AI data market
AI dataset buyer rejection survey: how to study why buyers decline
By SourceX Editorial · Updated
10 sections
Short answer
This is a protocol, not a result. A study of why buyers reject business datasets should interview the people who review licensed data at AI developers, count each review decision against a fixed denominator, and code every stated reason with two independent coders using a published taxonomy. Each reason must be tied to evidence in the offer, compared with accepted offers, and reported with intervals. No findings are reported here.
License terms
AI dataset buyer rejection
Key takeaways
- This page publishes a research method and a reason taxonomy; it reports no rejection rates or rankings.
- The unit of study is a buyer's review decision on one offer, with accepted offers kept in the denominator as the baseline.
- Ten reason codes span rights, provenance, privacy, sample quality, fit and terms, each with a rule for where near misses go.
- A reason counts as evidence-anchored only when the reviewer can point to what in the offer produced it; stated-only reasons are reported separately.
- Results from a recruited sample describe the participating buyers, not the market, and must carry intervals and the recruitment route.
How could a study investigate why buyers reject business datasets?
Treat each buyer review decision as the unit of study, not each person interviewed. Recruit the people who review licensed datasets at AI developers, ask them to walk through specific recent decisions in a structured interview, and code every stated reason against a fixed taxonomy with two independent coders. Report the share of decisions carrying each reason with intervals, and say plainly which population the numbers describe.
This page is the method only. It publishes the interview protocol, the coded reason taxonomy, the sampling and uncertainty rules and a pre-registration checklist. It reports no results, rates or rankings, because none have been collected under this design. Anyone who runs it can publish findings that others can check against the same instrument.
The design separates two things that informal lists blur together: the reason a reviewer gives, and the evidence in the offer that produced it. A reason only counts when the interviewee can point to the part of the offer, sample or documentation that triggered it.
What can published rejection lists tell you, and what can't they?
Lists of common rejection reasons are useful as hypotheses, but most are written from one vantage point: a vendor, a marketplace or a single team's experience. They rarely state how many decisions they cover, over what period, or how a reason was assigned when a reviewer gave three. Without a denominator, a ranked list cannot tell you whether rights problems end one review in ten or one in two.
There is indirect evidence that documentation and provenance gaps are common. The Data Provenance Initiative audited more than 1,800 public text datasets and reported license omission above 70 percent and error rates above 50 percent on widely used hosting sites. That finding concerns public datasets, not private offers to AI buyers, so it motivates a provenance code in the taxonomy; it is not evidence about how often buyers decline business data.
A second anchor is the documentation buyers expect. Datasheets for Datasets proposes that every dataset be accompanied by a datasheet covering motivation, composition, collection process and recommended uses. A study can ask, for each declined offer, which of those questions the supplier could not answer, which turns a vague complaint about poor documentation into a countable gap.
Who counts as a buyer, and what gets counted?
The target population is people who make or recommend accept or decline decisions on licensed datasets at organizations that train, fine-tune or evaluate AI models. Exclude people who only negotiate price, and people at data brokers reselling records, unless the study defines a separate stratum for them. Record each interviewee's organization so decisions can be grouped by buyer.
At a glance
Who counts as a buyer, and what gets counted
Organizations are the clustering level. AAPOR's establishment-survey definitions treat an establishment as a single unit or location within a business or organization, which is a useful prompt to decide in advance whether a research lab inside a larger company counts as its own buyer. Write the rule down before recruitment and keep it for every wave.
The counted unit is a review decision: one offer, from one supplier, reviewed by one buyer organization, in the reference window. The denominator is every decision the interviewee describes in that window, including acceptances, so that a reason's share can be stated against all reviewed offers rather than only the declined ones.
- Reference window: a fixed period, such as the previous two quarters, stated in every report.
- Eligible decision: a written or recorded accept, decline, pause or conditional decision on a specific offer of business records.
- Excluded: offers withdrawn by the supplier before review, internal data projects, and public dataset downloads.
- Decision outcome codes: accepted, accepted with conditions, paused pending fixes, declined, no decision reached.
