The short answer: three signals that rule most vendors out

Three signals remove most of the field before you read a single proposal: the vendor cannot open a working configurator on your phone during the first call, cannot describe how your CAD or product data will get into a browser, and cannot name the system that receives the order once a customer finishes configuring. Apply those three and a longlist of twenty shrinks to five or six. The stakes are ordinary software money: the typical software development project reviewed on Clutch costs $10,000–49,000, and the average total engagement runs $132,480 over roughly 13 months (Clutch pricing guide, data from client reviews, updated September 2026). A configurator that fails any of the three signals costs the same as one that works.

Signal one: a live demo on your own phone. Ask the vendor to send a link, not a video, and open it on the phone in your pocket over cellular. 3D fails at the performance layer more often than at the visual layer, and a phone over 4G is where it fails first. The commercial cost of that failure is documented outside the 3D world: Renault measured a 13% conversion lift for every one-second improvement in Largest Contentful Paint across 10 million visits in 33 countries, and a 14-point drop in bounce rate once LCP fell under about 1.6 seconds (web.dev case study, 2021). A median mobile page already weighs 2,311 KB (HTTP Archive Web Almanac, 2024); one uncompressed product model can weigh more than the rest of the page combined. If the vendor's own showcase stalls on your phone, yours will too.

Signal two: a clear data path. A configurator is a pipeline with a viewer on the end. Somewhere between your engineering files and the customer's screen, someone has to reduce CAD geometry, rebuild materials, compress textures and set up the option logic. A vendor who has done this asks about your file formats, part counts and material libraries before quoting. A vendor who says "just send us whatever you have" either has not done it before or plans to do it by hand, once, with no repeatable process for the next product line.

Signal three: a named destination for the order. Every configurator produces something at the end: a SKU, a bill of materials, a price, a quote request. The vendor should ask where that goes—Shopify cart, Salesforce CPQ, SAP, a dealer's inbox—in the first conversation. If they do not, the configurator will ship with its own price table, and you will have created a second source of truth. Buyers already notice this kind of gap: 69% of B2B buyers report inconsistencies between what a supplier's website says and what its sellers tell them (Gartner, survey of 632 B2B buyers, 2024).

The rest of this guide is the longer form of those three tests: weighted criteria, a first-call question list, the red flags that predict a bad outcome, how to read portfolios and reviews, and a table matching project types to vendor types.

Eight evaluation criteria, weighted by how often they decide the outcome

The weights below add up to 100. They reflect a simple observation: configurator projects rarely fail because the 3D looked wrong. They fail because the assets were too heavy, the option logic did not match the factory's rules, or the integration was scoped as an afterthought. Weight your scorecard toward the parts that break.

Criterion 1: track record on configurators, not on 3D in general. A studio with beautiful WebGL brand campaigns has solved rendering, not configuration. Configuration means dependencies: option A removes option B, changes the price of C and alters the geometry of D. Ask for three shipped configurators and, for one of them, a walkthrough of how the rules were expressed and who maintains them. Platforms make this easy to check because their customer pages are public. Expivi, a Dutch 3D CPQ platform, describes a trailer configurator for Eduard with more than 100,000 combinations and a Crocs shoe customizer with automated manufacturing output. Salsita, a custom-build studio in Prague with an Atlanta office, publishes its configurator work with dealer quoting, real-time pricing and CAD/BOM output. Either kind of evidence is what you are looking for. Logos alone are not.

Criterion 2: the asset pipeline. This is where most of the hidden cost lives. Engineering CAD is built for manufacturing tolerances, not for a phone's GPU. A vendor should be able to explain, in plain terms, how geometry gets reduced and compressed and what the target file size is. The reference numbers are public: Draco compression took the geometry of a Khronos sample vehicle from 7.6 MB to 0.82 MB, about 89% smaller (Cesium engineering blog, 2018), and KTX2 texture compression cut GPU memory on a sample model by about 82% (Khronos Group, KTX Artist Guide). A maintainer of Google's model-viewer component has said it is "rarely necessary to use more than a few MB" for high-quality mobile rendering and that files over about 20 MB should be treated as problematic (google/model-viewer GitHub discussion #2716). A vendor who cannot quote comparable numbers for their own shipped work has not measured them. One caveat: some compressed texture formats can be up to three times larger over the network than the source image, so "we compress everything" is not the same as "it downloads fast" (Soft8Soft, Verge3D manual). Companies whose whole business is this step exist—VNTANA, in Van Nuys, California, converts Siemens NX, SolidWorks, CATIA and STEP files into GLB and USDZ at scale—and a configurator vendor who knows the pipeline will know when to recommend one.

