AI data market
AI data buyer demand tracker methodology: how to measure requests
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
To research AI data buyer demand without unsupported claims, count only dated written requests, code each against a fixed taxonomy, state which channels and dates the sample covers, and publish aggregated shares with intervals after suppressing cells that could identify a requester. Register the method before coding starts. This page is that protocol; it reports no results and contains no demand figures.
Value drivers
AI data buyer demand tracker
Key takeaways
- This is a measurement protocol only; it reports no results and no demand figures.
- Count written, scoped requests; conversations, news and public statements are context, not demand.
- Code every request on the same ten dimensions, with explicit unclear values and a versioned codebook.
- Publish shares of the requests in the sample, with denominators and intervals, never shares of the market.
- Suppress small or dominated cells by distinct requester, and never publish names or verbatim request text.
- Pre-register channels, rules, thresholds and planned tables before the first request is coded.
How can you track AI data buyer demand without publishing unsupported claims?
Treat demand as something you measure, not something you assert. Define the unit you count (a dated, written request for data), code every request against a fixed taxonomy, state which requests your sample can and cannot see, and publish only aggregated shares with intervals, after suppressing any cell small enough to identify a requester. Write all of that down and register it before the first request is coded.
What follows is a measurement design rather than a finished study. It sets out the question, population, sampling, variables, coding rules, reporting rules, uncertainty and limits for an AI data buyer demand tracker. It reports no results, contains no demand figures, and should not be read as evidence that any particular kind of data is in demand.
The protocol suits a data marketplace, a broker, a company fielding inbound requests, or a journalist assembling requests from several sources. The common failure it is designed to prevent is the shopping list: a published inventory of what buyers supposedly want, compiled from conversations, never counted, and impossible to check.
Which signals count as demand, and which do not?
Only written requests with a stated scope enter the count. Everything else is context. A conversation where someone mentions interest in a data type, a press report about a model developer, or a public statement about data strategy may be worth noting, but none of them is a request, and mixing them into the count is how trackers end up measuring rumor.
Keep each signal type in its own column of the coding sheet so the count stays clean and the context is not lost.
| Signal | Evidence that it exists | Counted as | Never report it as |
|---|---|---|---|
| Written request with scope | Dated form submission, email or request document naming data type and use | One request, coded on every taxonomy dimension | A purchase or a commitment to buy |
| Formal request for proposal | Issued document with requirements and a response date | One request, flagged as formal | Evidence that a contract followed |
| Repeat or revised request | Same requester, same need, new date or changed scope | Linked to the original; counted once per wave | Additional independent demand |
| Signed agreement | Executed contract | Outcome of a request, coded separately | A demand count on its own |
| Verbal or meeting mention | Notes from a call or event | Context only, excluded from counts | A request |
| Public statement or news item | Published article or announcement | Context only, excluded from counts | A request from that organization |
The request taxonomy
Every counted request is coded on the same dimensions, using the definitions below. Categories within each dimension are mutually exclusive, with an explicit unclear value so coders never force a fit. The taxonomy is versioned: when a category is added or split, the change log records the date, and later waves report under both versions until the overlap is understood.
| Dimension | Categories (starting set) | Coding rule | If the request is silent |
|---|---|---|---|
| Data modality | text, documents, structured records, event logs, audio, image, video, code, mixed | Code what the requester asks to receive, not what the source system is called | unclear |
| Record domain | Industry and business function of the source records | Use one industry list for every wave | unclear |
| Intended use | pretraining, fine-tuning, evaluation, retrieval, analytics, other | Only the use stated in writing | not_stated |
| Volume band | Ordered bands of record counts fixed in the protocol | Lowest band that contains the stated minimum | not_stated |
| History depth | Ordered bands of years of history | Band of the stated minimum span | not_stated |
| Language | Language of records requested | Code each language named | not_stated |
| Geography of source | Country or region of the businesses or people in the data | Code the requirement, not the requester's location | not_stated |
| Licensing terms sought | exclusive, non-exclusive, evaluation-only, unspecified | Only terms stated in the request | unspecified |
| Requester type | model developer, enterprise team, research group, intermediary, other | Based on the requester's own description | unclear |
| Verification status | verified_identity, unverified | Verified only when the organization and contact were confirmed | unverified |
Who is in the sample, and what is the denominator?
