Skip to content

AI data market

SourceX licensed business data report: five ledgers, no rankings

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

The SourceX licensed business data report, planned as the State of Licensed Business Data, is a first-party study with no results published yet. It will keep five ledgers separate for one observation period: supplier inquiries, qualified datasets, accepted deliveries, paid transactions and buyer requirements. Figures appear only above a minimum cell size and after approval gates, are never summed across ledgers, and never become market-wide claims or rankings.

License terms

SourceX licensed business

This page covers sourceX licensed business data report: five ledgers, no rankings.Illustrative

Key takeaways

  • The report is a published design, not a result: no observation period has closed and no SourceX figures exist yet.
  • Five ledgers, from supplier inquiries to buyer requirements, keep their own units and denominators and are never added together.
  • Public evidence is thin and unreliable: a few filings, buyer wish lists and license metadata with omission rates above 70 percent on popular hosting sites.
  • Every figure must pass permission, confidentiality, a 10-or-fewer cell rule and privacy review, or be reported as withheld.
  • Findings describe companies that met SourceX's eligibility rules in the period; no rankings, prices or market-wide claims appear.

What will the SourceX licensed business data report measure, and what exists today?

It will measure what SourceX itself observes in one fixed observation period, kept in separate ledgers, and nothing beyond that. Today the report exists only as this design: no observation period has closed, no figure has been approved for release and no results exist.

Licensed business data here means records an operating company created in the course of its own work, such as support tickets, CRM histories, claims files or engineering tickets, offered for AI training or evaluation under a signed agreement. The scope excludes public web content, synthetic data, work produced to order by labeling vendors and consumer data sold by brokers, because each has a different supply chain and would blur every denominator.

Publishing the design before the data is the point. Readers can see which questions the report will answer and which it never will, and the definitions are fixed before any result could tempt a change. The SourceX research methodology page sets the general standard; this page applies it to one report.

Why public sources cannot describe licensed business data yet

Most licensing agreements are private, and the public record that exists is thin, concentrated and often mislabeled. A first-party report sees a different population: the companies and datasets that actually went through a licensing process, including the ones that dropped out.

The best public disclosures show how narrow the window is. Reddit's S-1 (2024) reported data licensing arrangements signed in January 2024 with an aggregate contract value of $203.0 million and terms of two to three years, and its annual report for 2024 says substantially all of the contract value associated with its licensing revenue comes from two partners. That is precise information about one platform and two buyers, not about the operating businesses whose records this report covers.

Buyers' own statements describe what they want rather than what they bought. OpenAI's data partnerships page asks for large-scale datasets that are not already easily accessible online and says it is not seeking sensitive or personal information or information that belongs to a third party. Useful as a requirement, it says nothing about how often such data is found, cleared or delivered.

Metadata is weaker still. In its 2023 audit of 44 data collections, which together hold 1,858 fine-tuning datasets, the Data Provenance Initiative found that widely used dataset hosting sites left out the license more than 70 percent of the time, with error rates topping 50 percent. A license label copied from a listing is a lead to check, not evidence, so this report codes rights from signed documents only. The public-source protocol on this site, the data licensing market report methodology, counts disclosed agreements; this report counts first-party process records, and the two are shown side by side, never merged.

Five ledgers that are never added together

The report keeps five ledgers, each with its own unit, entry event, resolution and denominator, and never adds one to another. A company can sit in the inquiry ledger and appear in no other, and a buyer request is evidence of interest, not of a purchase. Quoting the largest ledger as if it were the market is the error this structure exists to prevent.

Every ledger reports cases still open at the close of the period as open, neither dropped nor counted as failures, and states its entry window in calendar dates. Companies that applied on their own and companies SourceX sourced are reported separately, because they arrive with different odds of qualifying.

The figure shows the five ledgers as separate columns, the withheld cells that the disclosure rules create, and the boundary every finding has to stay inside.

Five ledgers that are never added together
LedgerUnit countedEnters whenResolves whenSupports claims about
Supplier inquiriesA company that completed the assessmentAssessment submitted in the periodQualified, declined, withdrawn or open at closeWho applies and why companies are declined; not supply
Qualified datasetsA defined dataset scope that passed rights, privacy, quality and fit reviewScope approved for presentationAccepted, rejected, withdrawn or open at closeCategory coverage and preparation effort; not demand
Accepted deliveriesA delivery a buyer accepted under its acceptance testDelivery submitted for acceptanceAccepted, rejected or open at closeAcceptance and rejection reasons; not payment
Paid transactionsA deal whose supplier payment was madeClosing recordedPaid or outstanding at closeElapsed time from closing to payment; not deal values
Buyer requirementsA written data requestRequest receivedMatched, declined, withdrawn or open at closeRequested record types, volumes and license scopes; not purchase commitments
Five separate ledger columns, inquiries, qualified, accepted, paid and requests, with some cells withheld under the disclosure rule, crossed plus signs showing they are never added together, and a dashed boundary keeping market-wide claims and rankings out.
The report's five ledgers and its boundary. Schematic: no counts are shown because none have been published.