Criterion 3: integration depth. Ask one question: in your last project, which system held the price? If the answer is "the configurator," you are looking at a demo tool. If the answer is "Salesforce CPQ" or "the customer's SAP instance," you are looking at a sales tool. Threekit, the Chicago platform founded in 2005, positions its configurator explicitly as a layer over existing CPQ, ERP and commerce systems, listing Salesforce CPQ and Oracle CPQ integrations. Tacton, in Stockholm, sells configuration with order handover to manufacturing and names SAP, Oracle, Microsoft Dynamics 365 and Salesforce among its connectors. Those are the platform answers. For a custom build, the equivalent is an engineer who describes the API contract between the configurator and your systems before the visual work begins.

Criterion 4: pricing and timeline transparency. The test is not whether the vendor gives you a number on the first call—nobody responsible does—but whether they publish anything about how they price at all. Todor3D, a 3D and immersive studio founded in 2020 with a presence in Culver City, California, lists three project brackets on its site: $10,000–50,000, $50,000–100,000 and $100,000–250,000+, with matching timelines of 4–10 weeks, 3–6 months and 4–12 months, on a stack of WebGL, Three.js, React Three Fiber and WebAR/WebXR. That does not tell you what your project costs, but it tells you which conversation you are entering and lets you check the eventual quote against a public reference. The homepage also lists a closet configurator in the $10,000–50,000 bracket delivered in four months and a jewelry configurator in the same bracket, which is the category-and-bracket pairing you want every vendor to give you for your own product type. The limitation cuts the other way: a studio founded in 2020 has a shorter track record than Threekit (2005), VividWorks (2006) or Program-Ace (1992), so weight criterion 1 accordingly and ask for the studio's 300+ delivered projects to be narrowed to the configurators among them. Salsita names three complexity tiers without prices; many platforms publish nothing and route you to a sales call. None of these is wrong, but the vendor that publishes brackets has given you a negotiating anchor for free.

Criterion 5: team shape. Configurators have a long tail. The person who wrote the option logic in month two is the person you want fixing it in month nine when a new fabric collection breaks the pricing rules. Ask who will be on the project by name and role, whether the 3D artist and the engineer sit in the same team, and what happens if someone leaves. Larger firms absorb turnover; Program-Ace, in Nicosia, states more than 150 in-house staff. Smaller studios cannot, and should instead show you the documentation that makes handover possible.

Criterion 6: ownership. Read the contract for three things: who owns the optimized 3D models, who owns the source code or configuration rules, and what happens to both if you leave. With a SaaS platform you generally rent the viewer and rules engine and should expect to; the question is whether your models and rules can be exported in a usable form. With a custom build, insist on the repository and the asset library being yours at final payment. Roomle, the Austrian furniture platform, sells an open API and SDK, which is one way a platform reduces lock-in; a custom studio reduces it by handing over the code.

Criterion 7: commercial model fit. The choice between a platform license and a custom build is mostly a function of catalog size, change rate and how unusual your UX needs are. A furniture brand with 2,000 SKUs that change every season is a platform customer: VividWorks in Oulu, 3D Cloud in St. Petersburg, Florida, Cylindo and iONE360 all built their businesses around that shape. A manufacturer with one flagship line, a bespoke UX and an unusual integration is a custom-build customer. Model the three-year cost for both, using the 15–25% of build cost per year that agencies budget for maintenance (GoodFirms app cost survey, 2026) against the platform subscription.