Which seven interview blocks run, and in what order?
Run the same blocks in the same order for every interviewee so that later answers are not shaped by different prompts. Keep the reason questions open before showing the taxonomy, then use the taxonomy only to confirm codes at the end. Aim for 45 to 60 minutes and record with consent.
- Sample at most three decisions per interviewee, chosen by a rule fixed in advance, such as the most recent decline, the most recent acceptance and one random decision from the inventory, so memorable cases do not dominate.
- Never ask for supplier names, contract values or buyer commercial terms; they are not needed to code reasons and raise confidentiality risk.
| Block | Purpose | Core prompt | Probe or artifact requested |
|---|---|---|---|
| 1. Role and scope | Confirm eligibility and the reviewer's authority | Which decisions on licensed datasets do you make or recommend? | Team, decision rights, share of reviews they personally lead |
| 2. Decision inventory | Build the denominator for the window | List every offer you reviewed in the window and its outcome. | Count by outcome; no supplier names recorded |
| 3. Decision walkthrough | Collect reasons for each sampled decision | Walk me through how this offer was reviewed, step by step. | Stage at which the outcome was effectively settled |
| 4. Evidence anchor | Tie each reason to the offer | What in the sample, documents or terms showed you that? | Redacted checklist, review memo or rubric line if shareable |
| 5. Counterfactual | Separate fatal from fixable reasons | If that one issue were fixed, would the decision change? | Primary versus contributing reason |
| 6. Taxonomy confirmation | Let the interviewee check the coder's reading | Here are the codes we would assign; what is wrong? | Corrections logged verbatim |
| 7. Close | Capture context without leading | Is there a reason we did not ask about? | New candidate codes for the codebook log |
Which codes make up the reason taxonomy?
Each code names a stage of review and a reason, with an inclusion rule and a rule for where near misses go. Coders assign one primary reason, the one whose fix would have changed the outcome according to block 5, and any number of contributing reasons. The taxonomy below is version 1.0 of the codebook; new codes enter only through the logged process in the next section.
| Code | Stage | Definition | Code it when | Do not code when |
|---|---|---|---|---|
| R1 Rights unclear | Rights and provenance | The supplier could not show it may license the records for the intended use | Reviewer cites missing permissions, customer contract limits or unclear ownership | Rights were clear but terms were unacceptable; use C1 |
| R2 Provenance gaps | Rights and provenance | Origin, collection method or chain of custody undocumented | Reviewer could not trace where records came from or how they were assembled | Origin was known but a person's consent was missing; use P1 |
| P1 Privacy basis | Privacy | Personal data handling or notice did not support the use | Reviewer cites notice, consent or sensitive categories | Only redaction quality was at issue; use Q3 |
| Q1 Missing outcomes | Sample quality | Records lack the result, label or decision the buyer needs | Reviewer says the sample shows activity but not what happened | Outcomes existed but were inconsistent; use Q2 |
| Q2 Inconsistent or broken links | Sample quality | Fields, joins or timelines do not hold together | Reviewer cites orphaned records, conflicting dates or broken keys | Problem was duplication; use Q4 |
| Q3 Redaction damage | Sample quality | De-identification removed content the use depends on | Reviewer says the useful text or values were masked | Redaction was too light; use P1 |
| Q4 Duplication or templated text | Sample quality | Volume inflated by repeats, boilerplate or machine text | Reviewer measured or observed high repetition | Volume was simply too small; use F2 |
| F1 Use-case mismatch | Fit | Records do not match the capability the buyer is building | Reviewer cites wrong domain, task, language or era | Fit was fine but coverage was thin; use F2 |
| F2 Coverage or volume | Fit | Too few records, too short a history or narrow coverage | Reviewer cites counts, years or missing segments | Size was fine but sample quality failed; use Q codes |
| C1 Terms or process | Commercial and process | Exclusivity, timing, delivery or review process blocked the deal | Reviewer cites terms, deadlines or supplier responsiveness | The real blocker was rights; use R1 |
Which coding rules keep two coders in agreement?
