Skip to content

AI data market

Credible AI data market research: 8 checks before any number

By SourceX Editorial · Updated

10 sections

Short answer

Credible AI data market research names what it counts, traces each input to a dated primary page or labels it illustrative, shows arithmetic anyone can rerun and states a range with the rule behind it. The rule matters most: in this page's illustrative four-input demand workbook, the widest single-input swing is 4x, the central 80% band 5.45x and the all-low to all-high bound 72x.

Value drivers

Credible AI data market

This page covers credible AI data market research: 8 checks before any number.Illustrative

Key takeaways

  • Before an AI data market estimate enters a model, check its unit, primary inputs, arithmetic and range rule; an estimate that fails those four belongs in context at most.
  • The range rule can matter more than any input: with six inputs each uncertain by a factor of 2, a tornado table shows 2x, an exact 80% band 3.57x and the all-low to all-high bound 64x.
  • Build supply and demand separately: in the illustrative workbook the transacted median falls to 67, below both the demand central of about 85 and the supply central of 120.
  • AI spending, data volume and venture funding enter a model only through labeled carve-out shares, and each share adds a multiplied input that widens the honest range.
  • Ask whether the publisher has ever compared an old estimate with the outcome; without that, any growth rate in the estimate is unscored.

What makes AI data market research credible?

Credible AI data market research does five things: it defines exactly what it counts, traces every input to a dated primary page or labels it illustrative, shows arithmetic a reader can rerun, states a range together with the rule that produced it, and checks itself against something independent. An estimate that does all five can be argued with line by line. One that does none can only be believed or ignored.

The decision for an investment associate is whether a given estimate may enter a model or memo, and in what form: as an input, as context, or not at all. The evidence that supports a useful answer is the estimate's own documentation. If the definition, sources, arithmetic and range rule are not published, the number cannot be checked, however large or precise it looks.

This page builds the test, applies it to an illustrative bottom-up workbook, and shows a point many estimates leave unstated: when several uncertain inputs are multiplied, the stated range depends more on the combining rule than on any single input. It states no market size of its own; every workbook value is illustrative.

Eight checks borrowed from statistical agencies and securities rules

The checks below turn general quality rules into questions about one estimate. Two U.S. statistical standards anchor the uncertainty checks: the Census Bureau's Statistical Quality Standard E2 requires information products to state that data are subject to sampling and nonsampling error and to accompany weighted estimates with measures of uncertainty such as standard errors, margins of error or confidence intervals, and Standard F2 asks documentation to discuss nonsampling errors such as coverage, nonresponse and processing error (both paraphrased). They govern Census products, not market reports, but they set the floor a reader can ask for.

Securities rules add the disclosure checks. Item 10(b) of the SEC's Regulation S-K, which covers management projections in U.S. filings, encourages projections that have a reasonable basis and an appropriate format, asks management to avoid presenting only favorable items, to give the most probable amount or most reasonable range without making the range so wide that the disclosure becomes meaningless, and to disclose an outside reviewer's qualifications and relationship to the company (paraphrased). The rule binds filers, not research firms; the questions travel well.

Apply all eight checks; an estimate that fails the first four should not enter a model at all:

Eight checks borrowed from statistical agencies and securities rules
CheckPasses whenFails whenAnchor
DefinitionThe unit, buyer and seller sides, geography and year are named"The AI data market" with no unitUnit-of-analysis rule in report design
Primary inputsEach input is read on its publisher's own page with year and definition, or labeled illustrativeInputs come from unnamed reports or press summariesReasonable basis, Regulation S-K Item 10(b)
ArithmeticEvery operation and input is printed and a reader can recompute the resultA total appears with no path to itReproducibility
Range and ruleA range is stated with its rule: single-input swings, a probability band with assumptions, or all-low to all-highA point estimate, or a range with no ruleCensus Standard E2; Item 10(b); JCGM 101
Error sourcesCoverage gaps, double counting and proxy bias are namedSilence on what the number missesCensus Standard F2
Two-sided checkA supply-side count and a demand-side count are both built and comparedOne side onlyReconciliation
Track recordPast estimates are compared with what happenedNo published comparisonEIA retrospective review
IndependenceFunding, reviewers and their relationships are disclosedUndisclosed sponsor or reviewerItem 10(b) outside review disclosure

Why the range rule moves the answer more than any one input

Because multiplying inputs multiplies their uncertainty, and the three common ways of reporting a range diverge fast. Take a chain of inputs, each known only to within a factor of 2, independent of one another, and with every value in its range equally likely on a log scale. A one-input-at-a-time sensitivity table, the familiar tornado chart, always shows a widest swing of 2x, however many inputs are chained. The all-low to all-high bound doubles with each input: 64x for six inputs. The central 80% band, computed exactly, grows slowly: 2.84x for four inputs and 3.57x for six.

