Skip to content

Deal economics

Dataset price benchmark methodology: four rules before any number

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

A transparent dataset price benchmark needs four published rules before any number: a sample of executed, arm's-length licenses; fixed definitions of price, unit and date; rights cohorts that compare only like grants; and disclosure rules that protect contributors. Cell size binds hardest: below six deals no distribution-free 95 percent interval for a cell's median exists, and at twenty deals that interval still spans half of them.

Evaluation split

Dataset price benchmark

This page covers dataset price benchmark methodology: four rules before any number.Illustrative

Key takeaways

  • Financial benchmark rules supply the template: IOSCO's 2013 principles anchor inputs in observable arm's-length transactions, and the EU Benchmarks Regulation puts transaction data first where available and appropriate.
  • Five or fewer executed deals cannot support a distribution-free 95 percent interval for a cell's median; at 20 deals the interval runs from the 6th to the 15th price and holds half of them.
  • Each rights term used as a stratum multiplies the sample: an illustrative 40-cell grid needs 240 executed deals for any interval in every cell and 800 for intervals that exclude half the deals.
  • Deal terms are confidential by design, so a benchmark depends on contributed data protected by threshold and concentration rules, set with antitrust counsel.
  • Asking prices and payout ranges can sit only in a labeled lower tier, never averaged with executed deals.

What would a transparent dataset price benchmark have to publish?

Four rule sets, published before the first number: which deals count, what "price" and "unit" mean, which rights cohorts are compared with which, and when a cell may be shown at all. Financial markets already wrote the template. IOSCO's 2013 Principles for Financial Benchmarks say input data "should be based on prices, rates, indices or values that have been formed by the competitive forces of supply and demand and be anchored by observable transactions entered into at arm's length" (Principle 7), and that an administrator "should establish and Publish clear guidelines regarding the hierarchy of data inputs and exercise of Expert Judgment" (Principle 8).

The EU Benchmarks Regulation (Regulation (EU) 2016/1011) turned similar ideas into law for benchmarks in its scope: input data must be "sufficient to represent accurately and reliably the market or economic reality that the benchmark is intended to measure", and "The input data shall be transaction data, if available and appropriate." Neither text covers AI training data. Both show what an investor should demand from anyone who claims to know what datasets sell for.

A data benchmark needs one thing a currency fix does not: rights cohorts. Two licenses of similar records can grant very different things, from evaluation-only use for a year to exclusive training rights with no end date, so the cohort key matters as much as the price field. Transparent here means a reader can rebuild each cohort assignment and each published statistic from the methodology. It does not mean individual deal prices become public; they should not. This page publishes no prices and no SourceX figures.

Why dataset deal prices are so hard to benchmark

The inputs are private by design, few in number and unlike one another. In the US, a company filing a material contract may withhold terms: under 17 CFR 229.601(b)(10)(iv), "The registrant may redact specific provisions or terms of exhibits required to be filed by this paragraph (b)(10) if the registrant customarily and actually treats that information as private or confidential and if the omitted information is not material." Private companies disclose less still.

SourceX is one example of the norm. Its FAQ says "We don't publish our commercial terms" and, on exclusivity, "Neither side publicizes the deal." Its investors page lists Acquisition Partner anonymity: labs can source data without revealing their roadmap or identity to the market. A benchmark built on such deals cannot be scraped; it has to be assembled from confidential contributions under rules every contributor accepts.

Heterogeneity compounds scarcity. Deals differ in permitted use, exclusivity, term, sublicensing, accepted volume, who prepares the records and whether refreshes follow. Exclusivity also limits repeat observations. Under SourceX's terms, for example, the supplier agrees not to license the selected data to anyone else for AI training, with the term set in each deal's Dataset Deal Confirmation. Methods that price an asset from repeated sales of the same item by the same owner have little to work with.

The benchmark specification worksheet

The worksheet turns the four rule sets into eight elements a methodology must fix, the evidence each contributed deal must carry, and the condition that disqualifies the benchmark. Use it to audit any published data price index, survey or "market rate" claim, including one prepared by a company you are evaluating.

