Skip to content

AI data market

AI data buyer interview research: three questions market reports can't answer

By SourceX Editorial · Reviewed by Noah Loul ·

11 sections

Short answer

AI data buyer interview research shows three things market reports cannot: who owns the budget, what test the data must pass and which alternative lost. Interview independent budget owners, place each organization on a six-rung ladder from public signal to renewal, confirm every rung above stated interest with a dated document, and report nonresponse, which in an illustrative case moves a category estimate more than threefold.

Checklist

AI data buyer interview

This page covers AI data buyer interview research: three questions market reports can't answer.Illustrative

Key takeaways

  • Market reports describe the environment; procurement interviews show the decision: the budget owner, the acceptance test and the alternative that lost.
  • Place each buyer organization on a six-rung ladder and count only rungs a dated document supports; an interview alone establishes stated interest.
  • Count distinct budget owners, not enthusiastic people: by Reddit's fiscal 2025 Form 10-K, substantially all of its licensing contract value rests with just two partners.
  • Interviews can size a category as a floor with a range: frame size times adjusted paying share times median committed spend, with exclusions listed.
  • Nonresponse is the assumption to watch; in an illustrative case it moves the paying share between 6 and 20 percent of the frame.

What can AI data buyer interview research show that a market report cannot?

Three things: who controls the budget for a data category, what test a dataset must pass before anyone pays, and which alternative the buyer weighed and rejected. A market report can estimate aggregate spending and list announcements, but it cannot see inside a single procurement decision, and the decision is where AI data demand turns into revenue or quietly disappears.

The gap is especially wide in AI data, where purchases are lumpy, contracts are confidential and a few buyers can hold most of the money. Reddit's Form 10-K for fiscal 2025 states that, to date, "substantially all of the contract value associated with our licensing revenue is derived from two of our partners." A category estimate assembled from announcements would not show that two organizations stand behind a public company's licensing line; the filing shows it, and so would interviews with the people who sign.

The decision this research supports is narrow and testable: does a data category have paying, repeatable demand from enough independent buyers to carry a company built on it? This guide gives an investor the tools to answer it: a six-rung demand ladder, three interview anchors, a procurement evidence matrix, a sampling standard and a sizing sheet that shows how far nonresponse can move a category estimate. It reports no interview results, and every worked number is labeled illustrative.

The demand-signal ladder: which rung has a buyer actually reached?

Place every buyer organization on the highest rung it has reached, and count only rungs you can document. The six rungs separate signals that cost the buyer nothing from commitments that cost money, and each rung demands stronger evidence than the one below it.

Evidence types differ in how high they can reach. A market report establishes rung 1 and can only model rung 5 in aggregate. A public filing can establish rung 5 for the company that files it, as Reddit's S-1 did when it put the aggregate transaction price of the data licensing arrangements it entered in January 2024 at $203.0 million, at least $66.4 million of it expected as 2024 revenue. An interview on its own establishes rung 2 and reports everything above it as claims. An interview backed by documents the buyer is free to share can establish every rung up to 6.

Two rules keep the ladder honest. Date each rung when it was reached, not when someone mentioned it, so cohorts compare like with like. And record falls: an evaluation that ended without a purchase sends the organization back to the rung its evidence still supports, with the reason coded.

The demand-signal ladder: which rung has a buyer actually reached?
RungWhat has happenedEvidence that establishes itWhat it still does not prove
1. Public signalA model release, benchmark, funding round or spending announcement points at a capabilityThe announcement itselfThat any organization will pay for data
2. Stated interestA named person says their organization would want data of this kindInterview notes with role, organization and dateThat a budget exists or a request will follow
3. Written requestThe organization sends a dated request naming data type, volume, format and rightsThe request documentThat a sample will be evaluated or bought
4. Funded evaluationA sample is tested against a written acceptance standard, with staff time assignedEvaluation license or sample agreement, plus the test planThat the dataset will pass or be purchased
5. Paid commitmentA license or order is signed with an amount, a term and acceptance termsSigned license, order form or purchase orderThat the buyer will buy again
6. Renewal or expansionAfter acceptance, the buyer renews, widens scope or buys a second datasetRenewal, amendment or second orderThat the need survives the next model generation
A column chart of six demand rungs against four kinds of evidence: market reports establish only rung 1, a public filing reaches rung 5 for the filer, an interview alone establishes rung 2, and an interview backed by documents reaches rung 6.
How high each kind of evidence can reach on the demand-signal ladder. Illustrative framework, not measured data.

