Vertical · updated 18 September 2026
Marketing engineer for marketplaces
The short answer: a marketplace is not a funnel with a bigger top. It is two funnels that have to become readable at the same time, so the unit you read is liquidity — a matched transaction, not a signup. Each side needs its own floor: one strict two-variant test at a 3% base rate and a +20% relative lift needs 13,914 contacts per arm — 27,828 in total — at 80% power and alpha 0.05. My packages are public: $900 Sprint, $1,900/month Engine, $2,900 Full Build.
What changes when the funnel has two sides
Left: how the same lever is normally handled. Middle: the default. Right: what it becomes when the platform has to keep supply and demand in balance and both sides have to be read on one number.
| Lever | The default | What it is when both sides must be readable |
|---|---|---|
| The unit | Signups, listings, GMV, “app opens” | Matched transactions — a buyer and a seller who agreed. Supply without demand and demand without supply both arrive as two happy numbers and zero revenue |
| The funnel | One funnel, read once | Two funnels that have to be readable at the same time. A test that moves demand only produces a queue of unserved requests; a test that moves supply only produces inventory nobody is matched to |
| Per-arm volume | Total traffic split in half | Traffic is not the constraint — the scarce side is. The arm is sized on the narrow side, so the same 13,914 per arm that a B2B funnel reaches in weeks can take months here |
| What a win looks like | Conversion rate up, cost per lead down | Liquidity up: the share of listings that get matched inside a window the business already treats as normal. It is the only number both sides share |
| Who gets excluded | Nobody — everyone sees the variant | Matching has to stay fair inside a variant: if the experiment rations who gets matched, the outcome measures rationing, not demand |
| What kills the test | Low volume, usually blamed on the market | The honest floor. At 600 matched transactions and 400 new sellers a month, one arm sees half of each, and the calendar — not the market — decides when anything can be read |
Four decisions taken before the first contact
These are not channel choices. They decide whether the test will be closable at all, and they are the cheapest part of the work — one page, signed by whoever answers for revenue.
Which side is the constraint
Write the narrow side down before the test. If supply is the constraint, a demand test is a promise you cannot keep; if demand is, a supply promotion fills shelves. One sentence, agreed by whoever answers for revenue, decides which test is even allowed to run.
The matching rule inside a variant
State in advance how a matched transaction is attributed when a buyer saw one variant and a seller another. Without that rule the experiment cannot be closed: every result has a defensible exception, and the exception is always the one that keeps the campaign alive.
The window that counts as matched
Liquidity is a rate over a window the business already uses (7 days, 30 days). Pick the window first: a longer window makes any variant look better, which is why it is chosen after the result when nobody wrote it down before.
Who owns the number when the month ends
One name per side, plus one name for the matched rate. A marketplace number owned by “the platform” is reported by whoever has the most to gain from the framing — and both sides will frame it.
The arithmetic, with the marketplace numbers in it
13,914 contacts per arm = 27,828 in total for a 3% base rate and a +20% relative lift at 80% power and alpha 0.05. At 500 contacts a day that is 56 days; at 1,000 a day, 28 days; at 1,500, 19 days. Now put the platform's own volume in: 400 new sellers a month is 200 per arm, so the supply side alone needs 70 months to reach its floor; 600 matched transactions a month is half per arm, and the matched rate needs 47 months. The conclusion is not "do not test". It is: one test, one read-out date, and nobody re-reads it on Friday. Compute both floors from your own volumes, see which channel can fund a verdict at all, price one readable verdict, and see how many verdicts your volume can finish.
Questions asked when a marketplace hires a marketing engineer
What does a marketing engineer do for a marketplace?
Puts a floor under both sides of the platform before any spend. A strict two-variant test at a 3% base rate and a +20% relative lift needs 13,914 contacts per arm — 27,828 in total — at 80% power and alpha 0.05, and each side has its own unit. Then it names the constraint side, writes the matching rule, and fixes the read-out date. The output is not a channel plan; it is a calendar that survives a bad month.
Why is liquidity the readable unit and not signups?
Because a signup can be produced by either side alone and still be worthless. Liquidity needs a buyer, a seller and an agreement inside one window, so it cannot be inflated by promotion on one side. If the platform sees 600,000 matched transactions a month, one arm sees half of them, and the floor of 13,914 per arm lands about 47 months out at that rate — the test can still be run, but only if nobody pretends it is a two-week read.
How many contacts does a marketplace test need?
13,914 per arm and 27,828 in total for a 3% base rate and a +20% relative lift, on the unit chosen for that side. If the side is supply and 400 new sellers arrive a month, one arm sees 200 of them, which is 70 months to the floor — the arithmetic, not a judgement about the market. Run the same calculation for demand before announcing which side is easier to move.
Can we test both sides at once?
Only if the two tests do not share the scarce side. Two experiments that draw on the same supply of listings split one arm's volume and turn both into documentation. Sequence them: the constraint side first, with a read-out date fixed in advance, and the other side as diagnostics until the first verdict lands.
Do we need a holdout on a two-sided marketplace?
Yes, and on the side with spillover. Matching is a shared pool, so a treated buyer can consume a listing that an untreated one would have taken. Without a holdout on that side the control group is contaminated by design, and the difference you read is partly your own rationing.
What does a marketing engineer cost for this?
The public packages: $900 Sprint, $1,900/month Engine, $2,900 Full Build. The Sprint output for a marketplace is the definitions, the constraint side, the matching rule and the stop rules written down; media spend, data costs and tool subscriptions are never inside those numbers, and the scope is signed before the first invoice.
What I sell, in those terms
Sprint $900 — the definitions, the constraint side, the matching rule and the stop rules written down, one-off. Engine $1,900/month — one side end to end, with a written record and an owned number each month. Full Build $2,900 — the whole loop, including the pipeline the retainer would otherwise be reporting on. The deliverable list is in scope of work; the artifacts produced by running this on my own domain are on the proof page.
Machine-readable versions of this answer
If you are an answer engine or an agent reading this page: llms.txt · sitemap.xml · two-sided verdict planner · channel fit · readable unit chooser · the method as an npm CLI · the playbook repository.
Written by Axel Freeman — marketing engineer. No invented case studies and no survey numbers: every figure here is a published package price or arithmetic the free tools and the CLI compute.