Insights · E-commerce

D2C ecommerce: what owning the customer costs

The commission you escape is not a margin you keep. It is the price of demand you now have to create yourself. And while the industry argued about it, the marketplace quietly stopped charging a contract rate and started charging an auction rate.

D2C ecommerce is selling your own product to the end customer through a channel you control, rather than through a retailer or a marketplace that takes a cut. The case for it is always put as a margin case: stop handing an intermediary 12 to 15 percent of the price and keep it. That case is wrong on both sides of the equals sign, and the numbers that show why were published this quarter by the two companies on either side of the trade.

The commission is not the price

A commission is a disclosed, contracted percentage of a transaction. It is the easiest cost in the business to model, which is exactly why it dominates the discussion. Nobody builds a board paper around a cost they cannot name.

But a marketplace does not sell you a transaction. It sells you a transaction and a position in a queue, and it charges for them separately. The first is a contract. The second is an auction, cleared against every competitor who wants the same position, which means its price is set by the willingness to pay of the firm in your category with the loosest payback discipline. You cannot negotiate it, you cannot forecast it, and it has no ceiling other than the margin of the most desperate seller in the category.

So the real comparison is not commission against zero. It is commission plus auction against zero commission plus a different auction, run by Meta and Google instead, plus the fixed cost of the shop, the stack, the service desk and the people. Put that way, the question stops being which channel has better unit economics and becomes which channel you can survive the fixed costs of. That is a cash question, not a margin question.

You never escape the cost of demand. You only change who invoices you for it, and whether the invoice is a contract or an auction.

What the marketplace actually charges now

Allegro reported its second quarter of 2026 on 17 September. Group GMV was 19,572 million zloty, up 14.4 percent. The group take rate, the share of GMV the marketplace keeps as revenue, was 12.30 percent, and it was down 0.56 percentage points year on year.

Read in isolation that looks like good news for sellers, and it is how a take rate line usually gets read. Now read the next line in the same deck. Advertising revenue in Poland grew 31.1 percent year on year and now equals 2.31 percent of GMV. Against GMV growth of 14.4 percent, advertising is growing at better than twice the rate of the thing it is charged on, which means the advertising cost per zloty of GMV rose about 14.6 percent in a year. Advertising is a component of that falling take rate, roughly a sixth of it, and it is the only component going up.

That is the whole mechanism in two lines from one slide deck. The disclosed, contracted, negotiable part of the cost of selling on a marketplace is going down. The undisclosed, auctioned, unnegotiable part is going up, faster than the market it sits on. A seller comparing this year's take rate to last year's take rate will conclude the channel got cheaper. A seller comparing this year's invoice to last year's invoice will conclude otherwise, and the seller is right.

One more number from that deck, because it decides the strategic question rather than the arithmetic one. Allegro had 20.9 million active buyers, up 1.6 percent. GMV grew 14.4 percent on a buyer base that grew 1.6 percent. In a country of under 38 million people, the marketplace is not a channel that reaches part of the market. It is close to being the market, and it has stopped adding people to it. Anything your own shop acquires, it re-acquires from there.

It stopped buying attention and started selling it

Amazon filed its second quarter on 30 July. The disaggregated revenue table is the most honest description of what a marketplace is that either side of this argument has produced.

Online stores, which is Amazon selling its own inventory, took 70,432 million dollars, up 14.6 percent. Third-party seller services, which is the commission and fulfilment business, took 46,780 million, up 15.9 percent. Advertising services took 19,809 million, up 26.2 percent. And Amazon's own sales and marketing expense, the money it spends to bring shoppers in, was 11,698 million against 11,416 million a year earlier. Up 2.5 percent.

Set those two lines against each other. In the second quarter of 2025, Amazon earned 1.37 dollars of advertising revenue for every dollar it spent on its own sales and marketing. A year later it earned 1.69. Worldwide paid units grew 17 percent and advertising revenue grew 26.2 percent, so advertising revenue per unit sold rose about 7.9 percent. Advertising is now 42.3 percent of the size of the entire third-party seller services business, up from 38.9 percent.

This is not a retailer with an advertising sideline. It is an attention business with a logistics attachment, and the attention it sells is attention to products that brands have already agreed to put on its shelves. The brand funds the demand generation, the marketplace prices it, and the marketplace's own cost of bringing the shopper in has effectively stopped growing because it no longer needs to.