The population is the set of requests your channels could receive during the wave, not the market. A tracker built from one company's inbound form sees only requesters who found that form; a tracker pooled from several intermediaries sees only their combined reach. Name every channel, its collection dates and what it structurally misses, and report results as shares of the requests in the sample, never as shares of the market.
At a glance
Who is in the sample, and what is the denominator
Survey research offers a ready disclosure standard. AAPOR's disclosure standards ask researchers to state the dates of data collection, the sample size for each frame or list used, how the frame was built and which parts of the target population it does not cover. They also say that reports from non-probability samples should not present measures of precision unless the model behind them is described and validated, and should otherwise state that precision measures are not provided. A request tracker is a non-probability sample by construction, so it inherits that caution.
- Unit of count: decide in advance whether a share is of distinct requests or distinct requesters, and report which.
- Deduplication: link the same request arriving through several channels before counting, using organization and scope, and report how many links were made.
- Inclusion: a request enters the wave on the date it was received, inside the wave's stated window.
- Exclusion: spam, solicitations from data sellers and requests with no data type are excluded and counted in an exclusion log.
- Denominator: every share is reported with its denominator, the count of in-scope requests in the same wave.
How do you keep coding consistent across coders and waves?
Write a codebook with a definition, an inclusion example and an exclusion example for every category, then have two people code a random subset of each wave independently. Compare their codes dimension by dimension, resolve disagreements by a named adjudicator, and revise the codebook where disagreements cluster. Report the agreement check, the size of the double-coded subset and the codebook version with each wave.
Document the dataset as well as the findings. Datasheets for Datasets, a widely used documentation framework, asks whether a dataset is a sample of a larger set and, if so, what the sampling strategy was and whether the sample is representative. Answering those questions for the request log itself makes the tracker's limits visible to anyone who later reuses the coded data.
The anonymized reporting method
Requests are commercially sensitive. A detailed request can reveal a requester's model plans, and a published cell with one requester in it is that requester's plan. The reporting method therefore publishes aggregates only, under rules fixed before the data is seen. NIST's guidance on de-identifying government datasets frames the same goal: limiting disclosure risk to individuals and establishments while still allowing meaningful statistical analysis.
At a glance
The anonymized reporting method
The rules below are the reporting method. The specific thresholds are set in the pre-registration and do not change after the data is in.
- No requester names, logos, contact details or verbatim request text, ever, including in examples.
- Minimum cell size: a cell is published only when it contains at least a pre-registered number of distinct requesters, not merely requests.
- Dominance rule: a cell is suppressed when a single requester accounts for more than a pre-registered share of its requests.
- Complementary suppression: when a cell is suppressed, suppress enough neighboring cells that it cannot be recovered by subtraction from totals.
- Coarsening before suppression: merge sparse categories into broader ones, such as a whole industry instead of a sub-sector, and say so in the table notes.
- No combinations that narrow to a requester: limit cross-tabulations to two dimensions unless every resulting cell passes the rules.
- Verification flag on every table: shares based on unverified requests are reported separately or labeled as such.
How should a share of requests carry its uncertainty?
Report every share with an interval and its denominator, and choose an interval method that behaves at small counts. NIST's Dataplot reference for proportion confidence intervals offers the Wilson, adjusted Wald and Jeffreys methods, uses Wilson by default, and notes the recommendation of Wilson or Jeffreys for small samples. The simple normal approximation tends to understate uncertainty when counts are small or shares are near zero or one.
Be clear about what the interval means. In a non-probability sample, an interval describes variation within the requests you observed under the stated design; it does not make the sample representative of all buyers. State that in the table note, following the AAPOR caution above, and compare waves only when the channels, taxonomy version and coding rules are unchanged or the changes are reported.
Which conclusions can this design not support?