Criterion 8: support terms. Weighted lowest because it is the easiest to fix by contract, but not zero, because browsers change. WebGL and WebXR behavior shifts with iOS and Chrome releases, and a configurator nobody is paid to test after launch will eventually break on the device your largest customer uses. Ask for response times, a rate card for new SKUs or materials, and an explicit statement on browser-update coverage.

A checklist of questions for the first call, and what to send before it

The first call is thirty to forty-five minutes. Vendors will spend the first ten on their deck if you let them; do not. Send the material below in advance and use the call to listen to how they reason about your product. The questions are grouped by the criterion they test; start with the group that carries the most weight for your project.

Send before the call:

• One representative product with its option list: every choice a customer can make, with dependencies and exclusions written out, even roughly.
• A sample of your source assets: one CAD file in its native format, or the product photography you have if CAD does not exist.
• The system that currently holds the price list, and the system that will receive the order.
• The devices and browsers your customers actually use, from your analytics, with the share of mobile traffic.
• Your catalog's change rate: how many new SKUs, materials or finishes arrive per quarter.
• The launch constraint, if there is one (trade show, season, ERP go-live).
• Your budget bracket. Withholding it does not get you a better price; it gets you a proposal scoped for the wrong bracket.

Questions on track record (criterion 1):

1. Send us three live configurators you shipped. Which one is closest to our product in rule complexity, and why?
2. For that project, how were the configuration rules expressed: in code, in a spreadsheet the client edits, in a platform's rules engine? Who maintains them now?
3. What went wrong on that project, and what did you change afterward?
4. Which projects did you turn down in the last year, and why?

A good answer to the third question is specific and slightly uncomfortable: a texture budget that was blown, a rules engine that had to be rewritten when the client's ERP changed. A vendor with no answer either has not shipped enough or is not going to be candid later either.

Questions on the asset pipeline (criterion 2):

1. Here is our CAD file. Walk us through what happens to it, step by step, before a customer sees it.
2. What is your target file size per product for mobile, and how did the last project land against it?
3. Which compression do you use for geometry and textures, and when do you not use it?
4. When we add a new product next quarter, who does the conversion, how long does it take, and what does it cost?
5. If we have no CAD, how do you build models, and at what per-model cost?

The fourth question is the one most buyers skip. The first ten products are inside the project budget; the next hundred are not, and the per-product conversion cost determines whether the configurator grows with the catalog or freezes at launch. Outsourced product modeling starts from about $42 per model for simple products (Modelry, CGTrader, pricing page, 2026), a floor to compare against a vendor's per-SKU figure.

Questions on integration (criterion 3):

1. In your last project, which system held the price, and how did the configurator read it?
2. Have you integrated with our specific ERP, CPQ or commerce platform before? Show us the connector or the API documentation.
3. What does the configurator output at the end of a session: a SKU string, a BOM, a quote PDF, a cart line item?
4. How do you handle a rule that only the factory knows, such as a minimum order quantity by material?
5. Who owns the integration layer if we switch ERP in three years?

Questions on commercials, ownership and support (criteria 4–8):

1. How is the quote structured: assets, logic, UX, integration, QA and support as separate lines, or one number?
2. Who is on the team by name, and are they the same people through launch?
3. Who owns the optimized models, the source code and the hosting account at final payment?
4. What does the support agreement cover, at what response time, and what is the rate for changes?
5. What happens when the next iOS release changes WebGL or WebXR behavior?
6. What is the three-year cost of this configurator, including your estimate of maintenance?

On the last question, a vendor who offers a number is more useful than one who refuses. A configurator built for $60,000 with maintenance at 20% costs $36,000 over the following three years; a platform at an illustrative $1,500 a month costs $54,000 over the same period with no separate maintenance line. Neither number fits every case, but a vendor who will not walk through the arithmetic is asking you to trust rather than check.

Seven red flags, each with what it leads to

Red flags are worth more than green ones because they are cheaper to observe. Each item below is visible in the first two conversations, followed by what it reliably produces six months later.

