Skip to content

AI data market

AI data buyer rejection analysis: sort every no before sizing demand

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

AI data buyer rejection analysis turns each declined dataset into evidence about the market. Sort every no into one of three classes: a boundary no, where the buyer does not want the category at all; a conversion no, where demand exists but the supply failed on rights, privacy or quality; and a measurement no, where the evaluation itself was flawed. Boundary rates redraw the market; conversion rates drive the forecast.

Evaluation split

AI data buyer rejection

This page covers AI data buyer rejection analysis: sort every no before sizing demand.Illustrative

Key takeaways

  • Request counts overstate demand unless each declined request is classified; the reason for a no matters more than the count.
  • Boundary declines shrink the market definition, conversion declines lower the forecast and point to supply work, and measurement declines are retested, not counted.
  • Some boundaries are public: OpenAI says it is not seeking sensitive or personal data or information that belongs to a third party.
  • Contamination is a documented reason evaluation buyers want unseen material: GSM1k reports accuracy drops of up to 8 percent on fresh problems in its current version, down from 13 percent in version 1.
  • Concentration magnifies every decline: per Reddit's FY2025 Form 10-K, substantially all of its licensing contract value comes from just two partners.
  • SourceX has not published rejection or demand metrics; ask any company for the dated cohorts and denominators behind its demand claims.

What is AI data buyer rejection analysis, and which decision does it serve?

AI data buyer rejection analysis records why buyers decline datasets, at which stage and out of how many cases, then uses those reasons to correct estimates of how much demand exists and how much of it turns into spending. For an investor, the decision it serves is whether a company's claimed demand is a real, bounded market or a pile of requests that could never have closed.

Most demand evidence in AI data counts the asks and the yeses: announced partnerships, inbound requests, pilots. A count of requests with no rejections attached cannot separate a large market with hard supply problems from a small market padded with requests nobody would have paid for. The declines carry the information that tells the two apart.

Useful evidence has four properties. Each case is one buyer decision on one dataset offer, dated to a cohort. Each decision is classed as a request, an evaluation or a paid commitment. Each rejection carries a coded reason with the evidence behind it. Every rate states its denominator. This page sets out the classes, the ladder on which declines happen, how to weigh the documents and the risks. It reports no measured rejection rates.

Three kinds of no: boundary, conversion and measurement

Every rejection belongs in one of three classes, and each class changes a different number in a demand model. Pool them into one rate and you will either shrink a market that is really supply-constrained or inflate one that does not exist.

A boundary no means the buyer does not want the category at any quality: the task is not one it trains or evaluates for, the modality is wrong, or the material is a kind it has said it will not take. OpenAI's data partnerships page, for example, says it is not seeking datasets with sensitive or personal information, and it asks for large-scale data that is not already easily accessible online. A dataset declined on those grounds shows where that buyer's market ends.

A conversion no means the buyer wants the category but this supply failed: rights could not be documented, personal data could not be removed without destroying the content, the history was too short or the format did not fit. These declines leave the market's edge where it was, lower the share that converts and mark where supply-side work can earn money. When the Data Provenance Initiative audited more than 1,800 public text datasets, popular hosting sites left out license information more than 70 percent of the time and got it wrong at error rates above 50 percent, which suggests rights documentation fails structurally rather than rarely.

A measurement no means the evaluation itself was flawed. Anthropic's engineering guide to agent evaluations notes that when a task fails, the transcript shows whether the agent made a genuine mistake or the graders rejected a valid solution. Dataset reviews fail the same way, for example when a sample is graded against the wrong rubric or tested on a task the buyer has since dropped. Retest these before they enter any rate.

A flow diagram in which every coded rejection splits into three classes, boundary, conversion and measurement, each leading to a different action: redraw the market edge, lower forecast conversion, or retest before counting.
Three kinds of no and the number each one changes. Editorial framework; no measured rates.

The demand-signal ladder, with a rejection rate on every rung

Declines happen on six rungs, from a buyer's published criteria to renewal, and each rung needs its own rejection rate over its own denominator. A single company-wide rejection rate mixes rungs with very different meanings, so ask for the ladder rather than the headline figure.

