Abstract
AI systems in marketplaces almost never create a market and almost always govern attention within one. They decide which supply units enter a candidate set, in what order they appear, which seller receives a lead, and how likely a response is. These are operations on allocation, not on the volume of demand.
The distinction matters operationally. If a rise in completed transactions after a model change is measured platform-wide and without a counterfactual, reallocation of attention between sellers cannot be told apart from newly created transactions. The first moves the outcome from one participant to another while the total stays fixed; the second raises the total. The operating and investment decisions implied by those two cases differ.
Central research question
What changes in the state of the market when the algorithmic layer changes, and under what conditions does that change increase the number of realised matches rather than redistribute them?
The layers AI acts on
The decomposition below is a working one. Its purpose is to avoid conflating interventions with different mechanisms and different effect horizons.
| Layer | What the system decides | First-order observable effect | The easy mistake |
|---|---|---|---|
| Eligibility and admission | which supply units participate at all | composition of eligible supply | changing composition changes the denominator of every exposure metric |
| Candidate-set formation | which units are considered for a request | share of supply with no exposure | narrowing the set improves precision and raises zero exposure |
| Ranking | order of presentation | exposure concentration | mean relevance rises while attention inequality rises with it |
| Lead routing | which seller receives an enquiry | time to response, qualified response rate | optimising for response shifts flow to the active rather than the suitable |
| Propensity prediction | likelihood a participant acts | processing priority | the model learns from the previous allocation of attention and reproduces it |
| Pricing and incentives | price, promotion, priority | composition of demand | the effect blends with the algorithmic one and needs separate identification |
The boundary between allocation and demand
Working hypothesis: the algorithmic layer changes how attention is distributed within given supply and demand and does not change the size of demand directly. It affects demand indirectly through experience — a participant whose requests are consistently satisfied returns more often — but that channel operates over months and is observed on the retention curve, not in ranking metrics.
The hypothesis produces testable expectations. In a micro-market with unmet demand, better allocation should raise the share of demand that closes. In a micro-market where supply cannot find demand, better ranking should not produce a durable rise in completed transactions, only redistribution between sellers. Departures from these expectations are the substantive result worth running an experiment for.
Feedback loops
Three loops make a naive effect estimate biased.
First, a ranking model learns from clicks and responses collected under the previous ranking. Units that were never shown have no positive examples and continue not to be shown. The observed quality of such a model improves as the considered set narrows.
Second, sellers adapt to ranking rules. After several iterations, a signal that once correlated with quality begins to reflect skill at optimising for the signal — Goodhart’s law operating at market level.
Third, routing to the most responsive sellers raises short-term response rates while lowering exposure for everyone else, accelerating their departure. The response metric improves over weeks and can degrade market thickness over quarters.
Cold start is two different problems
An absence of data about a new supply unit and an absence of demand for it look identical — no responses — but call for opposite actions. In the first case exploratory exposure helps: a reserved share of impressions that lets new units accumulate signal. In the second case exploratory exposure spends attention for nothing, because no ranking produces demand that does not exist.
The two can be told apart only per micro-market: if mature units in the same micro-market earn qualified responses and new ones do not, it is a data problem; if nobody earns them, it is an absence of demand. This is why separate measurement by micro-market precedes any decision about an algorithmic intervention.
Risks and controls
The NIST AI RMF defines four risk-management functions — GOVERN, MAP, MEASURE, MANAGE — and is treated here as a framework for organising the work rather than a source of claims about marketplaces. Applied to matching it means: record who owns decisions about the allocation of attention; describe the context and affected parties before deployment; measure effects on market outcomes rather than model metrics alone; and define the rollback path in advance.
The distributional effect deserves separate attention. Mean relevance and total conversion can rise at the same time as the share of supply receiving no attention. Both quantities are true; choosing between them is a product and editorial decision that should be made explicitly rather than left as a by-product of a loss function.
Research plan
First, assemble comparable definitions of the layers and observable quantities so that different platforms describe the same thing in the same way. Without this, comparison is impossible.
Second, specify the minimum set of observations needed to test the allocation-versus-demand boundary: share of demand that closes, share of supply with no exposure, exposure concentration, qualified response rate, and time to match, all per micro-market.
Third, describe identification designs available without access to internal systems — natural experiments arising from changes in display rules, discontinuities at threshold admission rules, and shifts during search-surface migrations. Each design carries a negative control.
Fourth, obtain independent methodological review before any quantitative claim.
What this work does not claim
It does not claim that AI in marketplaces is unhelpful, that ranking degrades markets, that any platform commits the errors described, or that the author observed these mechanisms at a particular company. It contains no effect sizes and proposes no ready thresholds.