The symmetry is the point. A brand that leaves for its own channel is not leaving an auction. It is swapping into Meta's, and that one took 60,801 million dollars in the same quarter against 47,516 million a year earlier, up 28.0 percent. Two auctions, both growing faster than the commerce underneath them. The choice is which auction you want to be a price taker in.

Disruption from below already finished

Bower and Christensen described the pattern in 1995: the entrant arrives worse on the dimension incumbents measure, good enough on a dimension they ignore, and takes the low end that nobody defends because nobody wants it. Then it moves up.

Amazon's own filing puts the end of that story at 61 percent. That is the third-party seller share of worldwide paid units in the second quarter of 2026, down a point from 62 percent. Unbranded, unknown, often unserious sellers move the majority of the units. The disruption is complete and it has been for years, and the brands that let it happen let it happen for the textbook reason: the low end looked unprofitable, so defending it looked irrational.

Here is the part that is usually missed, and it is the part that matters now. The disruption did not end with the marketplace taking brand volume. It ended with the marketplace owning the moment the category gets decided, and then selling that moment back to the brand by auction. When the shopper's first query happens inside the marketplace, the category entry point has moved inside a system the brand does not control and cannot see into. The 61 percent is where the first disruption finished. The 26.2 percent advertising growth is where the second one is running, and this time the brand is not losing share of shelf. It is renting its own shelf at a price its competitors set.

Which makes the fashionable conclusion, that D2C was a mistake and brands should retreat to where the volume is, exactly half right and dangerous in the half it gets wrong. Going back is correct on volume and wrong on power. You do not fix a pricing-power problem by moving further into the system that holds the pricing power.

Fifteen years of D2C, measured

The honest way to price the other side of the trade is to find somebody who has run it at scale for long enough that the fixed costs have had time to settle, and read their accounts. Warby Parker filed its 2025 results in 2026, fifteen years after launching as a pure direct-to-consumer brand that was going to make the intermediary unnecessary.

Net revenue was 871.9 million dollars, up 13.0 percent. Gross profit was 470.6 million, a gross margin of 54.0 percent. Selling, general and administrative expenses were 475.9 million, which is 54.6 percent of revenue, and which exceeds the entire gross profit by 5.3 million. Net income was 1.6 million dollars. Average revenue per customer was 324 dollars, up 5.7 percent, on an active customer count up 7.0 percent. And the company ended the year with 323 physical stores, 47 of them opened that year.

Take the store count seriously, because it is the finding. The brand whose premise was that you do not need shops answered the economics of direct demand generation by building 323 shops. Not because it changed its mind about retail, but because a shop is a cheaper and more durable way to make a stranger aware of you than renting that awareness by auction, month after month, forever. D2C did not fail. It converged on retail, which is what the cost of demand does to anybody who tries to carry it alone.

And look at where the growth came from: customers up 7.0 percent, revenue per customer up 5.7 percent. Fifteen years of owning the relationship, and the business still grows mainly by adding people rather than by compounding the ones it has. That is not a Warby Parker failing, it is the ordinary behaviour of brands, and we have set out the evidence for it in penetration, not loyalty. It matters here because the entire financial case for D2C rests on the opposite assumption. You pay once to acquire and then harvest repeat purchases that cost nothing. If the repeat does not compound, the acquisition cost never amortises, and the fixed cost is permanent.

The threshold is 22.5 percent

So cost it. Not the margin, the threshold, because the threshold is the only number that survives disagreement about your particular basket and your particular funnel.

Start with the marketplace side, where the number is published: 12.30 percent of GMV. On a 200 zloty order that is 24.60 zloty, and assuming 40 percent contribution margin before channel cost, it leaves 55.40 of the 80 zloty of contribution. The own-channel side keeps all 80, and then has to pay for the demand out of it. So 24.60 per order is the entire budget D2C has to create the demand the marketplace created for nothing.

Most a D2C order can spend on demand	Per customer
1 lifetime order	24.60
2 lifetime orders	49.20
3 lifetime orders	73.80
5 lifetime orders	123.00
Share of own overhead at which the saving disappears	22.5%

Basket and contribution margin are illustrative; the 12.30 percent and the 54.6 percent are reported figures. Currency is irrelevant, every line is a ratio. Contribution rather than revenue, for the reason we set out in the discount nobody costed.

