Blog » Industry

Marketplace KPI Tree: Two Sides of the Same GMV

Name: kpitree.io
Availability: InStock
Author: kpitree.io

August 27, 2026 · 12 min read

Marketplace KPI tree: buyer-side and supply-side GMV decompositions both reconcile and disagree. Worked arithmetic on a 600,000 dollar decline.

Two Correct Answers to One GMV Decline

A marketplace KPI tree can be built from either side of the market, and the side you choose changes which driver takes the blame.

Gross merchandise value is the total sales dollar volume transacting through a marketplace in a period, and platform revenue is the portion of that volume taken as fees.⁵

The demand-side identity is active buyers times orders per buyer times average order value. The supply-side identity is active sellers times orders per seller times the same average order value. Both are complete, and both reconcile to the dollar.

The example below runs one services marketplace across two quarters. GMV falls from 18.0 million dollars to 17.4 million, a decline of 600,000 dollars, while active buyers grow by 5,000 and active sellers fall by 2,500.

Every figure in this article reconciles exactly. None of it is a residual.

Why Do Buyer-Side and Supply-Side GMV Decompositions Disagree?

They disagree because each identity holds a different variable constant. The buyer-side chain prices a change in orders through orders per buyer. The supply-side chain prices the same change through orders per seller. Both reconcile to the same total, but they distribute that total across different drivers, so the largest term differs by side.

Orders is the shared node. Both sides divide the same 290,000 orders and both multiply by the same 60.00 dollar average order value.

What differs is the denominator above orders. Divide orders by buyers and you get 2.32 orders per buyer. Divide the same orders by sellers and you get 23.20 orders per seller.

Sellers fell 16.7 percent while orders fell 3.3 percent, so orders per seller had to rise. That rise is arithmetic, not performance. It is what a smaller denominator does.

The same mechanism runs in reverse on the buyer side. Buyers grew 4.2 percent while orders fell, so orders per buyer had to fall.

Neither ratio is a behavior. Each is a division result that absorbs whatever the count did not explain, which is why the two attributions rank drivers differently while both stay correct.

The Buyer-Side Attribution, Term by Term

The buyer-side chain is GMV equals active buyers times orders per buyer times average order value. Attribution runs sequentially in that order, holding earlier terms at current values and later terms at prior values.

Buyer count. Five thousand additional buyers at the prior 2.50 orders and prior 60.00 dollars contribute 5,000 times 150.00, or 750,000 dollars.

Orders per buyer. The rate falls from 2.50 to 2.32 across 125,000 buyers at 60.00 dollars, contributing 125,000 times negative 0.18 times 60.00, or negative 1,350,000 dollars.

Average order value. Flat at 60.00 dollars, contributing zero.

The three terms sum to negative 600,000 dollars with no residual. Sequential attribution removes the unallocated gap that single-factor effects leave in any multiplicative chain, worked through in the DuPont ROE decomposition.

Read alone, this tree says the marketplace acquired buyers and failed to keep them transacting.

The Supply-Side Attribution, Term by Term

The supply-side chain is GMV equals active sellers times orders per seller times average order value. Same method, same order, same period.

Seller count. Sellers fall from 15,000 to 12,500. Negative 2,500 sellers at the prior 20.00 orders and 60.00 dollars contributes negative 2,500 times 1,200.00, or negative 3,000,000 dollars.

Orders per seller. The rate rises from 20.00 to 23.20 across 12,500 sellers at 60.00 dollars, contributing 12,500 times 3.20 times 60.00, or positive 2,400,000 dollars.

Average order value. Flat, contributing zero.

The three terms sum to negative 600,000 dollars. Identical total, opposite story: the marketplace lost a sixth of its sellers and the survivors absorbed most of the volume.

Gross movement differs more than the net. The buyer side allocates 2.1 million dollars of absolute contribution. The supply side allocates 5.4 million. The largest single term is 1.65 million dollars larger on the supply side.

Ordering Inside a Side Moves Less Than the Choice of Side

Sequential attribution depends on the order terms are introduced, a known weakness of the method. Averaging each driver's marginal contribution over every possible ordering removes that dependence, the approach formalized as Shapley value decomposition in strategy research.⁶

Reverse the buyer-side chain and take orders per buyer first. The rate term becomes 120,000 times negative 0.18 times 60.00, or negative 1,296,000 dollars, and the buyer count term becomes 5,000 times 2.32 times 60.00, or positive 696,000 dollars. Both still sum to negative 600,000 dollars.

The rate term moved by 54,000 dollars. Set that against the 1.65 million dollar gap between sides.

Ordering inside one identity moves the answer by 54,000 dollars. Choosing the identity moves it by roughly thirty times that. Teams argue about ordering conventions and rarely record which side their tree is rooted on.