Budget, acceptance standard, rejected alternative: the three interview anchors

Build every interview around three anchors, because each answers a question no aggregate source can: whether money exists, whether the data can pass, and whether the buyer has a cheaper way out. Ask about specific recent decisions rather than opinions about the market, and ask for dates.

At a glance

Budget, acceptance standard, rejected alternative: the three interview anchors

Budget comes first. Find the line item that would pay, the person who approves it and the fiscal period it covers. A researcher who admires a dataset but has no line to charge it to sits on rung 2, however enthusiastic. Interest from three teams that draw on one central budget is one buyer, not three.

The acceptance standard is the test the data must pass before anyone signs, and it should be specific enough to run. Anthropic's engineering guide to agent evaluations treats a task as one test with its inputs and success criteria spelled out, and recommends several trials per task because model outputs vary from run to run. A buyer who can describe an acceptance test at that level has done the work that comes before a purchase; one who says the data must be high quality has not.

Buyers have good reason to prefer their own held-out tests to public scores. When researchers wrote GSM1k, a fresh set of grade-school math problems built to mirror the public GSM8k benchmark, leading models scored up to 8 percent lower on it, according to the paper's current version. A private test is how a buyer checks that data helps on its own task, so a written one is a strong demand signal in itself.

The rejected alternative is the anchor investors skip most often. Every data purchase competes with in-house collection, synthetic generation, public data, another supplier and simply waiting. Buyers sometimes publish their exclusions: OpenAI's data partnerships page turns away sensitive or personal information and data a third party owns, even as it looks for large-scale data the public cannot already get easily online. In an interview, ask what else was weighed and why each option lost.

Questions that work for each anchor, asked about one recent decision:

  • Budget: Which line would pay for this, who approves it, and is money allocated for the current fiscal period?
  • Budget: Do other teams draw on the same budget, and who signs if two teams want the same data?
  • Acceptance: What test did the last dataset you accepted or rejected face, what threshold applied and who signed the result?
  • Acceptance: Was the test written before the sample arrived, or after?
  • Alternatives: What else did you consider, including building it yourselves, synthetic data, public data, another supplier or waiting, and why did each lose?
  • Counterfactual: If this data were unavailable for a year, what would you do instead, and how much time would that cost?

The procurement evidence matrix: report, interview or document?

Use the matrix to decide where each answer must come from before it counts. A market report can fill the second column; it cannot fill the third, and an interview answer without the document in the last column stays a claim.

Read across a row to see the upgrade path. The interview tells you which document should exist; the buyer then shares it, describes it in enough detail to be checked later, or cannot. That third outcome is information too: a buyer who cannot point to a budget line, a written test or a comparison memo is usually below rung 4.

The procurement evidence matrix: report, interview or document?
Question an investor needs answeredWhat a market report can showWhat a procurement interview can showDocument that settles it
Is there money for this category?Aggregate spend estimates and announced AI budgetsWho owns the budget line, its size band and the fiscal periodApproved budget line, purchase order or signed order form
What must the data pass?General quality themesThe acceptance test, its threshold and who signs acceptanceWritten acceptance criteria or evaluation plan
What did the buyer choose instead?Vendor lists and estimated market sharesThe alternatives weighed and why each lostEvaluation memo or vendor comparison
How far has demand moved?Announcements, funding and benchmark newsThe highest rung reached and the date it was reachedDated request, evaluation license or signed license
How many independent buyers are there?Concentration estimates, where any existWhether budgets are central, shared or duplicated across teamsCount of distinct signing organizations
Will it repeat?Growth forecastsRenewal plans and what would trigger a second purchaseRenewal, amendment or second order

Who should be interviewed, and how is the sample kept honest?

Interview independent budget owners: people who can approve or veto spending on the data category at organizations that train, fine-tune or evaluate models, and who were not introduced by the company under diligence. Code everyone else, such as engineers who use the data and managers who influence the choice, as influencers in a separate stratum.

Write the frame before the first call: the list of organizations, the rule that put each one there and the channels used to reach them. Company-supplied references belong in their own stratum and never in the headline share, because a company picks them for satisfaction, not representativeness.

