Why Configurable Products Are Hard to Sell Online
In 2000, Sheena Iyengar and Mark Lepper published one of the most quoted experiments in consumer psychology. A tasting table in a California grocery store offered either 6 jams or 24. The larger table drew more visitors. The smaller one sold roughly ten times more jam. The finding, that abundant choice attracts attention and then suppresses the decision, has shaped twenty years of thinking about how commerce should present options.
It is a genuine result, and it is also routinely misapplied. The jam study describes a buyer who can evaluate every option on the table and struggles only with the number of them. Curation solves that problem: fewer jams, cleaner shelves, a confident default. An entire school of interface design descends from this insight.
But there is a class of products where the buyer's difficulty is different in kind, and no amount of curation reaches it. A power transformer, a matched audio system, a made-to-measure jacket. Here the buyer cannot pick the right product from any shelf, however elegantly arranged, because the right product does not exist until it has been derived from their situation. The load profile, the listening room, the shoulders. These purchases are not chosen. They are specified. And almost everything about how commerce works online assumes choosing.
The specification problem
A configurable product is defined across several variables that interact:
A power transformer's rating, voltage class, vector group, impedance, and loss capitalisation are not independent selections; a decision on one constrains the others, and some combinations are simply invalid.
An audio system's amplifier and speakers are not two purchases but one relationship, mediated by sensitivity, impedance, and the room they will share.
A garment's ease, cloth, and construction answer to a body and an occasion together.
The buyer, meanwhile, does not arrive speaking in parameters. They say roughly 160 MVA, and the honest answer depends on the impedance the network needs for parallel operation, which they have not mentioned. They say warm sound, which is a real preference wearing an unmeasurable word. They say like my old blazer but roomier, which contains a complete specification that no form field can receive. The buyer's language is a starting point for a derivation, and somebody, or something, has to conduct it.
For a century this was a person's job. The sales engineer who asks about the site before quoting. The dealer who asks about the room before recommending. The tailor who asks what the buyer means by roomier. What these people have in common is that they hold the product's rules in one hand and the buyer's vernacular in the other, and translate between them. The specification problem, stated plainly, is that this translation is the sale, and the instruments of online commerce cannot perform it.
Why the standard instruments fail
Commerce online has three instruments for complexity, and they fail against specifications.
The catalogue with filters assumes the buyer can express their need as attribute values. For products with thousands of valid configurations this collapses immediately: either the catalogue enumerates variants until no interface quite manages them without scroll after scroll, or it hides the variety behind filters the buyer does not know how to set. A filter panel asking for vector group is a question aimed at the wrong person.
The visual configurator, the industry's proudest answer, works beautifully where it works: colours, finishes, engravings, anything the buyer can see their way through. It fails structurally where the constraint is invisible. No configurator asks about the harmonic profile of a variable-speed-drive load, because the buyer cannot answer with a click, and the configurator cannot explain why the question matters. Configurators let buyers assemble what they can see. Specification is mostly about what they cannot.
The chat widget, in its common form, answers questions and pastes links. It has no access to the product's validity rules, no authority to hold a requirement across turns, and no output beyond a transcript. When the conversation ends, nothing exists that a quotation process or a cutting table can act on. It is company at the moment of confusion, which is not nothing, but it is not selling.
What all three share is an assumption inherited from the jam table: that the buyer's job is to pick, and the interface's job is to make picking pleasant. For specified products, the buyer's job is to be understood, and none of these instruments can conduct an understanding.
What the gap costs
The cost of leaving specification to email and patience is documented differently in each sector, and it is large in all of them.
In engineered equipment, the enquiry waits. A serious question sits in a shared inbox while a sales engineer finishes the previous serious question, and the buyer, being serious, has sent the same enquiry to three competitors. The specification then degrades in transit: told to one person, retold to another, retyped into a quotation, with each retelling admitting an error that surfaces at order review, at inspection, or on site. A mis-specified power transformer does not come back as a return. It comes back as a failed factory acceptance test, a delayed substation energisation, and a utility customer who remembers.
In apparel, the cost arrives as returns. NRF's 2025 figures put the overall US online return rate at 19.3%, with clothing running far above the average, and McKinsey's analysis attributes around 70% of fashion returns to fit and sizing. Behind those numbers is a simple mechanism: the buyer was asked to guess, in a sizing language that means something different in every house, about the one attribute that cannot be photographed. Made-to-measure exists precisely to remove the guess, and then, sold through a size-chart interface, reintroduces it.
In considered categories like engineered audio, the cost is quieter and shows up as wasted expertise. The buyer researches for weeks across forums and other people's rooms, arrives at a dealer with a shortlist assembled without anyone asking about their amplifier or their four-by-five-metre room, and the demonstration hours, the scarcest resource in the channel, are spent correcting the shortlist rather than closing it.
Three sectors, one structure. The purchase stalls, degrades, or bounces in the space between the buyer's first expression of interest and the first accurate, validated answer. We have written before about the purchase conversation as a sales function rather than a support cost; specification is where that argument becomes unavoidable, because for these products the conversation does not assist the sale. It is the sale.