Six exhibits, each with a rule it must pass

Six exhibits answer the questions this report was commissioned to address: category coverage, preparation effort, acceptance reasons, buyer needs, drop-out reasons and stage timing. Each has a publication rule, and an exhibit that cannot meet its rule is published as withheld, with the reason, rather than silently left out.

The acceptance reasons exhibit uses a closed list so that reasons can be counted rather than paraphrased. The draft list is short on purpose: required fields missing, links between related records broken, records outside the agreed date range, an agreed exclusion not applied, a de-identification defect, volume below the agreed scope, and accepted after cure. Any reason outside the list is coded Other and reviewed before the next edition, so the list grows only on evidence.

No exhibit reports deal values, prices, payouts, fee terms or named companies. They are excluded by design, and changing that would need a separate approval and a public fact record before any such figure appeared.

The report's anonymized evidence summaries follow the same discipline. Each summary describes a pattern seen in at least the minimum number of cases, written fresh in the report's own words, never one company's story with the names removed.

Six exhibits, each with a rule it must pass
ExhibitQuestion it answersLedgerPublication rule
Category coverageWhich record systems and workflows reached qualificationQualified datasetsCategories below the minimum cell size merge into Other
Preparation effortHow many hours each preparation stage took, as a median and rangeQualified datasets and accepted deliveriesShown only where at least the minimum number of datasets contributes
Acceptance reasonsWhy deliveries were accepted, cured or rejected, by coded reasonAccepted deliveriesReasons coded from a closed list; buyer standards never quoted
Buyer needsWhich record types, volumes, languages and license scopes buyers asked forBuyer requirementsReported as request counts, never as demand or commitments
Drop-out reasonsWhere companies and datasets left the process, including rights findingsSupplier inquiries and qualified datasetsNegative results published at the same detail as positive ones
Stage timingElapsed days between ledger events, with open cases countedAll ledgersMedians with the number still open; no averages of closed cases alone

Four gates every figure passes before it is public

A figure is published only after it passes four gates: contractual permission, confidentiality, disclosure limitation and privacy review. A figure that fails any gate is withheld, and the report states how many were withheld and why.

The cell-size rule borrows from public statistics practice. The data use agreement terms of the Centers for Medicare and Medicaid Services bar displaying any cell of 10 or fewer and bar using percentages or other mathematical formulas that would reveal such a cell. Licensed business data involves far fewer companies than Medicare involves patients, so the same threshold will suppress many cells in an early edition. That is the price of never exposing a single supplier or buyer.

  • Permission: counsel confirms that the relevant agreements allow aggregate statistical use; datasets whose agreements do not are excluded and reported as an excluded count.
  • Confidentiality: no supplier, buyer or partner is named, or identifiable from a combination of category, company size and region.
  • Disclosure limitation: no cell describing 10 or fewer companies, datasets or requests is shown, nor any percentage or difference that would reveal one; complementary cells are suppressed where needed.
  • Privacy review: no figure is derived from personal data fields, and evidence summaries are rewritten, never excerpted from records.
  • Approval record: each exhibit carries the date and role of its approver, kept with the versioned report.

Where SourceX experience ends and industry conclusions begin

Every finding describes the companies and datasets that reached SourceX under its eligibility rules in the period, not the market. The report states that boundary in each exhibit title, not only in a limitations section readers may skip.

The selection is not random, and its rules are public. SourceX works with U.S., Canadian and EU companies that had 50 or more full-time employees at their peak, and can consider companies in several other countries when the offered data is at least 90 percent English, measured on the records actually offered. Its general guide is that buyers look for two or more years of consistent history and tens of thousands of records in one system, though smaller sets can qualify when the work is rare. Small companies, non-English archives and short histories are therefore under-represented by design, and the report cannot speak for them.

The table turns that boundary into wording rules an editor can enforce line by line. Some questions stay out of reach however much data the report collects, and the list below says so in advance.

  • How large the market for licensed business data is, or how fast it grows: the ledgers have no market denominator.
  • What any dataset is worth: prices and deal values are excluded, and one deal does not set a price for another.
  • Whether buyers in general want a record type: requests reach SourceX through its own channels and reflect who chose to ask.
  • Why a delivery was accepted, in a causal sense: the exhibits show associations, such as preparation hours alongside acceptance, not effects.
  • How companies that never applied would fare: they are outside every ledger.
Where SourceX experience ends and industry conclusions begin
Statement typeAllowed formNot allowed
Share within a ledgerOf qualified datasets in the period, a stated share came from ticketing systemsMost business data sits in ticketing systems
Change between periodsBoth counts, both periods and any change to eligibility rulesA growth rate for the market
Comparison with public sourcesSide by side, labeled first-party and public-sourceA blended figure
Buyer requestsRequests received named these record types and volumesBuyers will pay for these records
RankingsNoneAny ranking of companies, categories, or SourceX against peers

How each dataset is coded so a second team could recount it

Each qualified dataset is coded from a closed codebook whose fields follow published documentation standards, and every coded value points to the document it came from. Coders mark a field unknown rather than infer it, because inferred values are where reports quietly inflate.