Can One KPI Tree Hold Both Marketplace Sides?

Not as one identity. Buyers and sellers are factors in two separate complete expansions of the same root, not co-factors of one. Hanging both under GMV implies GMV equals buyers times orders per buyer times average order value times sellers, which is false by a factor of the seller count. One side is the tree. The other is a parallel view.

This is the most common structural error in marketplace trees, and it survives review because every node name looks reasonable.

The tell is a root that stops reconciling. Multiply 125,000 buyers by 2.32 orders by 60.00 dollars by 12,500 sellers and the result is 17.4 million dollars times 12,500.

The correct arrangement puts orders at the junction. GMV equals orders times average order value. Below orders, exactly one expansion is wired as parent and child. The other sits beside it as a second view reading the same orders column.

That second view is the check, not decoration. If both views do not return 290,000 orders, the two active counts come from different windows or different definitions of active, and neither attribution can be trusted.

The general parent-child contract this respects is set out in the KPI tree template.

Three Ways to Root a Marketplace Tree

Each arrangement answers one question well and one badly, and the choice should be made once and written down.

The failure is not picking the wrong root. It is picking a root by accident, usually because buyer-side data was the data that existed, then reading the result as if it were causal.

The table names what each arrangement attributes, what it hides, and the condition that makes it correct.

Tree root	Identity	Attributes well	What it hides	Choose when
Buyer-rooted	GMV = active buyers x orders per buyer x AOV	Demand acquisition and repeat rate	Supply availability. A seller exit appears only as lower orders per buyer	Supply is abundant and never the binding constraint
Supply-rooted	GMV = active sellers x orders per seller x AOV	Supply retention and per-seller throughput	Demand acquisition. A buyer surge appears only as higher orders per seller	Supply is scarce, contracted, or churning
Transaction-rooted	GMV = orders x AOV, with buyers and sellers as derived intensities	The size and value of the transaction base	Which side moved. It defers the question rather than answering it	The causal side is contested and you need a neutral baseline
Both sides in one tree	GMV = buyers x orders per buyer x AOV x sellers	Nothing. The identity is false	Every driver, because the root no longer reconciles	Never

Which Side Should a Marketplace KPI Tree Be Built From?

Root the tree on the side that is scarce. If sellers are contracted, capacity constrained, or churning, build supply-rooted, because the count you can act on sits at the top. If supply is abundant and demand is the constraint, build buyer-rooted. When neither is clearly scarce, root on orders and hold both sides as views.

Scarcity is the test because the top term in a sequential chain carries the largest contribution by construction. Putting the abundant side at the top inflates a driver nobody manages.

In the worked example sellers fell 16.7 percent while buyers grew, so supply is the scarce side and the supply-rooted tree is the correct one. Its answer is that seller loss cost 3.0 million dollars against 2.4 million recovered through higher throughput per remaining seller.

The buyer-rooted reading of the same quarter, that repeat rate fell, is arithmetically true and operationally misleading. Orders per buyer fell partly because there were fewer sellers to buy from.

One caution the arithmetic cannot resolve. Neither decomposition establishes causality. It ranks magnitude. The claim that seller churn caused the buyer slowdown needs evidence outside the tree, such as the order in which the two series turned.

What Public Marketplaces Disclose About Each Side

Reported metrics shape which tree gets built, and the supply side is instrumented less often.

Etsy reports both. For the second quarter of 2025 it disclosed 87.3 million active buyers on the Etsy marketplace, down 4.6 percent year over year, GMS per active buyer of 120 dollars on a trailing twelve month basis, down 2.9 percent, and 5.4 million active sellers, read on August 27, 2026.¹

Airbnb defines gross booking value, nights and experiences booked, and active listings, which gives a countable supply node, read on August 27, 2026.²

Uber's Form 10-K names monthly active platform consumers and trips as operating metrics and describes drivers and couriers without an equivalent named metric.³ DoorDash's Form 10-K describes Dashers in the same narrative way.⁴

A team that roots on buyers because sellers were never counted has not chosen an arrangement. It has inherited one from the reporting layer.

Where the Take Rate Layer Sits

Net revenue is not a factor of GMV. It is a sum.

Build it additively: seller fees plus buyer fees plus advertising plus other revenue. Take rate is derived by dividing that sum by GMV, never stored as a percentage on a row.

Keep it additive because the fee bases differ. Seller commissions are charged on GMV dollars. Buyer service fees are frequently charged per order. Advertising responds to neither.

In the worked quarter orders and GMV both fell 3.33 percent, because average order value was flat, so a blended take rate holds with no change to any fee schedule.