Record what happened to every organization in the frame, not only the ones that talked. AAPOR's Standard Definitions sort every sampled case into a final disposition (complete, eligible but not interviewed, unknown eligibility or not eligible), define response, cooperation, refusal and contact rates from those counts, and ask researchers to name the formula they used and publish the disposition table. AAPOR applied the same structure to surveys of organizations in its Standard Definitions for Establishment Surveys, published in December 2025, which suits buyer research, where the unit is an organization rather than a person.

Separate requests, evaluations and paid commitments by dated cohort. Group organizations by the quarter in which they first reached rung 3, follow each cohort forward and report how many reached rungs 4, 5 and 6 by each later quarter, keeping organizations still in evaluation in the count as open rather than removing them. Pooling cohorts lets an old wave of purchases flatter a new wave of interest.

Can procurement interviews size an AI data category?

Yes, as a documented floor with a range, not as a data licensing TAM. AI data market customer research can size a category bottom-up when every input comes from the interview sample or from documents, the frame is written down, exclusions are visible and nonresponse is modeled rather than ignored.

The arithmetic is short. One year of category spend equals the number of organizations in the frame, times the adjusted share of them at rung 5 or above for the category, times the median annual committed spend per paying organization. The adjusted share combines what responders showed with an explicit assumption about everyone else: the response rate times the observed share, plus the nonresponse rate times the observed share times k, the assumed paying rate of non-responders relative to responders.

Two choices in that formula carry most of the risk. Use median committed spend, not the mean, because one large buyer can dominate a small sample, the pattern Reddit's concentration disclosure describes. And never set k to 1 by default: organizations with an active purchase are more willing to talk about it, so assuming non-responders buy at the same rate inflates the floor.

List exclusions beside the result: budgets for in-house collection and annotation staff, synthetic data generation, compute sold together with data, and one-time settlements. Each may be real spending, but none is addressable by a supplier licensing records unless the interviews show the buyer would switch.

Can procurement interviews size an AI data category?
InputDefinitionEvidence that sets itCommon inflation
F: frame sizeOrganizations that could buy the category, under a written inclusion ruleThe frame list and its ruleCounting every AI company instead of those with a task that needs the data
RR: response rateCompleted interviews over eligible organizations, with the formula namedA disposition table in the AAPOR formatDropping unknown-eligibility cases from the denominator
p: observed paying shareShare of completed interviews at rung 5 or above for this categorySigned licenses or orders dated inside the cohort windowCounting funded evaluations or verbal plans as purchases
k: non-responder ratioAssumed paying rate of non-responders relative to responders, from 0 to 1A stated scenario range, never a default of 1Treating silence as if it meant yes
s: committed spendMedian annual committed spend per paying organization in the categoryAmounts in shared documents or ranges confirmed by budget ownersUsing the mean, or totals that bundle compute or services
r: repeat shareShare of paying organizations that have reached rung 6Renewals, amendments or second ordersTreating a one-time purchase as recurring spend

How much can nonresponse move a category estimate?

Enough to change the answer. In the illustrative case below, the same interviews support a paying share anywhere from 6 to 20 percent of the frame, a range of more than three times, before spend per buyer is even estimated. Only the assumption about organizations that did not respond changes.

Illustrative case: an investor contacts every organization in a frame, 30 percent complete an interview, and 20 percent of completed interviews show a document-confirmed paid commitment in the category. If non-responders buy at the same rate, the paying share is 20 percent. If they buy at half the rate, it is 13 percent. If none of them buy, it is 6 percent. These figures are arithmetic on assumed inputs, not measurements of any market.

Report the estimate as a range across values of k and say which value you believe and why, for example by comparing early and late responders, or by checking non-responders against public signals such as model releases in the category. A single number without a range hides the assumption that matters most.

How much can nonresponse move a category estimate?
Assumed paying rate of non-responders (k)Adjusted paying share of the frameChange against assuming k = 1
Same as responders (k = 1)20.0 percentNo change
Three quarters (k = 0.75)16.5 percent17.5 percent lower
Half (k = 0.5)13.0 percent35 percent lower
One quarter (k = 0.25)9.5 percent52.5 percent lower
None buy (k = 0)6.0 percent70 percent lower

Why spending headlines and data stocks overstate licensing demand

Because they measure something next to the purchase rather than the purchase. AI capital spending, the size of the world's data and venture funding describe the environment a data deal happens in, and announcements and capability progress can climb for years without a matching line in anyone's data budget. That is the strongest alternative explanation for any bullish reading of the post-training data market, and interviews are how to test it.

