Deal economics
Revenue per dataset metric: five restatements before you trust a rise
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Revenue per dataset is a ratio with an undefined denominator, so a rising figure can reflect counting rules, bundled services, multi-year terms or exclusive scope rather than better prices. Before trusting it, restate both periods five ways: one counting rule, license revenue only, per license-year, matched scope and per accepted record. In this article's illustrative portfolio, a reported 60 percent rise shrinks to 9 percent.
Value drivers
Revenue per dataset metric
Key takeaways
- Revenue per dataset has no standard definition, so ask for the written counting rule before the number.
- Bundled services, multi-year terms and single large projects inflate the numerator; merges, splits, refreshes and relicensing move the denominator.
- Restate both periods five ways: one counting rule, license revenue only, per license-year, matched scope and per accepted record.
- In the illustrative portfolio, a reported 60 percent rise leaves 9 percent like-for-like, while revenue per candidate worked falls 13 percent.
- Pair the ratio with preparation hours, days to cash and buyer concentration; a rising average can sit on a falling margin.
What does a revenue per dataset metric actually measure?
It measures revenue recognized in a period divided by the number of datasets a company chose to count in that period, and both halves of the ratio are management choices. No accounting standard defines a dataset or asks a data licensing company to report revenue per dataset, so the figure is only as useful as its written definition.
The SEC's January 2020 interpretive release on key performance indicators in MD&A (Release 33-10751) sets a sensible floor: a company presenting a metric should consider disclosing a clear definition, how it is calculated and why it is useful to investors. A private company's data room should meet the same bar before anyone compares its ratio with another company's or with its own history.
A useful answer needs four pieces of evidence for each period: the counting rule for the denominator, the split between license and service revenue, the term behind each license and the usable records the buyer actually accepted. With those, the ratio can be restated five ways and the trend either survives or it does not. This article builds that restatement, applies it to an illustrative portfolio and lists what it cannot tell you.
Why can the denominator move when nobody sells more?
Because a dataset has no fixed unit: a company can merge, split, refresh or relicense the same records and count a different number of datasets each time. Four counting choices change the ratio without any change in price.
Merging is the most common. Three supplier extracts sold together can be counted as one dataset, which triples revenue per dataset for that deal. Splitting does the opposite: one extract cut into regional or yearly slices becomes several datasets. Refreshes are ambiguous: a quarterly update of records already delivered can be booked as a new dataset or as part of the original license. Relicensing produces a second license of the same records but not a second dataset.
Relicensing matters most, because it changes what the metric means. Revenue per dataset rises with every additional license of the same records, which can be the whole point of a data business, while revenue per license stays flat or falls. Ask for both ratios and the count of licenses per dataset; whether a second license is allowed, and what it costs to serve, is a separate rights and economics question.
Rejected candidates are the hidden part of the denominator. A ratio per delivered dataset ignores every candidate that consumed sourcing and preparation and then failed rights review or acceptance. A second ratio, revenue per candidate worked, keeps that work in view.
Which numerator choices inflate revenue per dataset?
Bundled services, multi-year terms and single large projects inflate the numerator most often, and public filings show each one at work.
Services first. Shutterstock's 2024 annual report puts AI training data and custom content in one revenue line, Data, Distribution, and Services, which the filing says includes access to its metadata for machine learning and generative AI model training as well as custom content produced by Shutterstock Studios. That line grew 28 percent to $175.3 million in 2024, while the data offering inside it grew 15 percent. A per-dataset figure built on the whole line would have shown nearly twice the growth of the data business itself.
Term second. According to Reddit's IPO registration statement, the data licensing arrangements it entered in January 2024 totaled $203.0 million in aggregate contract value and ran for terms of two to three years, and it expected at least $66.4 million of revenue from them in 2024. Contract value per deal and recognized revenue per dataset are two different numbers for the same contracts, and neither is a per-year figure until the term is applied.