Each element is read across its row; the last column is a pass-fail test:

The benchmark specification worksheet
ElementWhat the methodology must fixEvidence each deal carriesDisqualifying failure
Sample frameExecuted, arm's-length licenses of records for AI training or evaluation, signed inside a stated windowSignature date and an independence statement for the partiesLetters of intent, indicative offers or related-party deals counted as transactions
Price definitionNet consideration for the data grant only, separated from services, equity, compute and payments flowing backConsideration schedule split by componentBundled amounts reported as the price of the data
UnitOne unit per data family, counted as delivered and acceptedAcceptance record with the accepted countOffered volume mixed with accepted volume
Dates and currencyInclusion by signature date; one reporting currency with a stated conversion dateSignature and payment dates recorded separatelyPayment years and signature years mixed in one cell
Rights cohort keyThe grant terms used as strata: permitted use, exclusivity, term, sublicensingA rights summary mapped field by field to the keyA deal placed in a cohort its grant does not match
AdjustmentsOnly factors estimated from deals in the benchmark that differ in one key, published with their basisThe paired deals behind each factorJudgment factors with no paired evidence
Cell statistic and minimumMedian with a distribution-free 95% interval; a stated minimum count per cellCount of deals per cellA cell shown with fewer deals than the interval needs
Disclosure and governanceCount, threshold and concentration rules, reporting lag, contributor code, change log and auditContributor attestation for every submissionA published cell from which one contributor's terms can be read

How many deals does one cell need before its median says much?

More than most deal lists hold: a cell needs at least six executed deals before any distribution-free 95 percent interval for its median exists, and about twenty before that interval excludes half of the deals. The median suits prices that a few very large agreements can pull around, and its interval needs no assumption about how prices are distributed. With n deals sorted by normalized price, the interval from the k-th lowest to the k-th highest covers the true median with probability one minus twice the chance of k minus 1 or fewer heads in n fair coin flips.

With five deals the widest possible interval, lowest to highest, covers the median only 93.75 percent of the time, so no 95 percent statement can be made. At six it runs from the lowest to the highest price. At ten it runs from the 2nd to the 9th price and still holds 8 of the 10 deals. At twenty it runs from the 6th to the 15th and holds exactly half. At 21 and 22 the share inside is slightly above half again, and from 23 up to 60 it stays at or below half. At fifty the interval runs from the 18th to the 33rd price, 16 of 50 deals.

For an investor, the practical reading is blunt. A benchmark cell with eight deals can honestly say only that the median lies somewhere between its cheapest and dearest deal. Any narrower claim from a small cell rests on a distribution assumption, and the methodology should state it. The table gives the exact rank bounds:

How many deals does one cell need before its median says much?
Executed deals in the cellRanks bounding the 95% median intervalExact coverageDeals inside the interval
5None reaches 95%93.75% at most (ranks 1 to 5)Not applicable
61st to 6th96.9%6 of 6
102nd to 9th97.9%8 of 10
123rd to 10th96.1%8 of 12
206th to 15th95.9%10 of 20
3010th to 21st95.7%12 of 30
5018th to 33rd96.7%16 of 50
Bars for cells of 6, 10, 20, 30 and 50 executed deals showing the share of deals inside the distribution-free 95 percent interval for the median: 6 of 6, 8 of 10, 10 of 20, 12 of 30 and 16 of 50, with a note that 5 deals allow no 95 percent interval.
Exact order-statistic arithmetic, no prices: a cell needs about 20 executed deals before its median interval excludes half of them.

Stratify or adjust: what rights normalization costs in sample size

Every rights term used as a stratum multiplies the deals a benchmark needs, and every term handled as an adjustment needs its own evidence. Take an illustrative grid: permitted use (training or evaluation only), exclusivity (exclusive or non-exclusive) and term (perpetual or fixed) across five data families, such as support conversations, sales and CRM records, finance and accounting records, software engineering records, and contracts and documents. That is 2 x 2 x 2 x 5 = 40 cells.

At six deals per cell, the grid needs 240 executed, rights-documented licenses before every cell can show any 95 percent interval, and 800 at twenty per cell before every interval excludes half its deals. Turning term into a published adjustment halves the grid to 20 cells, 120 and 400 deals, but only if the benchmark holds deals that differ in term and nothing else, from which the adjustment can be estimated.

A defensible rule follows. Stratify on terms that change what the buyer may do with the records, chiefly permitted use and exclusivity, because no factor can credibly convert an evaluation-only grant into training rights. Adjust only where paired deals exist, and publish the pairs. Handle the brief's other normalizations in the definitions rather than as factors: usable scope through the accepted-unit count, preparation by stripping preparation services out of the price, and buyer purpose through the permitted-use key.

