Synthetic data sells — if you package it honestly

The most counterintuitive fact about the data economy: you don’t need to own real data to sell data. A synthetic dataset — one that preserves a real source’s statistical shape without exposing a single real record — is a legitimate, growing product category, and it widens who can be a data seller from “companies that own data” to “anyone who can model it.”

Why buyers want data that isn’t real

Four pressures push buyers toward synthetic, and none of them are fads: privacy regulation (GDPR and its siblings) makes real personal data hard to share across organizations; collecting and labeling real data is slow and expensive; real datasets systematically lack rare events (fraud, failures, edge cases) that models most need to learn; and sensitive domains — health, finance — often can’t be shared at all, while their shape can. The tooling matured accordingly: SDV, out of MIT’s Data to AI Lab, made tabular synthesis (GaussianCopula, CTGAN, TVAE) a standard open-source workflow, and the commercial end consolidated hard — NVIDIA acquired Gretel in 2025. Synthetic data for pipeline testing, ML prototyping, and privacy-safe sharing is a normal purchase now.

The trust problem — and why provenance is mandatory here

Synthetic data has a lemon problem: a buyer can’t tell a carefully-modeled statistical twin from noise with column names. The fix isn’t marketing — it’s declared provenance. On datapoint.market a synthetic listing must say:

create_dataset(
  …,
  synthetic = true,
  synthetic_method   = "CTGAN (SDV 1.x)",          # how it was generated
  synthetic_source   = "internal claims DB, 2019–2025",  # what it models
  synthetic_fidelity = "KS < 0.05 on all marginals",     # required reading
  synthetic_privacy  = "epsilon=1.0 differential privacy"
)

Method and source are required — publish flags a synthetic listing without them. Every listing then carries dataset_synthetic: true and the provenance block prominently: buyers always know they’re getting synthetic data and on what basis. Honesty is enforced by the platform because honest packaging is what lets the category command real prices. (The platform sells synthetic data and polices its provenance — it doesn’t generate it; that craft is yours.)

How to package one that sells

Name the source and era you modeled — “synthetic e-commerce transactions, modeled on 2023–2025 EU order flows” beats “synthetic transactions”.
Quantify fidelity — marginal KS scores, correlation preservation, downstream-task parity. Buyers making ML decisions read these first.
State the privacy guarantee — if you trained with differential privacy, say the epsilon; if you didn’t, say how you audited leakage.
Join card still applies — synthetic data gets fused too. If your synthetic customers key on region and month, declare it so agents can plan the join.
Not just files: a generator endpoint — “one synthetic record per call” — is a live API listing with the same provenance contract.

Buyers find the whole category with search_marketplace(type="synthetic").

The honest pitch to sellers

If you can model a domain — because you work in it, research it, or hold data you could never share raw — you have a sellable product with no cold-start inventory problem and no privacy exposure. The only rule is the one the platform enforces anyway: say exactly what it is and how it was made.

Guides: selling synthetic datasets · join cards