None of the three is wrong; they answer different questions. The single-input table says which input to measure first. The all-low to all-high bound is what you get if every input errs in the same direction, which is realistic when the inputs share a cause, such as one optimistic analyst choosing all of them. The probability band is honest only if the inputs really are independent and the assumed distributions are stated. The Joint Committee for Guides in Metrology's Supplement 1 to the uncertainty guide (JCGM 101:2008) describes propagating probability distributions through a model with any number of inputs and a single output and defines the probabilistically symmetric coverage interval (paraphrased); the band here is that kind of interval, computed exactly instead of by sampling.

The practical test for any estimate: which rule produced its range, and does the text say so? The figures for one to six inputs, each spanning a factor of 2:

Why the range rule moves the answer more than any one input
Inputs multipliedWidest single-input swingCentral 80% bandAll-low to all-high
12x1.74x2x
22x2.15x4x
32x2.48x8x
42x2.84x16x
52x3.20x32x
62x3.57x64x
Nested columns on a log scale for one to six multiplied inputs, each uncertain by a factor of 2: the outer column, all-low to all-high, doubles from 2x to 64x, the inner column, the central 80% band, grows from 1.74x to 3.57x, and a dashed line marks the widest single-input swing at 2x for every count.
Ratios of high to low for one to six multiplied inputs, each spanning a factor of 2, independent and log-uniform.

A two-sided workbook with every input labeled

A credible bottom-up estimate builds the demand side and the supply side separately and lets the smaller one bound the result. The illustrative workbook counts accepted business-record dataset licenses per year, a count rather than money, from four demand inputs and three supply inputs. Every value is illustrative, chosen to show the method; none is an estimate of the real market or of any company's activity, SourceX's included.

Each input spans a stated low and high, and its central value is the geometric midpoint. Demand is I1 x I2 x I3 x I4: central about 85 (60 x 4 x 0.5 x 0.707), all-low 10, all-high 720. Its widest single-input swing runs from 42 to 170, a factor of 4. Its central 80% band, with the inputs independent and log-uniform, runs from 36 to 198, a ratio of 5.45. Supply is S1 x S2 x S3: central 120, all-low 15, all-high 960, with an 80% band from 48 to 298.

Reconciling the two pulls the answer down. Transactions cannot exceed the smaller side, so the transacted count is min(demand, supply): its median is 67, below both the demand central (85) and the supply central (120), with an 80% band from 33 to 138. In 36% of the joint distribution supply is the smaller side. A one-sided estimate would have missed that bound entirely.

The inputs, their ranges and the evidence that would replace each illustrative value:

A two-sided workbook with every input labeled
Input (illustrative)LowCentralHighEvidence that would replace it
I1 Organizations licensing outside business records in a year3060120Counted buyers in filings, procurement notices or disclosed agreements, dated
I2 Business-record datasets each licenses per year248Buyer interviews with a stated frame and response count
I3 Share bought directly from the record holder1/31/23/4Coded disclosed agreements by source type
I4 Share accepted after delivery review1/20.7071Acceptance outcomes from buyers or intermediaries, with denominators
S1 Companies with licensable records that would consider licensing in a year200400800Official business counts filtered by stated rules
S2 Share passing rights, privacy and quality review15%30%60%Pass rates from a documented review process
S3 Datasets each can deliver in a year0.512Delivery records per supplier per year

What the workbook leaves out on purpose

Every estimate excludes something, and a credible one lists its exclusions where the number is quoted rather than in an appendix. This workbook leaves out the following, and any figure taken from it carries these limits:

At a glance

What the workbook leaves out on purpose

  • Money: it counts accepted dataset licenses, not spend, prices or revenue, so it cannot be turned into a dollar market size without a price input that would need its own source and range.
  • Geography and year: no region or calendar year is defined, which a real estimate must fix before any input is measured.
  • Other data types: web text, media, synthetic data and human labeling services are outside the unit; adding them would mix units.
  • Correlation: the probability bands assume independent inputs; if one cause drives several inputs, the honest range moves toward the all-low to all-high bound.
  • Timing: a yearly count says nothing about how fast a pipeline converts, and no growth rate can be read from it.
  • Any company's share: the count describes a category, not what any one participant, SourceX included, does or will do.