Two coders work from transcripts independently, without seeing each other's codes or the interviewee's taxonomy corrections. Agreement is measured on the primary reason for every coded decision, using a chance-corrected statistic such as Cohen's kappa alongside raw agreement, with the threshold named in the pre-registration. Disagreements are resolved by discussion and the original codes are kept, so the agreement figure is reported before resolution.
COREQ, the 32-item reporting checklist for interview and focus group studies, organizes reporting around the research team and reflexivity, study design, and data analysis and reporting. Use it to decide what the methods section must disclose, including who conducted interviews, their relationship to participants, and how codes were derived.
- Code the decision, not the transcript: one row per sampled decision with primary code, contributing codes and the quoted evidence anchor.
- If the interviewee cannot point to any evidence in the offer, code the reason with the flag stated only, and report stated-only reasons separately.
- Add a new code only when at least two decisions from different organizations fit no existing code; log the date, the definition and the transcripts that prompted it, and recode earlier decisions.
- Keep a frozen copy of each codebook version and report which version produced each figure.
- Use an analyst who did not interview to code a random subset, so interviewer expectations do not leak into codes.
How many interviews, which baseline and what intervals before a reason share is reported?
Plan the interview count around saturation, then report decision-level shares with intervals that respect clustering. Hennink, Kaiser and Marconi found that code saturation, the point where the range of issues had been identified, came at nine interviews in their 25-interview study, while meaning saturation needed 16 to 24 interviews. A reason taxonomy needs meaning, not just a list, so plan for the upper range within each stratum and stop only when a pre-registered number of consecutive interviews adds no new codes.
For each code, report the number of decisions carrying it as primary reason, the denominator of coded decisions, and an interval. The NIST Engineering Statistics Handbook describes the Wilson score interval for a proportion and notes that its worth does not depend strongly on the sample size or the proportion, unlike the common textbook interval, whose lower limit can fall below zero. Because decisions from the same buyer are correlated, also report how many organizations contributed each code and a cluster-aware interval, such as a bootstrap that resamples organizations.
The baseline is the accepted decisions. A reason that appears as a contributing issue in accepted offers as often as in declined ones is a common complaint, not a cause of rejection. Report each code's share among declines and among acceptances side by side.
- State the number of organizations, interviewees and coded decisions, and the share of decisions by outcome.
- Report codes with fewer than a pre-registered minimum of organizations as observed but not quantified.
- Report stated-only reasons and evidence-anchored reasons as separate columns.
- Give intervals for every proportion; never report a ranked list without them.
What can't this study design conclude?
The sample of buyers will be a recruited, nonprobability sample, so its shares describe the organizations studied, not the market. A finding such as a given share of declines citing rights should be written as a share of coded decisions among the participating buyers in the window, with the recruitment route stated.
Stated reasons are not the same as causes. A reviewer may name price or timing when the underlying issue was confidence in rights, which is why the counterfactual question in block 5 and the evidence anchor in block 4 carry weight. Even then, the study measures reviewers' accounts of decisions, not the decisions' true drivers.
The design also cannot rank suppliers, buyers or industries, estimate how many offers are never made because suppliers expect rejection, or say what a fix is worth. Survivorship matters as well: offers that never reached formal review are invisible to it.
What must be pre-registered and disclosed before the first interview?
Register the plan before recruitment and publish it with the results. The Center for Open Science describes preregistration as a time-stamped, read-only study plan posted before data collection, with changes allowed when they are specified and explained in the final report. AAPOR's disclosure standards add the reporting floor: name who sponsored and who conducted the research, the dates of data collection, and the sample and question wording used.
At a glance
What must be pre-registered and disclosed before the first interview
- Research question and the exact definition of a review decision, an eligible interviewee and a buyer organization.
- Reference window and data collection dates.
- Recruitment routes, invitation wording and incentives, if any, and how many people were contacted, screened, eligible and interviewed.
- Decision sampling rule within each interview.
- Codebook version, coder training, the agreement statistic and its threshold.