The tracker supports statements of the form: among in-scope written requests received through these channels in this period, this share asked for this category, with this interval. It does not support statements about market size, prices, how much any buyer will spend, which organizations are buying, whether a request led to a purchase, or what demand will be next year.
At a glance
Which conclusions can this design not support
The limits matter beyond method. The FTC's policy statement on advertising substantiation says a firm must possess and rely on a reasonable basis for objective claims before making them. A company that uses demand figures to market data services is making objective claims; a pre-registered protocol and reproducible tables are a large part of that basis, and claims beyond what the design supports have none.
- Not supported: total demand for any data type, or growth of the market.
- Not supported: which named companies want which data.
- Not supported: that coded requests were funded, approved or fulfilled.
- Supported with care: change in a category's share between waves, when the design is unchanged and intervals are reported.
- Supported: the mix of stated uses, modalities and terms among the requests in the sample.
Pre-registration checklist before the first wave
Register the protocol before coding starts. The Center for Open Science describes preregistration as specifying a study's plan in advance, before data are gathered or analyzed, so readers can tell which choices were made before the data were seen. For a request tracker, registering the plan prevents the most tempting failure: choosing categories or thresholds after seeing which ones produce a striking headline.
Checklist · 9 items
- Research question in one sentence, naming the unit of count.
- Channels in the sample, wave dates and what each channel structurally misses.
- Inclusion, exclusion and deduplication rules.
- Taxonomy version, codebook and the size of the double-coded subset.
- Reporting rules: minimum cell size, dominance share, suppression and coarsening order.
- Interval method and how verified and unverified requests are separated.
- Planned tables, written out with empty cells, and the comparisons between waves.
- Statements of what the design cannot support, copied into every publication.
- Correction policy: how errors found after publication are fixed and logged.
0 of 9 ticked. Tick each item as you go.
Reading demand claims about the kind of records you hold
Business owners are the people most often shown shopping lists of data that buyers supposedly want. Ask for the protocol behind any such list: the unit of count, the channels, the dates, the denominator and the suppression rules. A list without them is an opinion, however confidently presented.
SourceX's own model runs the other way around: it holds no stock and goes looking for records at U.S. businesses only when a buyer submits a request, and that request form captures scope, volume, language, format, timing and licensing needs, the same dimensions this taxonomy codes. A company curious whether its own records would qualify can begin with a short assessment that involves no upload.
How strong is each AI buyer data demand signal, from announcement to payment?
Rank each signal by what the buyer has committed. Announcements and new benchmarks sit at the bottom, the written scoped request this protocol counts sits in the middle, and only an accepted, paid delivery sits at the top. Code the higher rungs as outcomes attached to the original request ID, so any statement about AI buyer data demand signals names the rung it rests on.
The strongest alternative explanation for a rising signal is that announcements and capability progress never turn into spending. OpenAI's Data Partnerships page asks for large datasets that are not easily accessible online, particularly data that expresses human intention; a call for data says what a developer wants, not how many datasets it acquired or on what terms. OpenAI describes GDPval as measuring model performance on economically valuable tasks across 44 occupations, built from the work of professionals averaging 14 years of experience. That shows what a lab measures, not what it budgets to buy, and a model that already handles a task may need less data for it.
Labor statistics show how strict a definition can be. The Bureau of Labor Statistics counts a job opening only when a specific position exists, work could start within 30 days and the employer is recruiting outside the establishment, and it drops positions already filled by a hire who has not yet started. Hold each rung to the same standard: it counts only when its evidence exists. A business owner told that buyers want records like theirs can ask which rung that claim stands on.
| Rung | Evidence that places a request here | What it supports | What it cannot support |
|---|---|---|---|
| 1. Public signal | A dated announcement, call for data or benchmark release | Which data types a developer says it values | That a budget exists or that it buys from outside suppliers |
| 2. Written scoped request | The counted unit on this page: a dated request naming data type, use and volume | Stated interest among requests in the sample | Approval, budget or fit |
| 3. Budget owner confirmed | The requester names, in writing, the person who approves spend | The request has a sponsor inside the buyer | That spend has been approved |
| 4. Authorized evaluation | Executed evaluation or confidentiality terms and a sample or specification test scheduled | The buyer committed staff time to test fit | That a purchase will follow |
| 5. Commitment | A signed license or purchase order for a defined scope | Committed demand for one defined dataset | That the data will pass acceptance |
| 6. Accepted delivery | Written acceptance against the agreed criteria | Demand that met the buyer's bar | Repeat buying or price levels |
| 7. Paid | Payment received against an invoice | Realized buying for that scope | Market size or next year's demand |
What buyer activation metric is worth tracking, if introductions do not count?