The rungs are ordered by cost to the buyer: each one takes more of the buyer's time and money than the one before, so a no on a higher rung is stronger evidence that the category is wanted. The table gives the entry evidence, the class a decline usually falls in, the denominator and what the rate tells you.

The demand-signal ladder, with a rejection rate on every rung
RungEntry evidenceUsual class of a no hereRate denominatorWhat the rate tells an investor
1. Published criteriaThe buyer's dated public statement of what it seeks and excludesBoundaryNone: record the page, date and versionWhere the outer edge of that buyer's market sits
2. Written requestA dated request naming task, volume, format and termsBoundary or timingRequests received in the cohortHow much demand remains in bounds
3. Sample evaluationA sample tested against a written acceptance testConversion or measurementRequests that reached a sampleWhether supply quality or test design is failing
4. Rights and terms reviewBuyer counsel reviews provenance and license termsConversionSamples that passed evaluationWhether rights documentation is the binding constraint
5. Paid commitmentA signed agreement with a payment obligationConversion: price, budget or timingCases that cleared rights reviewHow much in-bound demand converts to spending
6. Renewal or expansionA renewal, a second order or a non-renewal noticeConversion or boundary shiftCommitments due for renewal in the windowWhether the demand lasts beyond one purchase

Which documents prove a rejection, and how much weight each deserves

Weigh rejection evidence by how close it sits to a decision with money behind it. A buyer's public criteria page sets edges but no rates; a signed non-renewal notice is close to conclusive; a sales team's summary of why a deal stalled is the weakest evidence of all.

Two standard documents help make reasons comparable across buyers and cohorts. The Data and Trust Alliance's Data Provenance Standards group 22 metadata fields into source, provenance and use, and the alliance says IBM tested them in early 2024 in its clearance process for datasets used to train foundation models. Mapping rights and provenance reason codes to those fields lets one cohort's documentation failures be compared with the next.

Which documents prove a rejection, and how much weight each deserves
EvidenceWhat it can show about a noWhat it cannot showHow to verifyWeight in a demand model
Buyer's published criteria pageThe categories a buyer excludes, as of a dateWhether the buyer applies them case by caseRecord the page version and date; keep a copySets boundaries; contributes no rate
Written decline from the buyerThe stated reason, the stage and the dateThe real reason when the stated one is a cover, such as priceMatch it to the evaluation recordCounts once coded with evidence
Evaluation recordWhich test failed, on which sampleWhether the test was the right oneRead the rubric and the graded examplesHigh when the rubric was written in advance
Interview with an independent budget ownerWhether budget existed and what would have changed the outcomeHow representative the buyer isState the sampling frame and the cohortHigh for telling boundary from conversion
Signed agreement, renewal or non-renewal noticePaid conversion and whether it lastedReasons, unless the notice states themCounterparty confirmation or a public filingHighest
Concentration statement in a public filingHow few buyers stand behind a licensor's revenueAny individual decisionQuote the named filing exactlyContext for weighting buyers, not a rate

Turning rejection rates into a market boundary and a forecast

Boundary rejections are subtracted from demand, conversion rejections set the conversion rate and measurement rejections are set aside until retested. That three-way split is the whole method; the work is in coding each case honestly and keeping open cases in the denominator.

An illustrative cohort, with invented numbers, shows the difference. A company logs 100 written requests in one quarter. Coding finds 30 boundary declines, so in-bound demand is 70. Of those 70, 25 fail on rights documentation, 10 fail on quality or fit, 5 fail on flawed tests and go back for retesting, 12 reach a paid commitment by the end of the window and 18 are still open, counted as censored rather than lost. A request count reports 100 units of demand and a paid share of 12 in 100. The classified view reports 70 in bounds, 12 paid out of 47 decided cases, about one in four, and 25 rights failures that a supplier with documented rights could contest.

Forecast quality improves for a reason investors can check: a model that tracks classes notices when the mix shifts. Dated policy events are the obvious break points. The European Commission's template for general-purpose AI model providers to summarize their training content is mandatory, has applied since 2 August 2025 and becomes enforceable by the AI Office from 2 August 2026, so documentation-related reasons may shift around those dates. Compare cohorts before and after each date instead of pooling them.