1. The demo is a video, a screenshot deck or a desktop-only link. What it leads to: you discover the performance problem at user acceptance testing, on the CEO's phone, and the fix is not a tweak but a rebuild of the assets, because heavy geometry cannot be optimized after the fact without redoing the pipeline. The commercial consequence is measured in the same currency as any slow page: the BBC lost an additional 10% of users for every extra second of load time, and Vodafone's 31% LCP improvement produced 8% more sales in an A/B test (web.dev, Google case studies, 2021). Insist on a link and open it on cellular.

2. A fixed quote arrives before they have seen your product data. What it leads to: the price is anchored to an assumption about your product, and the assumption is usually simpler than the truth. Once the real option list arrives, the fixed price becomes a floor and every dependency becomes a change order. The tell is a quote with a single number and no line for asset preparation. Ask for the quote to be re-issued after they have opened your CAD file and read your option list.

3. "We'll model everything from scratch" when CAD exists, or "we'll just export your CAD" when it does. Both are the same flag from opposite directions: no pipeline. Modeling from scratch when you already own engineering geometry doubles the asset cost and introduces dimensional errors the factory catches late. Exporting CAD without reduction produces the 20 MB-plus files that model-viewer maintainers call problematic. What it leads to: a budget overrun on assets, or a configurator that only works on the vendor's demo laptop. The right answer names the reduction step, the target size and who does it for product number eleven.

4. The price list lives inside the configurator. What it leads to: two price tables, one in your ERP or CPQ and one in the configurator, drifting apart from the first price change. Every promotion, currency update or material surcharge becomes a ticket to the vendor. Sales reps stop trusting the configurator's numbers and quote by hand, at which point the tool has become a cost with no output. The gap between website and seller is already the most common inconsistency B2B buyers report, at 69% (Gartner, 2024 survey). Ask where the price comes from, and if the answer is "we'll load your spreadsheet," ask what happens when the spreadsheet changes.

5. No engineer on the sales call. What it leads to: a scope written by someone who will not build it, handed to a team that inherits its assumptions. The symptom in month three is "that was not in the scope" about something you said clearly on the first call. It also signals bench-style staffing, where your team rotates with the vendor's utilization. Ask for the technical lead to join the second call at the latest, and put the pipeline questions to them directly.

6. ROI is promised with numbers from vendor pages. Configurator vendors publish aggressive results: Epicor's CPQ page claims that visual configuration boosts conversion rates by up to 40% and lists customer outcomes such as +108% annual sales (Epicor, product page, 2026, vendor data with no methodology); Threekit's homepage cites a 290% revenue increase within one month for one customer and 4× faster quoting for another (Threekit, 2026, vendor data, single-customer anecdotes). None of these are lies, but none are forecasts for you. What it leads to: a business case built on someone else's best month, a disappointed sponsor, and a vendor relationship that sours when the number does not arrive. Better-documented cases are still narrow: an analyst study of one manufacturer, CMTP, on Epicor CPQ recorded a 53% ROI, a 2.4-year payback and a 66% increase in quoting efficiency (Nucleus Research, July 2024)—one company, in a study commissioned around a vendor. Set your own baseline instead: current quote turnaround, sample-shipment cost, product-page conversion. A vendor who helps you measure those before launch is worth more than one who promises a percentage.

7. Ownership is vague, or a "custom" quote hides a platform license. Platforms are not the flag. Kickflip, a no-code configurator for Shopify, WooCommerce and Wix stores, and Zakeke, the Milan SaaS platform, are open about being subscriptions, and for the right catalog that is the correct answer. The flag is a proposal presented as a custom build that turns out to run on a third-party platform with its own license, where your models sit in the vendor's account and no export path is priced. What it leads to: a migration in year three that costs as much as the original build, because assets, rules and integration all have to be recreated. Ask for the IP and termination clauses in writing before kickoff.

How to read a portfolio and reviews without being steered by them

Portfolios and review sites are designed by the people they describe. That does not make them useless; it means you read them for what they cannot easily fake. Three habits cover most of it: test the work yourself, translate counts into relevance, and read results claims for their sample size.