Recognition timing adds a layer. Under FASB's licensing update, ASU 2016-10, a license whose promise is only to make finished intellectual property available is generally recognized at a point in time, once a copy is available and the license period has begun, while a license that includes supporting or maintaining the property is recognized over time. Depending on the contract and the auditor's analysis, a multi-year snapshot license can land in a single period's numerator.
Projects third. Wiley's quarterly report for the period ended July 31, 2024 attributed approximately $16 million of Learning revenue to an executed $21 million content rights project for training generative AI models, and said Learning revenue excluding that project fell 1 percent at constant currency. One contract turned a shrinking line into a growing one for the quarter; in a young company with a handful of deliveries a year, one project can set the average.
The eight inputs behind an editable restatement
Eight inputs per period are enough to restate revenue per dataset, and each should come from a document an investor can inspect rather than from the metric's own summary slide.
The values below are illustrative index units, not currency, and describe no real company. They were chosen so that each effect is visible on its own: one bundled deal, one block of services, one two-year license recognized at delivery, one exclusive grant and slightly smaller datasets in period 2. Replace each value with the company's records and hold every rule constant across both periods:
| Input | Period 1 | Period 2 | Document that verifies it |
|---|---|---|---|
| Datasets as the company counts them | 10 | 8 | Delivery log and the written counting rule |
| Supplier extracts delivered, one rule for both periods | 10 | 10 | Delivery receipts by supplier and source system |
| Revenue recognized, index units | 1,000 | 1,280 | Revenue schedule tied to the general ledger |
| Preparation and labeling revenue in the same contracts | 0 | 100 | Contract line items or statements of work |
| Multi-year licenses recognized at delivery | None | One two-year license, 240 | Signed agreement and recognition memo |
| Exclusive grants | None | One field-exclusive license, 160 a year | Grant clause in each agreement |
| Accepted records per non-exclusive dataset | 50,000 | 46,000 | Buyer acceptance reports |
| Candidates worked, including rejected | 14 | 17 | Qualification and rejection log |
Restating an illustrative portfolio: from 60 percent to 9
Applied in order, the five restatements turn a reported 60 percent rise in revenue per dataset into a 9 percent like-for-like rise in revenue per accepted record, and each step removes a different kind of noise.
The company reports 1,280 units over 8 datasets in period 2 against 1,000 over 10 in period 1, or 160 against 100. Counting the bundled deal as the three supplier extracts it contains brings period 2 to 128. Removing 100 units of preparation and labeling brings it to 118. Counting the two-year license at 120 a year instead of 240 brings it to 106. Setting aside the exclusive grant, which prices away future licenses of the same records and has no counterpart in period 1, leaves nine non-exclusive datasets at 100 each: no change at all.
The last step checks size. Period 2's non-exclusive datasets averaged 46,000 accepted records against 50,000 in period 1, so revenue per 1,000 accepted records rose from 2.00 units to 2.17, about 9 percent. That is the part of the headline that reflects better pricing of comparable usable records, and the headline hid it rather than revealed it.
The cross-check runs the other way. Spread over every candidate worked, including rejected ones, license revenue per license-year fell from 71.4 units to 62.4, about 13 percent, because period 2 rejected seven candidates against four. A higher price per usable record and a lower yield per unit of sourcing work can both be true, and the reported ratio shows neither.
Each row applies one rule to both periods; the figure plots the change that survives each step:
| Restatement | Rule applied | Period 1 | Period 2 | Change |
|---|---|---|---|---|
| Reported revenue per dataset | Company's own count and recognized revenue | 100 | 160 | +60% |
| One counting rule | One supplier extract is one dataset; the bundle counts as 3 | 100 | 128 | +28% |
| License revenue only | Remove 100 units of bundled preparation and labeling | 100 | 118 | +18% |
| Per license-year | Count the two-year license at 120 a year | 100 | 106 | +6% |
| Matched scope | Non-exclusive licenses only; exclusive grant set aside | 100 | 100 | 0% |
| Per 1,000 accepted records | Divide by accepted records: 50,000 vs 46,000 per dataset | 2.00 | 2.17 | +9% |
| Cross-check: per candidate worked | License-year revenue over all candidates, rejected included | 71.4 | 62.4 | -13% |
Is the rise real? Testing the strongest alternative explanation
Partly. The strongest alternative is not that the price gain is fake but that revenue per dataset, even restated, says nothing about whether the deals made money.