Spending figures describe compute paybacks. In September 2023, Sequoia's David Cahn estimated that each year of GPU capital spending at the then-current pace needed about $200 billion of lifetime revenue to pay it back. That is a question about whether AI applications earn enough to justify the infrastructure; it says nothing about what share, if any, reaches licensed data.

Data stocks describe supply, not budgets. Epoch AI sizes the effective stock of quality public human text, after adjusting for repetition, at about 300 trillion tokens, with a 90 percent confidence interval of 100 trillion to 1,000 trillion, and projects that on current trends it would be fully used between 2026 and 2032, earlier if models are heavily overtrained. A shortage creates demand only where a buyer has a task, a test and a budget, which is exactly what the interview anchors check.

Capability evaluations describe direction, not orders. OpenAI's GDPval covers 1,320 tasks across 44 occupations in the nine industries that contribute most to U.S. GDP, written and vetted by professionals averaging 14 years of experience, and OpenAI notes that it is one-shot and does not capture work that needs multiple drafts or accumulated context. A benchmark like this shows where developers want to measure progress, a rung 1 signal; whether anyone pays for expert task data is a rung 5 question.

The test for every such signal is the same: name the buyer organization, the rung it reached and the document that proves it. If the chain stops at the announcement, the signal is context for a thesis, not evidence of demand.

Where interview evidence misleads, and what would weaken the thesis

Interview evidence fails in predictable ways, and a thesis built on it is only as strong as the checks against those failures. Name them before reading any results, including your own; each one below could turn a promising category into a thin one.

  • Concentration hides inside counts. Reddit's 2024 IPO prospectus said substantially all of its licensing contract value came from one partner, and its Form 10-K for fiscal 2025 says two. Ten enthusiastic interviews can rest on one or two budgets, so report the number of distinct signing organizations next to every share.
  • Buyers answer strategically. A buyer may talk demand up to attract suppliers or down to anchor price, so weight documents above statements and ask the same question at two different rungs.
  • Selection flatters. Organizations with active purchases agree to talk more readily, and company-supplied references are chosen for satisfaction; both inflate the paying share unless strata and nonresponse are reported.
  • Capability can erase the need. A model generation that closes the gap a dataset was bought to fill turns a renewal into churn, so ask every paying buyer what would end the need.
  • Substitutes move faster than studies. If synthetic or public data starts passing a buyer's acceptance test, the rejected-alternative anchor flips; re-interview each cohort at least once a year.
  • Interviews have legal and confidentiality limits. Ask for ranges and documents the buyer is free to share, never for a counterparty's confidential terms, and have counsel set the rules when buyers or targets are public companies.

Questions to ask any AI data company about its buyer evidence, SourceX included

Ask for the ladder, the matrix and the sizing inputs from the company's own records, using your definitions rather than theirs. SourceX should answer them on the same terms as anyone else.

Read SourceX here as a worked application, not as evidence for the method. Its business is converting chosen business records into licensed AI datasets, and it does each step itself, from sourcing and qualifying supply through rights review and contracting to preparing, delivering and paying for the data. For AI buyers nothing comes out of inventory: each dataset is sourced from U.S. businesses once a buyer asks, with buyers stating scope, volume, language, format, timing and licensing needs through a request form, so a written request, rung 3, is where its sourcing starts. Data moves to SourceX only after the supplying company has signed off on the scope and terms and put its signature on the agreement.

SourceX has not published interview, request, evaluation or commitment counts, and investors receive its company and commercial metrics directly. Its Rule 506(c) raise accepts accredited investors only; materials are available on request through the investor page, and any of the questions below is a fair place to start.

Checklist · 6 items

  1. How many distinct organizations sit on each rung today, and on what date did each reach it?
  2. Which documents support rungs 3 to 6, and can a sample be reviewed under a confidentiality agreement?
  3. What acceptance tests have buyers written, and how many datasets passed or failed them?
  4. What did lost or stalled buyers choose instead, and how was each reason coded?
  5. What share of committed value comes from the largest one and two buyers?
  6. If you sized a category from interviews, which frame, response rate and non-responder assumption did you use?