Contamination and task realism: why evaluation buyers say no

Evaluation buyers decline material that models may already have seen and material that does not look like real work, and both reasons are documented in public research. Those two reasons put a hard boundary around evaluation demand that training demand does not share.

On contamination, the GSM1k study built 1,205 new grade-school math problems matched to the widely used GSM8k benchmark and found accuracy drops of up to 8 percent when evaluating leading open and closed models, according to the paper's current version; its first version reported drops of up to 13 percent. For a buyer of evaluation data, a dataset that resembles widely published material is a boundary no for testing even if it is acceptable for training. The change between versions carries its own lesson: record the version and date of every source a reason code relies on.

On realism, OpenAI's GDPval, published in September 2025, spans 44 occupations from the nine industries that contribute most to U.S. GDP, with tasks built and checked by professionals averaging over 14 years of experience. Anthropic's evaluation guide likewise advises sourcing realistic tasks from observed failures and writing unambiguous success criteria. A supplier declined because its records do not reflect real work has a conversion problem if it can document who produced the work and how outcomes were judged, and a boundary problem if it cannot.

What could make rejection-based demand research wrong

Rejection data can mislead as easily as request counts if the thesis behind it is wrong. The strongest objection is that announcements and capability progress may never turn into commercial data spending, in which case classified rejections only describe how a small market says no. Each risk below describes one way the method could fail.

  • Buyer concentration can swamp any rate. Reddit's annual report on Form 10-K for fiscal 2025 states that "substantially all of the contract value associated with our licensing revenue is derived from two of our partners." When two buyers dominate, their decisions are the market, and a rejection rate pooled across many small requesters hides that.
  • Stated reasons can be covers. Buyers often name price or timing when the real issue was rights or fit, so a reason without evidence should not move a case between classes.
  • One channel is not the market. Rejections logged from a single company's inbound requests describe that channel, and a company may be tempted to code its declines as conversion failures to imply a larger market than it has.
  • Rules and measurements move. The EU training-content template changes what buyers must document from fixed dates, and GSM1k's own headline figure fell from 13 percent to 8 percent between versions, so reason mixes and the evidence behind them drift over time.
  • Small cohorts mislead. A rate over a few dozen decisions can swing widely from one quarter to the next; ask for counts and intervals, not percentages alone.

Interviewing budget owners so each no lands in the right class

Interview the people who approve spending at buyers, not the supplier's sales team, and ask the few questions that decide the class of each decline. The unit of analysis is one buyer decision on one offer, sampled from every decision in a dated cohort across all channels, with requests, evaluations and paid commitments kept apart.

At a glance

Interviewing budget owners so each no lands in the right class

A full interview protocol for studying rejection reasons already exists on this site; for demand research, the questions below are the minimum that separate boundary, conversion and measurement declines. Code each answer with a primary class, a reason and an evidence grade, keep open cases in the denominator and report every rate with its count:

  • Boundary test: would you have bought this category of data from anyone, at any reasonable price, during this period?
  • Conversion test: if the problem you named had been fixed, would the purchase have gone ahead, and who would have approved it?
  • Measurement test: is the test you ran on the sample the one you would run today, and was it written before you saw the sample?
  • Budget test: was money approved for this purchase, by whom and for which quarter?
  • Substitute test: what did you use instead, in general terms, such as another source, synthetic data or your own records?

Questions to ask any company citing buyer demand, SourceX included

Ask for rejection evidence at the rung where a company claims its demand, using the classes defined here rather than the company's own labels. The same questions apply to SourceX.

For SourceX, the question has a concrete starting point. Its request form asks buyers to state scope, volume, language, format, timing and licensing needs, and datasets are sourced per request instead of being kept in stock, so every sourcing effort starts from a written request, the rung where a rejection record can begin. Its request, rejection and conversion counts are not public, and company and commercial metrics are disclosed to investors directly rather than published. SourceX is seeking capital from accredited investors under Rule 506(c); a reader who asks for materials on its investors page can bring any of the questions below.