Has the publisher ever tested an old estimate against the outcome?

Ask for the comparison, because a forecaster that has never checked itself has an untested error rate. The U.S. Energy Information Administration shows what that looks like: its Annual Energy Outlook Retrospective Review compares Reference case projections from past editions with the values later observed and reports the average absolute percentage difference, the simple mean of the absolute percent differences for each projected year, while noting that the Reference case is not meant as a forecast (paraphrased from the review page and report).

AI data licensing is too new for a long record, but the habit can start now. A publisher that dates each estimate, archives it and later reports how the realized figure compared gives investors something no headline can: a measured error band for its own method. Where no comparison exists, treat any growth rate in the estimate as unscored, and weight the definition and the range rule more heavily than the number.

Why AI spending, data volume and venture funding inflate licensing demand

Because each measures something larger than licensing, and the share that reaches licensed business records is itself an unknown that needs its own range. AI spending counts compute, staff, software and services; total data volume counts bytes, most of which no one has the right or the need to license; venture funding counts money raised by companies, not money spent by buyers. None of the three is a buyer budget for licensed records.

A proxy can still enter a model, but only through a labeled carve-out share with a low and a high, and that carve-out adds one more multiplied input. By the arithmetic above, each added input widens the honest range: apply three illustrative shares to a proxy, each spanning a factor of 2, and the single-input tornado still shows 2x while the all-low to all-high bound is 8x; if the proxy figure is itself uncertain by a factor of 2, the bound reaches 16x. That is why the workbook here starts from counted units instead.

Risks that could make this credibility test mislead

The test itself can mislead in several ways, and a reader should weigh these before relying on it:

  • A well-documented estimate can still be wrong. The checks reward transparency, not accuracy; a clean definition and visible arithmetic make errors findable, not absent.
  • The independence assumption flatters probability bands. Market inputs often share causes, such as one buyer cycle or one regulatory change, which pushes the true range toward the all-low to all-high bound.
  • Distribution choices matter. The log-uniform assumption used here is one reasonable choice among several, and a different shape changes the band.
  • Survey and filing standards were written for other purposes. Census standards govern official statistics and Regulation S-K governs filings; using them as yardsticks for private market research is an analogy, not a requirement.
  • Concentrated demand breaks averages. If a few buyers account for most purchases, inputs like I2 hide a skewed distribution and a small change at one buyer moves the total.
  • Supply and rights can shift fast. Changes in privacy rules, platform terms or substitute data could change S2 or I3 between estimates, so any figure needs a date and a revision plan.

Questions to ask any company that cites a market estimate, SourceX included

The same eight checks work as diligence questions for any company that cites a market estimate in a deck, and SourceX belongs on that list too. SourceX's investors page describes it as licensing business data directly from companies and licensing it onward to AI developers in its own name; the same page says company and commercial figures go to investors directly. SourceX is raising from verified accredited investors under Rule 506(c), and its materials are requested through the investors page.

Questions worth asking, in order:

Checklist · 8 items

  1. What unit does your market estimate count, on which side of the transaction, in which geography and year?
  2. Which inputs were read on a primary page, with year and definition, and which are assumptions? Show both lists.
  3. Can I rerun the arithmetic from what you have shown me, and get your number?
  4. What rule produced your range: single-input swings, a probability band with stated distributions, or all-low to all-high?
  5. What does the demand-side count look like next to a supply-side count, and which one binds?
  6. Which proxies, if any, entered the model, and what carve-out share and range did each get?
  7. Have you compared an earlier estimate with what happened, and what was the error?
  8. Who funded and who reviewed the estimate, and what is their relationship to you?

0 of 8 ticked. Tick each item as you go.

Methods note: sources, arithmetic and what was not measured

Sources, checked 2026-10-28: Census Bureau Statistical Quality Standards E2 and F2; SEC Regulation S-K Item 10(b) on the eCFR; EIA's Annual Energy Outlook Retrospective Review page; JCGM 101:2008 at the BIPM; SourceX's investors and how-it-works pages. Regulator and standards statements are paraphrased from domain-limited search excerpts, because direct fetches were blocked; a reviewer should open each page before release. No market size, growth rate, share or company figure from any outside source is used.