- Saturation stopping rule and the minimum organizations per reported code.
- Interval method, clustering approach and the comparison between accepted and declined decisions.
- Confidentiality rules: no supplier names, no contract values, transcript storage and deletion dates.
- Sponsor and any commercial interest of the team, stated on the first page of the report.
- The tables that will be released, including the full codebook and per-code counts.
How can a supplier use the taxonomy on its own records before an offer goes out?
Suppliers do not need to wait for study results to use the codebook. Read each code as a question about your own offer: can you show permission to license the records, can you trace where they came from, does the privacy basis support the use, do the records carry outcomes, and do the links hold. Any code you cannot answer is a likely reason for a decline.
When a supplier licenses records through SourceX, qualification, rights review, contracting, preparation and delivery are part of the process, so several of these codes are tested before a sample reaches a buyer. Whether a supplier qualifies at all is answered first by a brief company questionnaire, with no records handed over.
Which records should follow a decision ID after the buyer says no?
Four, kept under the same ID: whether the offer was appealed or resubmitted and with what result, how much preparation effort the decline stranded, whether any was recovered, and which corrective action followed and whether a later decision confirmed it. The reason code says why an offer failed; these records say what the failure cost and whether anyone learned from it.
Buyers supply only part: the interview captures the resubmission route offered and any second outcome, while hours, recovery and fixes sit with whoever presented the offer. A supplier or intermediary can keep the same IDs and codes; its log then describes its own offers, not the market.
Recovery is a category, never an amount, so the protocol's rule against collecting contract values holds. Before preparing anything, a supplier can ask a buyer two questions: is there a route to resubmit after a decline, and who pays for requested rework? The fields:
| Field | Recorded as | Source record | What it reveals |
|---|---|---|---|
| Decision ID | One ID per reviewed offer, reused on every later record | Decision inventory from the interview | Whether reasons, costs and fixes describe the same decision |
| Appeal or resubmission | None, appealed, resubmitted after fixes, resubmitted unchanged | Buyer interview and the presenting company's log | Whether a decline is final or one step in a longer review |
| Second outcome | Accepted, declined under the same code, declined under a new code, still open | A second decision linked to the first ID | Whether the first stated reason was the real one |
| Stranded effort | Preparation and review hours spent before the decline, by side | Time records of the presenting company | Cost a decline leaves behind that no revenue line shows |
| Cost recovery | None, partial, full or unknown, plus who bore the rest | Presenting company's records, with no amounts kept | Whether declines are absorbed by suppliers, buyers or an intermediary |
| Corrective action | Code addressed, change made, owner and date | Presenting company's change log | Whether screening changed or the same offer was simply sent again |
| Verified fix | Yes only when a later offer carrying the fix clears the same code | A later decision on a corrected offer | Whether a fix worked, as opposed to being announced |
Why do growth claims need rejection cohorts, not just accepted datasets?
Because a count of accepted datasets cannot separate more success from more attempts. Group every offer by the quarter it first entered buyer review, follow each one to accepted, declined, rescued on resubmission, withdrawn or still open, and measure growth on that fixed denominator, counting stranded effort against it.
Illustrative counts, not SourceX data: accepted datasets double from 6 to 12 between two quarterly cohorts, while offers reviewed triple from 12 to 36. Acceptance per reviewed offer falls from one in two to one in three, and offers that ended declined or withdrawn rise from 6 to 24. The headline doubled while the offers behind each acceptance rose from two to three.
Three distortions recur: a first cohort clearing a backlog of easy offers (a launch spike, not a rate), a small base that turns one extra acceptance into a big percentage, and a drifting definition, such as counting conditional acceptances one quarter but not the next. The SEC's January 2020 guidance on key performance indicators in MD&A notes that such metrics vary significantly between companies and industries, so a figure compares only alongside its definition.
Dropping rejected offers and departed suppliers also biases the picture. Brown, Goetzmann, Ibbotson and Ross showed in 1992 that a mutual fund sample truncated by survivorship can produce the appearance of predictable performance; a supplier sample cut to those that reached acceptance can likewise make preparation look cheaper and buyer criteria steadier than they are.