Checklist · 6 items

  1. What share of declines in the last four quarters fell in each class, boundary, conversion and measurement, with counts and denominators by rung?
  2. Which conversion reasons have you cured, and how many cured cases were later accepted?
  3. How many distinct budget owners stand behind your written requests, and what share of paid commitments comes from the largest two?
  4. How do you keep open cases in the denominator, and how many are open now?
  5. Who codes each decline, and does anyone outside the sales team check the coding?
  6. Which dated rule or published buyer criterion changed your rejection mix most, and when?

0 of 6 ticked. Tick each item as you go.

Methods note: sources, date and what this page did not measure

The three rejection classes, the six-rung ladder, the evidence weighting matrix and the interview tests are editorial frameworks built for this article on 2026-10-26. They draw on OpenAI's data partnerships page and its GDPval announcement (September 2025), Anthropic's engineering guide to agent evaluations, the GSM1k paper (arXiv 2405.00332, current version, compared with version 1), the Data Provenance Initiative's licensing audit, the European Commission's page on the training-content summary template, the Data and Trust Alliance's Data Provenance Standards and Reddit's annual report on Form 10-K for fiscal 2025. The illustrative cohort is invented and describes no real company.

Not measured: rejection rates at any buyer or supplier, SourceX included; the share of declines in each class; buyer budgets; how often cured conversion failures are later accepted; and how the EU template dates have changed real rejection mixes. None of these can be measured from outside without decision-level records that nobody has published.

Frequently asked questions

Why should investors care about rejected AI data deals?

Because declines show where a market ends and why supply fails inside it. A request count cannot tell a large, supply-constrained market from a small one padded with requests nobody would pay for; classified rejections can. They also show whether the binding problem is rights, quality or demand itself.

What is the difference between a boundary rejection and a conversion rejection?

A boundary rejection means the buyer does not want the category at any quality, so it shrinks the market definition. A conversion rejection means the buyer wants the category but this supply failed, for example on documented rights or de-identification, so it lowers the conversion rate and marks a fixable gap.

Is a high rejection rate a bad sign for an AI data company?

Not by itself. A high rate of conversion declines on documented rights can mean real demand that the company is not yet able to serve, while a high rate of boundary declines means many requests were never in its market. Ask for the rate by class and by rung, with counts.

How can rejection data be gamed?

A company can code declines as conversion failures to imply a bigger market, drop open cases from denominators, or pool rungs into one flattering figure. Ask who codes each decline, what evidence supports each reason and how open cases are handled.

Does SourceX publish rejection or conversion rates?

No. Request, rejection and conversion metrics are unpublished; SourceX reserves company and commercial metrics for investors and shares them directly. Through sourcex.si/investors, accredited investors can request materials and ask how each class and rung is defined.

Sources

  • OpenAI's data partnerships page seeks large-scale datasets not already easily accessible online and says it is not seeking datasets with sensitive or personal information, or information that belongs to a third party. Source
  • The Data Provenance Initiative audited more than 1,800 text datasets and found license omission rates above 70 percent and error rates above 50 percent on popular hosting sites. Source
  • Anthropic's guide to agent evaluations says the transcript of a failed task shows whether the agent made a genuine mistake or the graders rejected a valid solution, and advises realistic tasks sourced from observed failures with unambiguous success criteria. Source
  • The Data and Trust Alliance's Data Provenance Standards organize 22 metadata fields into source, provenance and use standards; IBM tested them in early 2024 in its dataset clearance process. Source
  • The Commission's template for general-purpose AI model providers' training content summary is mandatory, applies from 2 August 2025 and is enforceable by the AI Office from 2 August 2026. Source
  • GSM1k, 1,205 new problems matched to GSM8k, found accuracy drops of up to 8 percent when evaluating leading models, per the current paper version; version 1 reported up to 13 percent. Source
  • OpenAI's GDPval (September 2025) spans 44 occupations from the nine industries contributing most to U.S. GDP, with tasks built and vetted by professionals averaging over 14 years of experience. Source
  • Reddit's FY2025 Form 10-K states that substantially all of the contract value associated with its licensing revenue is derived from two of its partners. 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