has never had to hit a number like that will discover it on your budget.

Speed is not a vanity metric here. Renault found each one-second improvement in Largest Contentful Paint came with a 13 percent higher conversion rate, across 10 million visits in 33 countries (web.dev case study, December 2020 to March 2021). The BBC measured losing an additional 10 percent of users for every extra second its site took to load (web.dev, "Why does speed matter?"). A heavy model does not just look slow; it removes visitors before they see it.

Weight the criteria: pipeline first, showreel last

Most buyers weight a 3D vendor on how the demo reel looks. That is the criterion that transfers least, because a reel is rendered offline with no budget at all. Weight the things that decide whether your assets survive the trip into a browser.

The pipeline criterion carries the most weight because it is where the money goes. Your CAD file describes manufacturing intent: every fillet, every internal fastener, every duplicate part in an assembly tree. A browser needs a low-polygon mesh with baked detail, packed UVs and glTF materials. If the vendor assumes you will supply "a clean model" and you assume they will make one, that assumption becomes a change order in week three.

The published-brackets criterion is easy to test, and it is worth an example. Todor3D, a Culver City, California studio founded in 2020 with 40-plus engineers across three continents and 300-plus projects delivered, publishes three budget brackets — $10,000 to $50,000, $50,000 to $100,000 and $100,000 to $250,000-plus — with matching timelines of 4 to 10 weeks, 3 to 6 months and 4 to 12 months, and names its stack as WebGL, Three.js, React Three Fiber and WebAR/WebXR; you can read more on their site. What matters is not the figures but that they exist in public: you can place your project in a bracket before the first call and learn at once whether you are in the same conversation. A studio founded in 2020 also has a shorter track record than vendors shipping browser 3D for fifteen years — the kind of trade-off a published bracket lets you weigh openly.

For context, most software companies listed on Clutch report charging $25 to $49 an hour, US firms most commonly $50 to $99, with the typical project between $10,000 and $49,000 (Clutch pricing guide, updated 20 September 2026). Real-time 3D work sits at the upper end of its regional band.

Ten questions for the first call, and the answers that should worry you

Send your file inventory before the call, then work down this list. The point is not to test their knowledge. It is to find out whether they have done this particular kind of work before.

• What format do you want my files in, and what happens first? A good answer names your format, asks about units and scale, and treats decimation and UV work as their job.
• What file size and load time are you targeting? Any number is better than none. "A few megabytes on mobile" is a real answer.
• Which compression will you use, and where does it fall down? Listen for Draco for geometry, KTX2 for textures, and a caveat about when each is the wrong choice.
• Which device will you test on? A named mid-range Android, not "modern browsers".
• Who does the modelling: you, a subcontractor, or a marketplace? All three are fine. Not knowing is not.
• What do I get at handover? Source models, textures, export scripts, and documentation on adding the next SKU.
• Who owns the models, and what licences ride along? Ask specifically about marketplace assets, HDRI environments and fonts.
• How do I add product number 51 next year? If the answer is "send it to us", ask what that will cost per model.
• What does this integrate with? Your PIM, ecommerce platform, CPQ or ERP. Integration is usually more work than the 3D.
• What have you built that I can open right now on my phone? Two live URLs, ideally in your category.

On the second call, ask the same questions of a different person on the team. Consistent answers beat impressive ones.

Seven red flags, and the invoice each one turns into

They never ask about units or scale. Models arrive in millimeters, centimeters, inches or unitless numbers, and assemblies often sit thousands of units from the origin. A vendor who does not raise this discovers it after the viewer is built, and re-anchoring a 200-part assembly is not a five-minute fix.

No performance budget at any point. This ends with a beautiful page a mid-range phone cannot load. You pay twice: once for the build, once for the optimization pass that should have been the plan. The Renault and BBC figures above are what that second invoice buys back.

A fixed price before anyone opens your files. Nobody can price mesh cleanup sight unseen. A number quoted blind is either padded for the worst case or destined to become change orders. Ask for a paid discovery on two representative SKUs instead, with a per-model price that follows from it.

The portfolio is all rendered video. Stills and turntables come from offline renderers with unlimited budgets. They prove artistry, not that anyone on the team has shipped an interactive model to a browser. Ask for links; if none exist, you may be hiring a visualization studio for an engineering job.