Disclosure rules that keep each contributor's terms private

Publish a cell only when it passes a count test, a threshold test and a concentration test, and set those tests with antitrust counsel. Official statistics offer working designs. The Federal Committee on Statistical Methodology's Statistical Policy Working Paper 22 describes flagging a table cell as sensitive when its count is "less than some specified threshold (the threshold rule)", notes that "For frequency data, most linear sensitivity rules are equivalent to a threshold rule of 3 respondents", and names an (n, k) concentration or dominance rule that limits how much of a published value any one respondent's data can make up.

Price information shared among competitors is also an antitrust question, and in the US the Justice Department's old safety zones for such exchanges are no longer there to lean on. On February 3, 2023 the Justice Department withdrew three health care antitrust policy statements, including safety zones for information exchanges, saying the guidance had become overly permissive on certain subjects, such as information sharing (paraphrase of the release). A data benchmark administrator should treat aggregation, lag and contributor limits as decisions for counsel, not defaults. This is general information, not legal advice.

The table applies three illustrative rules in order (at least six deals; at least three distinct buyers and three distinct sellers; no single buyer or seller party to more than half of a cell's deals) to five invented cells, and reports the first failure:

Disclosure rules that keep each contributor's terms private
Illustrative cellDeals, buyers, sellersBusiest partyOutcome
Support conversations, exclusive, training, perpetual14 deals, 6 buyers, 12 sellersOne buyer in 4 dealsPublish: interval from the 3rd to the 12th price
Sales and CRM, non-exclusive, training, fixed term9 deals, 2 buyers, 9 sellersOne buyer in 6 dealsSuppress: fewer than three buyers
Finance and accounting, exclusive, evaluation only, fixed term5 deals, 4 buyers, 5 sellersOne buyer in 2 dealsSuppress: no 95% interval below six deals
Software engineering, exclusive, training, perpetual22 deals, 5 buyers, 18 sellersOne buyer in 13 deals (59.1%)Suppress: one buyer in more than half
Contracts and documents, non-exclusive, evaluation only, fixed term31 deals, 8 buyers, 25 sellersOne buyer in 7 dealsPublish: interval from the 10th to the 22nd price

Could asking prices and payout ranges fill the empty cells?

No. They can sit in a labeled lower tier, but pooling them with executed deals turns a benchmark into a survey of hopes. IOSCO anchors benchmark data in observable arm's-length transactions, and the EU regulation admits non-transaction inputs such as estimated prices and quotes only where transaction data is insufficient or inappropriate, under guidelines the administrator publishes. A data benchmark can copy that structure as a ladder: executed and paid licenses, executed but unpaid, written offers, then asking prices and marketing ranges, each tier reported separately and never averaged together.

Payout ranges deserve particular care. A range attached to a program describes possible outcomes before scope, rights and acceptance are known; it is not a price anyone paid. SourceX's own process makes the point: its how-it-works page says "Pricing stays indicative until a Dataset Deal Confirmation is signed." An indicative figure, from any company, belongs in the bottom tier or nowhere.

The strongest case for using them anyway is that the alternative is silence, since executed deals are scarce. The better answer is a smaller benchmark with honest intervals and visible empty cells, which tells an investor exactly where evidence runs out.

Six ways a dataset price benchmark could mislead investors

A benchmark that follows every rule above can still mislead. These are the failure modes to test before relying on one:

At a glance

Six ways a dataset price benchmark could mislead investors

  • Selection: contributors report the deals they are proud of, so the median drifts upward; ask what share of eligible deals each contributor submitted.
  • Survivorship: deals that were signed but never reached acceptance and payment disappear unless the sample frame counts them in a separate tier.
  • Concentration: one large buyer active across many cells sets the level everywhere; the concentration rule protects contributors but does not remove the bias, so ask for the count of distinct buyers per published cell.
  • Cohort drift: rights terms change faster than the cohort key, so deals from different years with the same label may grant different things; check the change log.
  • Staleness: the reporting lag that protects contributors also delays the signal, so a published median can describe a market that has moved.
  • Regulation and antitrust: an administrator found to facilitate price coordination could withdraw the benchmark or narrow it; jurisdiction-specific counsel review is a standing cost, not a one-off.

Questions for any data company that cites a price benchmark, SourceX included

Put these questions to any company in the data licensing market that cites a price benchmark, a market rate or a "typical" deal size, and expect written answers.

  • Which deals are in the sample, and are any of them letters of intent, indicative offers or asking prices?
  • How is price defined, and how were services, equity, compute and reciprocal payments separated out?
  • What rights cohort key is used, and which terms are adjustments rather than strata, with what paired evidence?
  • How many executed deals sit in each published cell, and which cells were suppressed or merged?
  • What threshold, concentration and lag rules protect contributors, and who reviewed them for antitrust risk?
  • For SourceX: it does not publish commercial terms, and its licenses are exclusive for AI training with each term set in the deal's Dataset Deal Confirmation, so ask how it evidences any value range it discusses with you. Its capital raise is a Rule 506(c) offering limited to accredited investors; company and commercial metrics go to investors directly, on request through sourcex.si/investors.