Test every portfolio link on a phone over cellular, and time it. Note the time to first interaction, whether the model appears progressively or after a spinner, whether the phone warms up within two minutes, and whether rotating the model stutters. Then try the option logic: pick the most expensive combination and the cheapest, and see whether the price and the image both update. Do this for three projects per vendor. The exercise takes half an hour and tells you more than any deck. A portfolio of desktop-only experiences means the vendor has been selling to brand teams, not to customers on phones.

Look at product-type similarity, not at the logo. A cloud-streamed automotive configurator of the kind ZeroLight builds for OEMs on Unreal and Unity, rendered in a data center and streamed as video, tells you nothing about how a vendor would handle a kitchen cabinet line with 400 finish combinations delivered as glTF in a browser. Both are configurators; they share almost no engineering. Similarity that matters: rule complexity, the material realism required (a jewelry configurator lives or dies on metal and gemstone shading; a garden shed does not), integration depth, and device mix. Ask the vendor for the project that matches on those four axes, and if none does, ask what they would need to learn.

Translate counts into a shortlist. Vendors publish scale: Threekit reports more than 150 manufacturers live; Zakeke claims more than 25,000 brands; custom studios cite project totals in the hundreds. The number is a signal about survival, not about fit. Convert it with one request: of those, show us three that resemble ours. If a platform with 25,000 customers cannot produce three in your category, the platform is optimized for a different category. If a studio with hundreds of projects can only show configurators built two stacks ago, the current team has not built one.

Read reviews for cost band, recency and the "improvement" line. On Clutch and similar directories, each review carries the project's cost band and the reviewer's role, and ends with what could be improved. Those three fields are worth more than the star rating. A 4.8 across 46 reviews, as Program-Ace states on its site, and a 5.0 across 25 tell you roughly the same thing: enough clients were willing to be interviewed. What you want is whether the projects near your cost band went well, whether the reviewers lived with the result (an ecommerce manager) or signed for it (a CEO on a two-week engagement), and whether the improvement line repeats. "Communication across time zones" three times in a row is information. Check the date on the most recent review, too; a profile whose last review is two years old describes a different team.

Read every results claim for its sample. The claims in this space cluster into three kinds. Vendor marketing aggregates ("up to 40% conversion", "+105% deal size") come without methodology and describe the vendor's best customers. Single-customer testimonials ("quote turnaround from about two days to five or ten minutes," as one CMTP manager says on Epicor's site) are real for that customer and say nothing about yours. Analyst case studies, such as the Nucleus Research assessment of the same manufacturer, come with a method and a payback period but still cover one company. None of these belong in your business case as a forecast; they belong as a range of what happened elsewhere, with the source and the caveat next to the number. A vendor who volunteers the caveats is one who will tell you when your own project is off track.

What a real case study contains. A useful case study has five things: the product category and rule complexity, the source assets they started from, the integration on the other end, one measured number with its baseline, and something that went wrong. Hapticmedia's write-ups of Baccarat lighting and Longchamp bag configurators and Zea's success stories for vehicle makers such as Motrec and Taiga Motors at least publish the category and the starting point. When a case study has a logo, a hero image and an adjective, it is an advertisement; ask for the version with numbers.

Which type of vendor fits which type of project

The market splits into roughly eight vendor types, and most bad engagements come from matching a project to the wrong type rather than to a bad vendor. The table pairs common project shapes with the vendor type that usually fits, examples of companies in that type, how they tend to charge, and the thing to watch. Examples illustrate the type, not an endorsement; pricing is not published unless stated.

Two patterns in the table deserve a note. First, the largest catalogs are usually a two-vendor problem: a pipeline vendor to get the CAD into web-ready form and a configurator vendor to put it in front of customers. Trying to make one vendor do both is where the "we'll model everything from scratch" flag tends to appear. Second, the custom-studio row is where budgets are most visible, because studios that publish brackets exist and the general software benchmarks apply. Most software firms on Clutch charge $25–49 per hour, with US firms most commonly at $50–99 (Clutch, 2026); a GoodFirms survey of more than 100 software companies found 56.3% charging $20–50 per hour and 10.9% charging $100–250 (GoodFirms, survey September–October 2025). No reliable published rate exists specifically for Three.js or WebGL specialists; in our experience they sit at the upper end of their regional band, which is an observation, not a statistic.