Treat a buyer account as activated when it reaches rung 4, executed evaluation terms with a named budget owner, within a pre-registered window after its first written request. Introductions, calls and meetings never move an account up a rung.
Group accounts into cohorts by the month of their first in-scope request, count the organization rather than the contact, and keep intermediaries apart using the taxonomy's requester type. Every rate takes the whole cohort as its denominator, including accounts that went quiet, and cohorts are compared only at the same age, so a two-month-old cohort never sits beside a year-old one. Engagement, committed demand and realized buying stay three separate rates, never one blended conversion figure.
Suppliers can put the same question to any intermediary that brings them a request: of requests like this one, how many reached authorized evaluation, and how long did it take? The dashboard below is a template only and holds no measured values.
| Dashboard column | What qualifies an account | Denominator | Read at (example) |
|---|---|---|---|
| Buyer accounts | Distinct organizations whose first in-scope written request fell in the cohort month | None: this is the base | Cohort entry |
| Engaged | Budget owner confirmed in writing (rung 3) | All accounts in the cohort | Day 60 |
| Activated | Executed evaluation terms and a sample or test scheduled (rung 4) | All accounts in the cohort | Day 90 |
| Committed | Signed license or purchase order for a defined scope (rung 5) | All accounts in the cohort | Day 180 |
| Realized | Delivery accepted in writing and payment received (rungs 6 and 7) | All accounts in the cohort | Day 365 |
| Days to activation | Median days from first request to executed evaluation terms, with still-open accounts shown separately | Activated accounts | Day 90 and day 180 |
| Largest-account share | Share of committed scope held by the single largest account | Committed scope in the cohort | Day 180 |
What must an AI data demand index methodology hold fixed before it reports growth?
It must hold the channels, period lengths and stage definitions fixed, publish requests and accepted deliveries as separate series, and refuse percentage growth from a tiny base. Without those controls a rising line can come from a new channel, a seasonal peak or a base of three requesters.
Public indices show the disclosure to expect. Indeed Hiring Lab states that its Job Postings Index is a seven-day trailing average of seasonally adjusted postings, set to 100 on February 1, 2020: the base, the smoothing and the seasonal treatment are stated, and the series covers postings on one job site rather than every opening in the economy. An index built from one company's requests is in the same position, so call it an index of requests through named channels until it has a sampling frame for the market.
Definitions will change, so make each change visible. The SEC's 2020 guidance on metrics in MD&A, Release 33-10751, says a company that changes how it calculates a metric should consider disclosing the change, its reasons and its effect on current and previously reported amounts. A demand index meets that bar by recomputing earlier periods whenever its channels or stage definitions move.
An owner shown a chart of rising demand for records like theirs can use the right-hand column below as a checklist: a figure published without those items cannot be checked.
| Index rule | What it prevents | Publish with every figure |
|---|---|---|
| Matched channels: compare only channels live in both periods, and link the series when one is added | A new form or partner read as new demand | Channel list with start dates |
| Same-span comparison: set a period against the same span a year earlier, or adjust for seasons | A budget-cycle or seasonal peak read as growth | Period dates and adjustment method |
| Minimum base: no percentage change below a pre-registered count of distinct requesters | Growth rates from a near-zero start | Raw counts for both periods |
| Two series: requests and accepted deliveries, each by request cohort | Rising interest presented as rising buying | Both series on the same chart |
| Provisional tail: mark recent periods provisional until the acceptance window closes | Young cohorts read as falling conversion | Revision date and size of each revision |
| Restated history: recompute earlier periods after any definition change | A break in method read as a break in demand | Old and new values side by side |
Where could these demand measures mislead an investor, and which records settle it?