What the conversation has to do
If the translation is the sale, then the answer is to engineer the translation: a conversation that can be held the moment interest appears, at any hour, in the buyer's own language, and that ends in the structured thing the seller's process needs next. This is what we built Cornea to conduct, and the requirements are worth stating in general terms, because they define the category whether or not the instrument is ours.
It must treat the buyer's words as starting points. Roughly 160 MVA is a hypothesis to verify against the load, not a number to transcribe. Warm sound is a preference to translate into amplification and pairing. The conversation has to know when an answer changes the engineering, and say so.
It must hold the product's rules inside the conversation. Validity is checked as the specification forms, so an impossible configuration dies as a sentence rather than surviving to a quotation. Commercial ceilings are enforced the same way, invisibly, so the proposal that emerges is one the seller can actually honour.
It must produce structure. The test of the conversation is what exists when it ends: a quotation-ready specification sheet, a measurement set a tailor can cut from, a matched-system proposal with the constraints recorded. A transcript is a by-product. The structured output is the point.
It must hand over warm. Considered purchases still close on human judgement, and the system's job is to make sure that judgement is spent well: the sales engineer opens a verified brief instead of a cold email trail, the dealer opens a thread that already knows the room, the fitter is flagged in when a fitting is genuinely worth the buyer's time.
And it must know its boundaries. It speaks from the knowledge the brand configured and does not improvise past it. When it does not know, it says so and escalates with the context attached. The rules for that handover belong to the seller, and the buyer always knows when a person has joined. Everything we argued in our guide to evaluating agentic AI about mandate and authority applies with particular force here, because this system speaks to your customers in your name.
In engineered equipment (read: transformers)
The intake becomes one structured conversation, held when the enquiry arrives rather than when the inbox is reached. The buyer describes the site in their own terms; the system asks for what is missing, verifies what is stated, and explains why the parallel-operation impedance matters before sizing around it. What reaches the sales engineer is a quotation-ready specification with the full thread attached, in the format the quotation process expects. The enquiry-to-quotation cycle moves from days toward hours, which matters because the first accurate answer usually shapes the shortlist. And the errors that used to surface at inspection now end inside the conversation, where a correction costs a sentence.
In high-end audio
The conversation starts from the room, the music, and the equipment the buyer already owns, because in a system product those are the specification. Synergy knowledge, amplification matching, sensitivity, source pairing, operates as configured expertise with hard claim boundaries: the system recommends within what the brand's own knowledge supports and declines to speculate past it. The shortlist that reaches the dealer has been specified before it is demonstrated, so the demonstration hours convert instead of triaging. The final judgement stays where it belongs, with the dealer and the buyer's own ears; the system's contribution is to make sure that moment is spent on the right equipment.
In bespoke and made-to-measure apparel
Fit is settled from a garment that already fits. The buyer describes it or photographs it, says what worked and what did not, and the conversation turns that into a measurement set the tailor can act on, flagging honestly where a physical fitting is worth the trip. The leap of faith that drives fit returns becomes a conversation about something the buyer already trusts. And the second order, where made-to-measure economics actually live, becomes one message, because the profile holds the pattern.
The record the conversation leaves
There is a second return on this architecture, and it compounds. Every specification conversation, converted or not, leaves structured data about what the market asked for: which loads, which rooms, which cloths, which budgets, which objections. A sales ledger records what sold. It has never recorded what was wanted, in the buyer's own words, validated against the seller's own rules. Over time that record becomes demand intelligence of a kind the sector has not had, and it belongs to the seller, in their environment, consistent with the principle that runs through everything we build: the system does the work, and the organisation keeps the record of how.
Where this does not fit
A brand selling a standard catalogue of low-consideration goods does not have a specification problem; it has a merchandising problem, and lighter instruments serve it well. The full specification engine earns its keep where products are specified, matched, or made to measure, where an error is expensive, and where expert time is the constraint. For everything short of that, the conversational layer alone, reading the live catalogue, answering accurately, escalating warm, is usually the right size, and we say so on a walkthrough rather than sell the heavier build to an operation that does not need it.
In closing
The jam study taught commerce to fear abundance, and the lesson was learned so well that an entire industry now curates, filters, and simplifies by reflex. For configured products the reflex misses, because the difficulty was never the number of options. It was that the right option has to be derived, from constraints the buyer holds and rules the seller holds, and the derivation is a conversation that the instruments of online commerce were never built to conduct.
The test for any operator reading this is short. Look at where your serious enquiries wait, where your specifications degrade between the person who heard them and the process that acts on them, and where your best people's hours go to questions a structured conversation could have settled. That is the gap. It has been paid for in quotation delays, returns, and demonstration hours for as long as these products have been sold at distance. It no longer has to be.
—
Mitochondria is an agentic AI product company based in Amsterdam and Pune. ISO 27001:2022 certified. Classified within the limited and minimal risk tiers of the EU AI Act, with controls aligned to the GDPR, UK GDPR and India's DPDP Act.