"We'll just generate the models with AI." Generative 3D is genuinely cheap: Tripo's own pricing works out to about $0.16 per model on its $20 Pro plan and $0.09 on the $90 Max plan, while an outsourced, human-checked product model from Modelry starts at $42 (vendor pricing pages, checked 20 September 2026). That nominal gap of roughly 260 to 470 times is real and also misleading, because generated meshes usually need retopology and UV cleanup before they can carry material swaps or exact dimensions. Generation fills a background shelf; it does not produce the hero SKU a customer configures.

Compression treated as a switch to flip at the end. KTX2 textures save GPU memory dramatically: on the Khronos StainedGlassLamp sample, 13 MB of JPG and PNG textures occupied 96 MB of GPU memory against 21 MB for the KTX2 version, and the Duck sample drops from 1.5 MB to 277 KB, about 82 percent (Khronos Group, KTX Artist Guide). The same guide is explicit that file sizes stay similar to JPEG and PNG, and engine documentation warns that UASTC textures can come out up to three times larger than the source over the network (Soft8Soft Verge3D manual). A vendor who promises compression always shrinks downloads has not measured it.

Vague handover terms. If the contract does not name the deliverables — source models, texture sets, export configuration, documentation — you have bought a rendering, not an asset. The cost lands a year later, when a second vendor has to rebuild everything to change a color.

A portfolio shows the render; open the network tab instead

Evaluate live work the way a visitor would. Open each link on your phone, on cellular data, and count the seconds until something appears. Then open it on a laptop with the browser's network panel visible and check the size of the largest file. Rotate the model and watch for popping textures and stutter. Try a two-year-old Android if the office has one — that is the device your hardest-to-convince buyer is holding.

Ask what the vendor actually did on each piece. Web 3D projects often involve three parties: the modelling shop, the development team and the client's own front-end developers. A straight answer about scope reassures more than a claim of total ownership.

Reviews deserve the same skepticism. On directory profiles, read the longest reviews rather than the average, and look specifically for what happened when scope changed, how deadlines moved, and whether the reviewer names the same people the vendor named to you. Sample size matters more than score: Todor3D's Clutch profile, for instance, carries 25 reviews at a 5.0 rating, which is a small enough sample that the individual write-ups tell you far more than the headline number. The same is true in the other direction — a single detailed two-star review about a missed integration deadline is worth more than fifty short five-star ones.

Finally, ask for a reference whose project has been live a year: the question is not how the launch went, but what maintenance has cost since.

Match the vendor type to the job, not to the showreel

There is no single category called "3D developer". These are different businesses with different cost structures, and picking the wrong type costs more than picking a mediocre example of the right one.

Marketplaces fill the gaps around your products. CGTrader lists more than two million models and Sketchfab more than eight million (vendor figures, checked 20 September 2026) — a sensible source for props and environments, never for your own SKUs, and always with the licence read first.

Questions buyers ask before their first 3D project

What file format should I hand a web developer?

Give them your native source file — the SolidWorks, Rhino, 3ds Max or Blender original — plus textures at full resolution. glTF, usually as a .glb, is the browser delivery format, not something you should convert to yourself. An already-decimated export removes detail the vendor may need. Include a note on units and on which parts must stay separate.

How large should a 3D model on a web page be?

Aim for a few megabytes, not tens. Google's model-viewer maintainers describe a few megabytes as enough for high-quality mobile rendering with proper compression, and treat files above roughly 20 MB as a problem. The median full web page was about 2.3 MB on mobile in 2024 (HTTP Archive Web Almanac). Your model is one element on a page, so it should not outweigh an average page several times over.

Should I use AI-generated 3D models to save money?

Not for products customers will configure or measure. Generation runs at roughly $0.09 to $0.16 per model on Tripo's published plans, against $42 and up for an outsourced human-checked model from Modelry. The gap closes once you account for retopology, UV work and dimensional accuracy. Generated assets suit background objects and concept exploration.

Do I own the source files when the project ends?

Only if the contract says so. Name the deliverables explicitly: source models, texture sets, export or conversion scripts, and documentation for adding the next product. Ask separately about third-party assets, because a purchased HDRI environment or marketplace prop may carry a licence that does not transfer. That clause is the difference between owning an asset library and renting one.

How long does it take to prepare assets before development starts?

For a handful of SKUs with usable CAD, expect days to a couple of weeks of preparation alongside early development. For a catalog, treat preparation as its own phase with its own budget. Published timeline brackets in this market run from roughly 4 to 10 weeks for a contained build up to 4 to 12 months for a large integrated system, and asset preparation usually decides where in that range you land.