Four risks deserve a check before any demand reading is trusted. Concentration: one large account can hold most committed scope, which is why the dashboard shows its share. Channel drift: a campaign or a new intermediary can lift requests while activation stays flat. Substitution: synthetic data, public data or a stronger base model can meet a need that once required licensed records. Lag: acceptance and payment trail requests by months, so the newest cohorts always look weakest.
Anyone weighing an AI data company, SourceX included, can ask for five records: each cohort month's account count with activated, committed and realized rates at fixed ages; the channel list with start dates; every definition change and whether earlier figures were restated; the largest account's share of committed scope; and, for one cohort, the dated documents behind each rung, so a sample of stage claims can be traced.
Methods note: the ladder, dashboard and index rules were drafted on October 26, 2026 from this page's protocol and the BLS, SEC, Indeed Hiring Lab and OpenAI pages listed below. Nothing here is measured, and no cohorts, rates or index values are reported. SourceX's own company and commercial metrics go to investors directly, in the materials requested from its investor page.
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
- OpenAI's Data Partnerships page asks for large-scale datasets that are not easily accessible online, particularly data that expresses human intention. Source
- OpenAI describes GDPval as an evaluation of model performance on real-world, economically valuable tasks across 44 occupations, built from the work of professionals with an average of 14 years of experience. Source
- The Bureau of Labor Statistics counts a job opening only when a specific position exists, work could start within 30 days and the employer is actively recruiting from outside, and excludes positions filled by a hire who has not yet reported. Source
- Indeed Hiring Lab's Job Postings Index is a seven-day trailing average of seasonally adjusted postings, indexed to 100 on February 1, 2020. Source
- SEC Release 33-10751 says a company that changes how it calculates a metric should consider disclosing the change, its reasons and its effect on current and previously reported amounts. Source
Frequently asked questions
What should the request taxonomy and anonymized reporting method include?
The taxonomy codes each written request on modality, record domain, intended use, volume band, history depth, language, source geography, licensing terms, requester type and verification status. The reporting method publishes aggregates only, with pre-registered minimum cell sizes, a dominance rule, complementary suppression and no requester names or verbatim text.
Does this page report how much demand there is for any data type?
No. It is a protocol only and reports no results. Any figures would need a registered wave of coded requests, published with the denominator, intervals and the limits stated in the design.
Which sample, baseline and interval would make a tracker figure defensible?
A named set of channels over stated dates, a denominator of in-scope written requests, a fixed taxonomy version, a double-coded subset with its agreement result, and an interval for every share. A baseline comes from an earlier wave run under the same design.
Why count distinct requesters for suppression rather than requests?
One organization can send many requests, so a cell with several requests may still describe a single buyer's plan. Counting distinct requesters for the minimum cell size, and applying a dominance rule, keeps one requester from being identifiable in a published table.
Can a tracker based on one company's inbound requests represent the whole market?
No. It represents the requests that reached that company's channels. That can still be useful, provided every published share is described as a share of those requests and the channel's blind spots are stated.
Sources
- AAPOR's disclosure standards call for collection dates, sample sizes by frame and frame coverage, and caution against precision measures for non-probability samples without a described, validated model. Source
- Datasheets for Datasets asks whether a dataset is a sample of a larger set, how it was sampled and whether it is representative. Source
- NIST SP 800-188 addresses limiting disclosure risk to individuals and establishments while still allowing meaningful statistical analysis. Source
- NIST's Dataplot reference lists Wilson, adjusted Wald and Jeffreys intervals for proportions, defaults to Wilson, and notes Wilson or Jeffreys are recommended for small samples. Source
- The FTC requires firms to possess and rely on a reasonable basis for objective claims before making them. Source
- The Center for Open Science describes preregistration as setting out a study plan before seeing the data, so readers can tell planned choices from later ones. Source
See if your company qualifies
A short company assessment. No data uploads are needed.