Methods note: the interval arithmetic, illustrative inputs and what was not measured

Interval arithmetic: for each n from 2 to 60, a script computes exact binomial tail probabilities with rational numbers and takes the largest k for which the interval from the k-th to the (n - k + 1)-th order statistic covers the median with probability at least 95 percent, assuming no two normalized prices are tied. "Deals inside" counts ranks k to n - k + 1 inclusive. Coverage and shares are rounded half up to one decimal (93.75 percent is shown exactly). The first n with any 95 percent interval is 6; the first with an interval holding at most half the deals is 20; between 20 and 60 only n = 21 and n = 22 hold more than half.

Illustrative inputs: the five data families, the three binary rights keys, the five example cells and their buyer, seller and busiest-party counts are invented to show the rules working, and the publication rules (six deals; three distinct buyers and sellers; no party in more than half of a cell's deals; checked in that order, first failure reported) are design choices for illustration, not regulatory thresholds. The 59.1 percent share is 13 of 22. Script and output: bin/xs0222_benchmark_cells.py and bin/xs0222_benchmark_cells_output.txt in the writer's scratch folder.

Not measured: no deal prices, no market levels, no SourceX transaction data and no survey of existing price indices. Regulatory texts were read through domain-limited search excerpts on iosco.org, eur-lex.europa.eu, nces.ed.gov, justice.gov and ecfr.gov, dated 2026-10-28; SourceX pages were fetched directly the same day.

Frequently asked questions

How many deals does a dataset price benchmark need?

Per cell, at least six executed deals before a distribution-free 95 percent interval for the median exists, and about twenty before that interval excludes half the deals. Multiply by the number of rights cohorts and data families: an illustrative 40-cell grid needs 240 deals for any interval and 800 for intervals that narrow.

Why use the median instead of the average deal price?

A few very large agreements can pull an average far from a typical deal, while the median does not move with them. Its confidence interval can also be built from ranks alone, without assuming any price distribution, which matters when cells are small.

Can rights differences be adjusted instead of stratified?

Only where the benchmark holds deals that differ in that one term and can estimate the factor from them. Terms that change what the buyer may do, such as permitted use and exclusivity, are better kept as separate cohorts, because no factor credibly converts one grant into another.

Are asking prices useless for valuing data?

They show what sellers hope for and can sit in a separately labeled tier. They should never be averaged with executed deals, because a quote that nobody accepted says little about what records actually clear at.

Does SourceX publish dataset prices?

No. Its FAQ says it does not publish its commercial terms; its licenses are exclusive for AI training, and each deal's terms are fixed in that deal's signed Dataset Deal Confirmation. Investors can request materials at sourcex.si/investors.

Sources

  • IOSCO Principles for Financial Benchmarks, Final Report (July 2013): benchmark data should be anchored by observable arm's-length transactions (Principle 7), and administrators should publish guidelines on the hierarchy of data inputs (Principle 8). Source
  • Regulation (EU) 2016/1011, Article 11: input data must be sufficient to represent the market accurately and reliably, should be transaction data where available and appropriate, and may otherwise include estimated prices and quotes under published guidelines. Source
  • Federal Committee on Statistical Methodology, Statistical Policy Working Paper 22 (second version, 2005): threshold rules and (n, k) concentration or dominance rules for deciding which table cells are too sensitive to publish. Source
  • US Justice Department release, February 3, 2023: withdrawal of three health care antitrust enforcement policy statements, including information-exchange safety zones, described as overly permissive on subjects such as information sharing. Source
  • 17 CFR 229.601(b)(10)(iv): a registrant may redact specific terms of filed material contracts it customarily treats as private or confidential if the omitted information is not material. Source
  • SourceX FAQ: SourceX does not publish its commercial terms; licenses are exclusive for AI training with the term set in the Dataset Deal Confirmation; neither side publicizes the deal. Source
  • SourceX how-it-works page: pricing stays indicative until a Dataset Deal Confirmation is signed. Source
  • SourceX investors page: company and commercial metrics are shared with investors directly; Acquisition Partner anonymity lets labs source data without revealing their roadmap or identity. Source

Request SourceX investor materials

SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.

See if your company qualifies