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

Methods note: how the ladder, matrix and sizing sheet were built

The six-rung ladder, the evidence matrix, the three interview anchors and the interview-to-category sizing sheet are this article's editorial frameworks, prepared on 2026-10-26. The public evidence comes from Reddit's S-1, its 2024 IPO prospectus and its Form 10-K for fiscal 2025, OpenAI's data partnerships and GDPval pages, Anthropic's guide to agent evaluations, the GSM1k paper on arXiv, AAPOR's Standard Definitions and its establishment-survey edition, Epoch AI's data-stock study and Sequoia's 2023 essay on GPU payback. The nonresponse case uses assumed inputs and describes no real market, company or buyer.

Not measured: the number of organizations on any rung, response rates, paying shares, committed spend or repeat rates for any category, at SourceX or anywhere else. No interviews were conducted for this article, and none of these quantities can be estimated from outside without records that have not been published.

Frequently asked questions

How many interviews does AI data buyer research need?

Enough to cover every stratum in the frame, with the stopping rule written before the first call. The count of interviews matters less than the count of distinct budget-owning organizations: ten interviews inside two organizations are two observations of budget. Report the frame, completed interviews, the response rate and the number of organizations on each rung.

Are expert network calls the same as procurement interviews?

Usually not. Expert calls often reach former employees or practitioners who can describe a market but do not control a current budget, so their answers count as stated interest at most. A procurement interview targets the person who approves or vetoes the spend and asks for dated documents.

What is the difference between an AI data request and a purchase?

A request is a dated, written description of the data an organization wants, and it costs the buyer little. A purchase is a signed license or order with an amount, a term and acceptance terms. Between them sits the funded evaluation, where the buyer spends staff time testing a sample against a written standard.

Is a market report still useful for AI data diligence?

Yes, as context: which capabilities developers are pursuing, how much capital is moving and which categories draw attention. Use it to choose whom to interview and which categories to test, then require interview and document evidence before treating any figure as demand.

Does SourceX publish buyer interview or demand results?

No. Interview, request and commitment results are unpublished, and company and commercial metrics are shared with prospective investors directly rather than publicly. Materials are available on request at sourcex.si/investors, and SourceX makes any offering only to verified accredited investors.

Sources

  • Reddit's Form 10-K for fiscal 2025 states that, to date, substantially all of the contract value associated with its licensing revenue is derived from two of its partners, and that those agreements may not be renewed or may be renewed on less favorable terms. Source
  • Reddit's 2024 IPO prospectus stated that substantially all of the contract value associated with its licensing revenue was derived from one of its partners. Source
  • Reddit's S-1 disclosed data licensing arrangements entered in January 2024 with an aggregate transaction price of $203.0 million, of which at least $66.4 million was expected to be recognized as revenue in 2024. Source
  • The GSM1k paper (arXiv 2405.00332, current version) reports accuracy drops of up to 8 percent when leading open and closed models are evaluated on fresh GSM1k problems instead of the public GSM8k benchmark. Source
  • OpenAI's GDPval spans 1,320 tasks across 44 occupations in the top nine industries by U.S. GDP, created and vetted by professionals averaging 14 years of experience; OpenAI notes it is one-shot and does not capture multi-draft work or accumulated context. Source
  • OpenAI's data partnerships page asks for large-scale datasets that are 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
  • Anthropic's engineering guide to agent evaluations defines a task as a single test with defined inputs and success criteria, and recommends running multiple trials because model outputs vary between runs. Source
  • AAPOR's Standard Definitions (10th edition, 2023) classify each sampled case into final dispositions (complete, eligible nonresponse, unknown eligibility, not eligible), define response, cooperation, refusal and contact rates, and ask researchers to state the formula used and publish the disposition table. Source
  • AAPOR published Standard Definitions for Establishment Surveys in December 2025, applying the same disposition framework to surveys of businesses and other organizations. Source
  • Epoch AI estimates the effective stock of quality, repetition-adjusted public human text at about 300 trillion tokens (90 percent confidence interval 100 trillion to 1,000 trillion), fully used between 2026 and 2032 on current trends, earlier with heavy overtraining. Source
  • In AI's $200B Question (September 20, 2023), Sequoia's David Cahn estimated that each year of GPU capital spending at the then-current pace implied about $200 billion of lifetime revenue needed to pay back the investment. 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