Reported deal margin can disappear after rework, overhead or collection delays. A portfolio that shifts toward bundled, bespoke deliveries can lift revenue per dataset while each delivery takes more preparation cycles, more review and longer to collect. The 100 units of services in the illustration are real revenue, but they came with labor, and whether they earned anything depends on hours the revenue line does not show.
Test the alternative with three requests that sit beside the ratio: preparation hours per accepted dataset in both periods, days from delivery to cash for each license, and the share of revenue from deals that needed more than one acceptance cycle. If revenue per accepted record rose while hours and days held steady, the gain is real. If hours rose faster, the company sold more work, not more value per record.
The opposite error is possible too. Normalizing can strip out genuine progress: a well-priced exclusive grant or a buyer's multi-year commitment is real evidence of demand. Restating does not discard it; it moves it to its own line so it is not mistaken for a trend in comparable deals.
Who is inside the average? Concentration, medians and the largest deal
Ask for the distribution, not just the mean, because a data licensing portfolio with few deals is usually dominated by one or two of them.
Filings show how concentrated licensing revenue can be even at a large company. Reddit's Form 10-K for fiscal 2024 states that "substantially all of the contract value associated with our licensing revenue is derived from two of our partners." For a company with ten deliveries a year and two large buyers, mean revenue per dataset will move with every renewal those buyers decide.
Three numbers fix most of the distortion: the median per period, the largest deal's share of period revenue, and the count of distinct buyers behind the revenue. Where one deal is more than a third of a period's revenue, read the mean as that deal's price rather than as a portfolio trend, and compare medians across periods instead.
What could make this restatement mislead
Restating revenue per dataset removes noise but adds judgments of its own, and an investor should know where each one can go wrong:
- Recognition judgments differ. Whether a license is recognized at delivery or over its term depends on contract facts and the auditor's analysis under ASU 2016-10, so two companies with identical deals can report different period revenue. Restate from the agreements, not from the revenue schedule alone.
- Records are not equal units. Accepted records measure size, not value; 46,000 linked case histories can be worth more than 50,000 unlinked rows. Use the per-record step to detect size effects, never as a price list.
- Scope classes can get too thin. With few deals, separating exclusive from non-exclusive or training from evaluation-only licenses can leave one deal per class. Report the classes and their counts instead of one normalized figure.
- Exclusive grants borrow from later periods. An exclusive license can lift this period's revenue per dataset while taking the same records off the market; a rising share of exclusive revenue is a risk to future periods, not only a gain.
- One buyer or one project can reverse the trend. Reddit's two-partner concentration and Wiley's single $21 million project show how a few contracts can decide a period, so no normalization substitutes for knowing who the buyers are and when their terms end.
- The inputs may not exist. Extract-level delivery logs, line-item service revenue and rejection logs are operating records; if a company cannot produce them, that tells you how its ratio was built.
Questions to ask about revenue per dataset, SourceX included
Ask for the definition first and the number second, and put the same questions to every data licensing company, SourceX among them.
SourceX is one application of the question. It helps companies turn selected business records into licensed AI datasets, handling everything from finding and qualifying suppliers to reviewing rights, contracting, preparing data, delivering it and managing payment. Data is delivered only once the supplying company has approved scope and terms and signed the agreement; SourceX may then use, sublicense and distribute the de-identified dataset to the extent that agreement permits, which is why the questions that follow include licenses per dataset.