The fields follow Datasheets for Datasets (Gebru and colleagues, first posted in 2018), which proposes that every dataset be accompanied by a datasheet documenting its motivation, composition, collection process, recommended uses and so on. For licensed business records that means the source system and workflow, the date range, the record types, the preparation applied, the permitted uses and the maintenance plan.

Source categories follow the European Commission's template for the public summary of training content, which general-purpose AI providers must use from August 2, 2025, with models already on the market given until August 2, 2027. Its list of data sources separates public datasets, private datasets, scraped web data, user data and synthetic data. Coding to the same categories keeps the report compatible with what providers already have to disclose.

Two coders work independently on the rights and category fields. Disagreements go to an adjudicator who applies the written rule, and agreement is reported per field, so readers can see which fields were coded consistently and which were not.

A versioned research agenda, dated from today

The agenda is versioned and dated, and each version names the gate it must pass rather than a promised release date. This page is version 0, dated October 26, 2026.

Four questions lead the first edition, each chosen because its answer could reflect badly on the process as easily as well: which record systems most often fail rights review, how preparation hours are spread across stages, which acceptance reasons recur, and which buyer requests go unmatched. Null and negative results are published at the same level of detail as positive ones.

A versioned research agenda, dated from today
VersionContentsGate before release
v0, October 26, 2026This design: scope, ledgers, exhibits, gates and boundary rulesEditorial, legal and privacy review of the design
v1First observation period, with every exhibit whose cells passPeriod closed, open cases counted, all four gates passed
v1.xCorrections and cases that resolved after the closeDated correction note; earlier figures kept online
v2Second period and the first period-to-period comparisonUnchanged definitions, or old and new figures shown side by side

Using the design now if you need business data

Write data requests in the same fields the report counts, so a request can be matched, tested and, in aggregate, counted. Datasets at SourceX are found at U.S. businesses after a request arrives, not drawn from stock. The request form covers scope, volume, language, format, timing and licensing needs; a request that also names its intended use and an acceptance test becomes precise enough to act on.

At a glance

Using the design now if you need business data

Suppliers see the same structure from the other side. For them the process opens with a short company assessment, no data uploads required, and nothing leaves the company before it has approved the scope and terms and signed. A dataset-level quality report, covered in this site's guide to quality report contents, is a different document: it describes one dataset for one buyer, while this report describes patterns across many, under the gates above.

  • Record type and the system it lives in, such as tickets from a help desk or claims from a claims platform.
  • Date range and the minimum history that would make the records useful.
  • Volume, with the smallest quantity that would still answer the question.
  • Language and format, including whether attachments and links between records must survive.
  • License scope: training, evaluation or both, and any exclusions the use requires.
  • Intended use and the acceptance test, written so a delivery can pass or fail it.

Frequently asked questions

Has SourceX published the State of Licensed Business Data report?

No. This page is the report's design: scope, ledgers, exhibits, gates and boundary rules. No observation period has closed and no results have been published.

Will the report rank data companies, categories or SourceX?

No. The design forbids rankings of any kind, including SourceX against peers, because a first-party sample has no comparable peer denominator. It reports shares and counts within SourceX's own ledgers only.

Will the report publish prices or deal values?

No. Deal values, prices, payouts, fee terms and named companies are excluded by design. Changing that would require a separate approval and a public fact record before any such figure appeared.

How is this different from a market report built on public deals?

A public-source report counts agreements someone chose to announce, which is a small and skewed slice of activity. This report counts first-party process records, including companies and datasets that dropped out, so its population and its blind spots are different; the two are shown side by side, never merged.

Could a supplier or buyer be identified from the report?

The design is built to prevent it: no names, no combinations of category, size and region that point to one company, and no cell describing 10 or fewer companies, datasets or requests. Counsel also confirms that the underlying agreements permit aggregate statistical use before anything is published.

Sources

  • Reddit's S-1 disclosed data licensing arrangements entered in January 2024 with an aggregate contract value of $203.0 million and terms of two to three years. Source
  • Reddit's annual report for 2024 says substantially all contract value associated with its licensing revenue comes from two partners. Source
  • OpenAI's data partnerships page seeks large-scale datasets not already easily accessible online and says it is not seeking sensitive or personal information or information that belongs to a third party. Source
  • The Data Provenance Initiative audited 44 data collections comprising 1,858 fine-tuning datasets and found license omission of 70%+ and error rates of 50%+ on widely used dataset hosting sites. Source
  • The Data Provenance Initiative's repository describes a large-scale audit of 44 data collections spanning 1,800+ fine-tuning text-to-text datasets, documenting sources, licenses and creators. Source
  • CMS data use agreement terms bar displaying any cell of 10 or fewer and bar using percentages or other mathematical formulas that would reveal such a cell. Source
  • Datasheets for Datasets proposes that every dataset be accompanied by a datasheet documenting its motivation, composition, collection process, recommended uses and so on. Source
  • The European Commission's mandatory template for the public summary of training content lists data sources by public datasets, private datasets, scraped web data, user data and synthetic data, applying from 2 August 2025, with models already on the market given until 2 August 2027. Source

Tell us what data you need

Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.

See if your company qualifies