Suppliers choosing a licensing partner can ask for this dashboard before committing staff time:
| Metric | Numerator | Denominator | What a growth headline hides without it |
|---|---|---|---|
| Acceptance per reviewed offer | Offers accepted, first time or after resubmission | Offers that entered review in the cohort quarter | Growth bought with more attempts |
| Rescue rate | Declined offers later accepted after a fix | Declined offers in the cohort | Whether declines are recoverable or terminal |
| Stranded effort per accepted dataset | Hours on offers that ended declined or withdrawn | Datasets accepted in the cohort | Preparation cost that never reaches revenue |
| Repeat-reason rate | Resubmissions declined again under the same primary code | Resubmissions in the cohort | Corrective actions announced but not effective |
| Reason mix shift | Share of declines under each primary code | Declines in the cohort, set against the prior cohort | A new rubric or rule turning many decisions at once |
| Open share at cutoff | Offers without a final decision on the reporting date | Offers that entered review in the cohort | Recent cohorts that look clean only because they are unfinished |
Where can a buyer still decline, and when do declines move together?
At almost every stage before acceptance and again at renewal, so record what each stage commits the buyer to and which decline right remains. A large buyer concentrates risk twice: through its share of accepted datasets and through its share of open decisions that could still go against the company, which only a decision record shows.
Declines are correlated when they share a cause, not only a buyer: one rubric change, one reviewing team, one rights basis common to many offers or one rule can turn several decisions in a quarter. Article 10 of the EU AI Act, which applies to high-risk AI systems, requires their training, validation and testing data to follow data governance practices suited to the intended purpose and to be relevant, sufficiently representative and, as far as possible, free of errors and complete; buyers building such systems may apply those tests to every offer at once. Count declines by buyer and by shared cause, and treat a cluster on either axis as concentration.
Mandates are no exception: micro1 says it plans to spend $1 billion over 12 months acquiring and licensing enterprise operational data for reinforcement learning environments: a stated plan, not money spent or a commitment to any company's dataset. For a supplier its value appears only as decisions moving down the matrix, which also shows where preparation effort stops being at risk.
| Stage | What the buyer has committed | Decline right still open | Effort at risk if it declines | Correlation to test |
|---|---|---|---|---|
| Announced program or mandate | A stated budget or intent for a data category | Any offer can still be declined | Preparation started on the announcement alone | How many companies are preparing for the same program |
| Sample review | Reviewer time on a sample | A full decline, often with no cost terms agreed | Sample extraction and de-identification hours | Whether one team or rubric reviews most samples |
| Conditional acceptance | A written list of fixes | A decline if the fixes miss the list | Rework on the listed fixes | Whether the same fix is being asked of several offers |
| Signed terms with an acceptance test | Scope and terms in a signed agreement | A decline under the acceptance test as the contract defines it | Full preparation of the dataset | Whether many signed deals share one buyer, test or rights basis |
| Accepted delivery | Acceptance of the delivered dataset | Only the post-acceptance remedies the contract keeps | Little for that dataset itself | Whether the next refresh depends on the same buyer program |
| Refresh or renewal | Whatever the renewal terms set | A new decision each cycle | Refresh work done before renewal is signed | Whether renewals cluster in one quarter or program |
What could mislead an investor reading these records, and what should be asked?
Four things could mislead. Suppliers with spare staff resubmit more, so a rescue rate partly measures supplier resources. Uneven time records make stranded effort a floor. A buyer's rubric change can make cohorts on either side of it incomparable. And confidentiality may leave an investor only aggregates, whose method must be explained.
Questions for any company that presents datasets to buyers, SourceX included: does every reason, appeal, cost and fix carry one decision ID; what is acceptance per reviewed offer by quarterly cohort, open offers apart; how many hours sat in declined or withdrawn offers per accepted dataset, and who bore them; which buyer, rubric or rights basis accounts for most declines; and when, if ever, did the definition of accepted change?