A typical custom engagement looks like this: a manufacturer of modular outdoor structures, with CAD for every panel and a dealer network that quotes by email, hires a studio to build a browser configurator that outputs a BOM and a dealer quote request. The studio reduces the CAD to web assets, writes the option logic against the manufacturer's rules, connects the output to the CRM and delivers in a 3–6 month window inside a published $50,000–100,000 bracket. Year two adds a new panel system at a per-SKU rate and a support retainer at roughly a fifth of the build cost. The same project on a visual CPQ platform would carry a license instead of a build fee and put the rules in the platform's engine: better if the sales team already lives in Salesforce, worse if the manufacturer wants a UX that looks like nothing else on the market.

Frequently asked questions

How much does a 3D configurator cost?

Between roughly $10,000 and $250,000 or more, depending on rule complexity, asset preparation and integration, and that range is the honest one. Studios that publish brackets put simple configurators in the $10,000–50,000 range and integrated, multi-product builds at $100,000–250,000+. The general software benchmarks agree: the typical Clutch project runs $10,000–49,000 and the average total engagement $132,480 (Clutch, 2026), while GoodFirms puts small-to-mid custom software at $30,000–100,000 and large projects at $100,000–200,000 (GoodFirms, survey of 100+ companies, 2025–2026). Platforms charge a subscription instead and rarely publish it. Budget separately for assets per SKU and for maintenance at 15–25% of build cost per year.

How long does it take to build one?

From about four weeks for a single-product configurator with existing clean assets to a year for a multi-line, ERP-integrated build. Published studio timelines run 4–10 weeks, 3–6 months and 4–12 months by bracket, and GoodFirms' app survey reports 3–6 months for basic and 6–9 months for mid-level applications (GoodFirms, survey of 267 companies, updated August 2026). Asset preparation is usually the critical path, not the code: if your CAD is heavy or your product photography is inconsistent, add weeks before the first working screen. Ask the vendor for a schedule with the asset milestone shown separately.

Should we choose a platform or a custom build?

Choose a platform when your catalog is large and changes often and your UX needs are ordinary; choose a custom build when the product line is focused, the UX is part of the brand, or the integration is unusual. Platforms such as VividWorks, Roomle, Threekit or Expivi have solved catalog onboarding and common integrations many times, which is what you pay the license for. A custom studio gives you the code and the assets and a UX nobody else has, at the price of a maintenance line you own. Model both over three years, including the per-SKU cost of growth, before deciding.

Can a configurator connect to our ERP or CPQ?

Yes, and it should, because the alternative is a second price list that drifts from the first. Visual CPQ platforms list connectors for Salesforce, SAP, Oracle and Microsoft Dynamics; custom studios build against your system's API. The practical questions are which system is the source of truth for price and availability, what the configurator outputs (SKU string, BOM, quote request, cart line), and who owns the integration layer if you change ERP. Get those three answers in writing before the visual work starts. Sales teams already spend 60% of their time on non-selling tasks (Gartner Sales Survey 2024, as reported by Salesforce), and a configurator that reps have to re-key into the CPQ adds to that rather than removing it.

Who owns the 3D models and the code after launch?

With a custom build, you should, at final payment: the optimized models, the source repository and the hosting account. With a platform, you typically own your source assets and rent the viewer and rules engine, so the question becomes whether your optimized models and configuration rules can be exported in a form another vendor could use. Ask the vendor to describe, in one paragraph, exactly what you receive on the day you stop paying. Vagueness in that paragraph is the most reliable predictor of a painful migration later.

What does maintenance cost after the configurator is live?

Plan on 15–25% of the build cost per year, which is the range agencies report for software maintenance generally (GoodFirms app cost survey, 2026), plus a per-SKU rate for new products, materials or finishes. The recurring work is real: browser and operating-system releases change WebGL and WebXR behavior, the catalog grows, prices move, and the analytics will show option combinations customers want that the rules do not allow. A vendor who quotes zero maintenance is quoting a demo. Ask for a support agreement with response times and a rate card as part of the original proposal, and compare its three-year total against a platform subscription for the same catalog.