Let average order value move and the bases separate. Blended take rate then shifts while every published rate stays fixed. Treating that shift as a pricing event is the error the additive structure prevents.

How Do You Test That a Two-Sided Tree Is Correct?

Run four checks. Both identities must return the same GMV to the dollar. Both must divide the same order count. Exactly one side may appear as a multiplicative child of the root. And every ratio node must be derived from two summable columns rather than stored as a rate.

Test 1. Dual reconciliation. Compute GMV from the buyer chain and the seller chain independently. Both must return 17,400,000 dollars. A gap means the two active counts cover different date windows or different definitions of active.

Test 2. One order column. Both expansions must divide the same order count. A buyer-side tree counting orders against a supply-side tree counting shipments produces two numbers that were never comparable.

Test 3. One multiplicative side. Sellers and buyers may not both hang under GMV as factors. Multiply the leaves and compare the product to reported GMV.

Test 4. Derived ratios only. Orders per buyer, orders per seller, average order value and take rate are divisions. Stored as rates they cannot be aggregated across categories or countries, because summing an average returns the average of averages.

A tree that fails any one of the four still renders. It is simply wrong in a way nobody notices.

Where a Marketplace KPI Tree Is the Wrong Instrument

Four situations defeat this method, and naming them is cheaper than finding them in a board review.

Pre-liquidity volume. With thin order counts, active buyer and active seller totals move on single transactions. The identity holds and the attribution means nothing.

Match quality as the binding constraint. When requests are not being filled, both count-based expansions describe symptoms. The variables to model are requests and matched requests as two summable columns, with fill rate derived between them.

Cross-side causality. The arithmetic ranks contribution size. It never establishes that a seller exit caused a buyer slowdown, however strongly the numbers suggest it.

First-party inventory in the mix. If the platform also sells its own goods, the seller count does not span the whole root, and the supply-rooted identity breaks silently for the portion it does not cover.

In each case the honest move is to state the limit rather than keep decomposing.

Where kpitree.io Sits in a Two-Sided Tree

kpitree.io is a self-service KPI tree builder for finance, business and product analysts. This is the only section of the article about the product.

It holds the structure described above rather than choosing the side for you. Node identities are addition and subtraction, so orders as a sum of category orders and net revenue as a sum of fee lines are native shapes.

Derived KPIs are computed inside the tree by dividing two summable columns, so no row stores a ratio. Orders per buyer, orders per seller, average order value and take rate are all divisions at the node. That is what Test 4 requires, and it is what lets the same tree aggregate correctly across categories.

Three honest limits. The evidenced ingest path today is CSV upload, so a tree is refreshed by uploading a new file. Capabilities the site marks as coming soon are not available. And a tree output is an arithmetic decomposition, not financial, investment or accounting advice.

The working format is one CSV: period, buyers, sellers, orders, GMV. Starting structures across business models are in the KPI tree examples.

Frequently Asked Questions

What is a marketplace KPI tree?

A marketplace KPI tree is a decomposition of gross merchandise value into countable drivers, where parent-child relationships are arithmetic identities rather than themes. Its distinguishing feature is that two complete decompositions exist, one per side of the market.

Why does orders per seller rise when GMV falls?

Because the seller count fell faster than orders did. Orders per seller is a division result, so a shrinking denominator raises it regardless of how any individual seller performed.

Can I build both sides and average them?

Averaging two attributions produces a number that matches neither identity and reconciles to nothing. Build one side as the tree, hold the other as a view, and compare them.

Does GMV or net revenue belong at the root?

GMV if the question is volume and liquidity. Net revenue if the question is monetization. Do not stack them as parent and child through a take rate node, because take rate is derived, not a driver.

Is this the same as a price, volume and mix bridge?

No. A price, volume and mix bridge splits one revenue variance into three additive terms inside a single identity. This article compares two competing identities over the same root.

What is the smallest useful version?

One CSV with five columns, two periods, and one side chosen deliberately. Two rows that reconcile beat a full two-sided model that does not.

Closing: Root the Tree on the Scarce Side and Say So

Most marketplace trees are rooted on the buyer side because buyer data was the data that existed, and the choice was never recorded as a choice.

The fix is one line in the tree's documentation. Name the side the tree is rooted on, name why, and name the condition that would flip it. If sellers become the constraint, the root changes, and everyone should know that in advance rather than during the review.

Then check the arithmetic once. Compute GMV from both sides for one closed period. If the two totals do not match to the dollar, the definitions of active buyer and active seller are not aligned, and every contribution below is unreliable.

Start with one metric and five columns. Upload a single CSV with period, buyers, sellers, orders and GMV, and decompose the quarter your team already argues about.

kpitree.io is a self-service KPI tree builder for finance, business and product analysts.