At SourceX, buyer datasets are sourced on request, not held in stock, and a company approves scope and terms and signs before any data is delivered, so a rejection cohort for its offers would start at the buyer's request. No SourceX decisions, cohorts or costs are reported here; SourceX shares company and commercial metrics with investors directly, on request through the investor page.
Methods note: drafted on October 26, 2026 from this page's protocol, SEC Release 33-10751, the 1992 survivorship study, EU AI Act Article 10 and micro1's announcement. All counts are illustrative; no buyer decisions, appeals, costs or cohorts were measured.
This article is for information only and is not an offer to sell or a solicitation of an offer to buy securities. Any offering by SourceX is made only to verified accredited investors through its offering documents.
Sources for this section
- The SEC's January 2020 interpretive guidance (Release 33-10751) addresses key performance indicators and metrics in MD&A and notes that such metrics vary significantly from company to company and industry to industry. Source
- Brown, Goetzmann, Ibbotson and Ross (Review of Financial Studies, 1992) show that a mutual fund sample truncated by survivorship can produce the appearance of predictable performance. Source
- Article 10 of the EU AI Act (Regulation (EU) 2024/1689), which applies to high-risk AI systems, requires their training, validation and testing data sets to be subject to data governance practices appropriate to the intended purpose and to be relevant, sufficiently representative and, to the best extent possible, free of errors and complete. Source
- micro1 announced in October 2026 a plan to spend $1 billion over 12 months acquiring and licensing enterprise operational data to build reinforcement learning environments; it is a stated plan, not money spent. Source
Frequently asked questions
Does this page report how often buyers reject datasets?
No. It publishes a study protocol, a coded reason taxonomy and a pre-registration checklist, and it reports no rates, rankings or findings. Any figures would need interviews collected under this design, with the sample, dates and intervals disclosed.
How many interviews does a buyer rejection study need?
Plan by saturation within each stratum rather than a fixed headline number. Published work on interview studies found that code saturation came earlier than meaning saturation, which needed roughly 16 to 24 interviews in one study, so plan toward the upper range and pre-register a stopping rule.
Why count review decisions instead of people?
Rejection is a property of an offer reviewed by a buyer, and one reviewer may describe several decisions with different reasons. Counting decisions gives a clear denominator, lets accepted offers serve as a baseline, and allows intervals that account for decisions clustered within the same buyer.
Can a single supplier run this protocol on its own declines?
A supplier can borrow the taxonomy and the evidence-anchor question, but it cannot produce a market-level finding from its own declines. Its sample is one supplier's offers, so the result describes that supplier's experience and should be reported that way.
How do you keep buyer and supplier information confidential?
Never record supplier names, contract values or commercial terms, since none are needed to code reasons. Store transcripts under a stated retention period, report counts only above a minimum number of organizations, and let interviewees review the codes assigned to their decisions.
Sources
- The Data Provenance Initiative audited more than 1,800 text datasets and reported license omission above 70 percent and error rates above 50 percent on widely used hosting sites. Source
- Datasheets for Datasets proposes that every dataset be accompanied by a datasheet documenting its motivation, composition, collection process and recommended uses. Source
- AAPOR's establishment-survey definitions treat an establishment as a single unit or location within a business or organization. Source
- COREQ is a 32-item checklist for reporting interview and focus group studies, grouped into research team and reflexivity, study design, and data analysis and reporting. Source
- In a 25-interview study, code saturation was reached at nine interviews while meaning saturation needed 16 to 24 interviews. Source
- The NIST Engineering Statistics Handbook describes the Wilson interval for a proportion, recommended for virtually all combinations of n and p, and notes the common textbook interval can give a negative lower limit. Source
- Preregistration posts a time-stamped, read-only study plan before data collection; changes are acceptable when specified and explained in the final report. Source
- AAPOR's disclosure standards require naming the sponsor and the party that conducted the research and disclosing the dates of data collection. Source
See if your company qualifies
A short company assessment. No data uploads are needed.