Now the last line, which is where the argument actually ends. You do not need to apportion Warby Parker's SG&A precisely, and you could not. You need the ratio between the commission you avoided and the overhead you took on: 12.30 divided by 54.58 is 22.5 percent. If more than 22.5 percent of your own overhead exists because nobody else is generating your demand, the commission was the cheaper option, and you have paid a premium for the privilege of owning the customer.

Run it on your own numbers and the answer will usually be uncomfortable, because the overhead of a direct channel is not mostly technology. It is the people and the media that exist to make strangers aware of you, and in a business that was previously sold through retail, that function did not exist at all. It was free, bundled into the retailer's margin, and invisible until the moment you stopped paying for it.

What you buy back, and what each channel is for

If the margin case for D2C does not survive the threshold, something else has to justify it, and there is something. It is just not margin and it is not data.

On a marketplace you do not set the price you are compared on. The page sets it. Your product appears beside a reference price, a unit price, a cheaper variant and a private label, in a layout built to make substitution effortless, and the comparison is the product the marketplace is really selling. Every pricing lever that works through context rather than through the number on the label gets disabled: pack architecture, good-better-best laddering, bundle framing, the entire toolkit we use in Pricemore. On your own channel those levers work, because you own the shelf they sit on.

That is a real asset and it is worth paying for. The mistake is paying for it while telling the board you are doing it for margin, because the margin promise is testable and it fails, and when it fails the pricing-power argument dies with it. Say what you are buying. You are buying the right to set the context of your own price, and the price of that right is the 22.5 percent threshold above.

The first-party data argument deserves the same treatment. The data is real, and it is worth exactly what it changes. If it changes a decision, cost the decision. If it feeds a dashboard nobody has ever acted on, you are paying a channel premium for a reporting feature, and the measurement problem underneath it is the one we set out in incrementality testing: the fact that a number arrives in your warehouse rather than a vendor's does not make it causal.

Which leaves the allocation, and almost nobody should pick one channel. The useful question is what each channel is for, and the allocation falls out of the asymmetry in the numbers above.

The marketplace is where the category is decided, so it gets the job of being present at the decision: availability, the pack the comparison is won with, and the minimum advertising required to hold position rather than to chase incremental volume. That last clause does most of the work. On-marketplace advertising bought to grow looks efficient and usually is not, for the reasons we set out in retail media, where eight times reported commonly lands between one and two times incremental. Bought to defend position, it is a shelf-space cost and should be approved like one, against a position, not against a return.

The own channel is where you do the things the marketplace page will not let you do: launch at a price the comparison has not yet anchored, test pack architecture before it reaches a public price history, carry the range that is too slow for the marketplace's economics, and sell subscriptions, which are the only form of repeat that actually amortises an acquisition cost. Judge it on those jobs. Judging it on total revenue share is how it gets killed at the first budget review, long before any of them have paid off.

And set one rule before either budget is approved: the same incremental standard applies to both. The reason the argument has swung twice in five years, from build D2C to abandon it, is that each side was measured by the method that flattered it. The marketplace reported attributed returns from its own ad server. The direct channel reported revenue growth bought with venture money. Neither number was a measurement.

How we count it

This sits across two of our lenses. Cashstream, our portfolio profit and loss lens, prices the channel decision in cash rather than in margin: payback on the marginal customer, contribution rather than revenue, and the threshold ratio above, which converts an argument about strategy into one published percentage against one line of your own overhead. Shopperology asks the prior question, which is where the decision happens. If the shopper's first query is inside the marketplace, your own shop is not competing with the marketplace for the decision. It is competing for people who have already made it somewhere else, and you are paying an auction to re-acquire them.

Five things to do, in order. Work out your real cost of selling on each marketplace by adding the commission, the fulfilment fee and the advertising invoice, then divide by GMV on that platform, and compare that number to the take rate the platform reports, because the gap is the auction. Pull your own overhead and mark the share of it that exists to generate demand, honestly, including the agency and the team. Run the threshold: platform cost divided by total overhead, and see which side of it you are on. Then define what each channel is for in one sentence each, and if the own channel's sentence is own the customer relationship, you have not defined anything, because that is a cost, not a job. Finally, hold one of them out for six weeks and measure the other, because until somebody does that, the entire allocation rests on two numbers produced by the people selling you the media.