Arithmetic, part one: for k multiplied inputs each spanning a factor of 2, independent and log-uniform, the base-2 logarithm of the result relative to its central value is the sum of k independent uniform variables on minus one half to one half. Its 10th and 90th percentiles come from the exact Irwin-Hall distribution, inverted by bisection; the 80% band ratio is 2 raised to the distance between them. The widest single-input swing is 2 for any k and the all-low to all-high ratio is 2 to the power k. Band ratios are rounded to two decimals.

Arithmetic, part two: each workbook input is log-uniform between its low and high; the central value is the square root of low times high; inputs are independent. The distributions of the base-10 logarithms of demand, supply and min(demand, supply) were computed by numerical convolution on a fixed grid with step 0.0001, without sampling; percentiles are read from the cumulative grid by linear interpolation. Single-input swings hold the other inputs at their central values. Workbook results are rounded half up to whole numbers in the text (for example 36.3 becomes 36 and 198.1 becomes 198), the probability that supply is the smaller side (0.36) to two decimals, and the demand band ratio of 5.45 is computed from unrounded percentiles. Scripts: xs0373_ranges.py and its rerun xs0373_ranges_v2.py, which corrects how percentiles are read from the grid; results in xs0373_ranges_v2_out.json, in the writer's scratch bin folder.

Not measured: the size of any real market, any buyer's or seller's activity, prices, and any SourceX metric. The workbook demonstrates a method; its outputs are illustrative counts and should not be quoted as estimates.

Frequently asked questions

How can you tell whether an AI data market estimate is credible?

Check whether it names its unit, sources each input to a dated primary page or labels it as an assumption, shows arithmetic you can rerun, and states a range with the rule behind it. Then look for a supply-side cross-check, a comparison of past estimates with outcomes and disclosed funding and reviewers. An estimate that fails the first four checks should not enter a model.

Why do market estimates with the same inputs report such different ranges?

Because the rule for combining uncertain inputs differs. With six inputs each uncertain by a factor of 2, a one-at-a-time sensitivity table shows a 2x swing, an exact 80% band for independent inputs is 3.57x wide and the all-low to all-high bound is 64x. A credible estimate says which rule it used.

Is a top-down estimate from AI spending ever usable?

Only as a labeled carve-out: the share of spending that reaches licensed business records must enter as its own input with a low and a high, and the result carries the wider range that adds. Spending also includes compute, staff and software, so the starting figure is never a buyer budget for records.

Why build both a demand-side and a supply-side estimate?

Because transactions cannot exceed the smaller side. In this page's illustrative workbook, reconciling the two moved the median from a demand central of about 85 and a supply central of 120 down to 67, a bound a one-sided estimate would miss.

Does this page estimate the size of the AI data market?

No. Every workbook value is illustrative and chosen to demonstrate the method. The page uses no outside market size, growth rate or share, and states no SourceX metric.

Sources

  • The Census Bureau's Statistical Quality Standard E2 (Reporting Results) requires information products to state that data are subject to sampling and nonsampling error and to accompany weighted estimates with measures of uncertainty such as coefficients of variation, standard errors, margins of error or confidence intervals. Source
  • The Census Bureau's Statistical Quality Standard F2 asks documentation to discuss potential nonsampling errors such as nonresponse, coverage, processing and measurement error and to give indicators of statistical precision. Source
  • Item 10(b) of Regulation S-K (17 CFR 229.10) encourages management projections with a reasonable basis and appropriate format, warns against showing only favorable items, asks for the most probable amount or a reasonable range that is not so wide as to be meaningless, and asks filings to disclose an outside reviewer's qualifications and relationship. Source
  • JCGM 101:2008, Supplement 1 to the Guide to the expression of uncertainty in measurement, describes propagating probability distributions through a measurement model with any number of input quantities and a single output by a Monte Carlo method, and defines the probabilistically symmetric coverage interval. Source
  • EIA's Annual Energy Outlook Retrospective Review compares Reference case projections from past editions with realized values using the average absolute percentage difference, and notes that the Reference case is not meant as a forecast. Source
  • SourceX's investors page says SourceX licenses business data directly from companies and licenses it onward to AI developers in its own name, and that it shares company and commercial metrics with investors directly. Source
  • SourceX's how-it-works page describes how SourceX licenses business data directly from companies for AI training and evaluation, with the company approving scope and terms and signing before any data is delivered. 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