SourceX has not published revenue, dataset counts or pricing, and this article reports none. It shares company and commercial metrics with investors directly, and it is raising from accredited investors under Rule 506(c), with materials available on request through its investor page. Questions worth bringing to that conversation, or to any other company's:
Checklist · 7 items
- What is your written counting rule for a dataset, and did it change between the periods shown?
- How much of the revenue in the ratio is preparation, labeling or other services, by contract line?
- Which licenses run longer than a year, and how was each recognized?
- What share of revenue came from exclusive grants, and which records did they take off the market?
- How many licenses does each dataset carry, and what is revenue per license next to revenue per dataset?
- What were accepted records per dataset and candidates worked, rejected ones included, in each period?
- What are the median, the largest deal's share and the number of distinct buyers in each period?
0 of 7 ticked. Tick each item as you go.
Methods note: the restatement model, its sources and blind spots
The eight-input sheet and the five-step restatement are editorial frameworks built for this article on 2026-10-26. Every portfolio value is an illustrative index unit chosen to isolate one effect; none describes a real company or a SourceX transaction. Public evidence comes from Reddit's 2024 registration statement and fiscal 2024 Form 10-K, Shutterstock's fiscal 2024 Form 10-K, Wiley's quarterly report for the period ended July 31, 2024, FASB's ASU 2016-10 and SEC Release 33-10751, each checked on its publisher's own site through search excerpts on 2026-10-26.
Not measured: revenue per dataset at any real company, how often counting rules change in practice, typical service shares in data licensing contracts and the size of exclusivity premiums; none is public for private data licensing companies. The order of the steps is a convention: running them in another order changes the intermediate rows but not the final like-for-like figure.
Frequently asked questions
Is revenue per dataset a GAAP or standard metric?
No. No accounting standard defines a dataset or requires the ratio, so each company sets its own definition. SEC guidance on key performance indicators expects a public company to define any metric it presents and explain how it is calculated, and investors can ask a private company for the same.
What is the best denominator for revenue per dataset?
There is no single best one. Fix one counting rule for every period, then show revenue per license-year and per accepted record within matched scope beside the headline ratio. Add revenue per candidate worked, rejected candidates included, to show how much sourcing work each unit of revenue needed.
Should multi-year data licenses be annualized in this metric?
For trend comparisons, yes. Multi-year terms make contract value and annual revenue diverge: Reddit disclosed $203.0 million of contract value on two-to-three-year terms with at least $66.4 million expected in the first year. Restate each license per license-year and keep the recognized figure on its own line.
Does licensing the same dataset to several buyers inflate revenue per dataset?
It raises it, and that can be real value rather than inflation. Show licenses per dataset and revenue per license as well, so investors can see whether a gain came from reuse or from higher prices, and confirm that the rights behind the records allowed each extra license.
Does SourceX publish revenue per dataset?
No. SourceX has not published revenue, dataset counts or pricing. It shares company and commercial metrics with investors directly through its investor materials request, and any offering is made only to verified accredited investors.
Sources
- The SEC's 2020 interpretive release on KPIs in MD&A says companies presenting a metric should consider a clear definition, how it is calculated and why it is useful to investors. Source
- Reddit entered data licensing arrangements in January 2024 with an aggregate contract value of $203.0 million on two-to-three-year terms and expected at least $66.4 million of revenue from them in 2024. Source
- Reddit's fiscal 2024 Form 10-K states that substantially all of the contract value associated with its licensing revenue is derived from two partners. Source
- Shutterstock's Data, Distribution, and Services revenue, which includes metadata access for AI model training and Shutterstock Studios custom content, grew 28 percent to $175.3 million in 2024 while the data offering grew 15 percent. Source
- Wiley attributed approximately $16 million of quarterly Learning revenue to an executed $21 million GenAI content rights project; excluding it, Learning revenue fell 1 percent at constant currency. Source
- Under ASU 2016-10, a license that only makes finished IP available is generally recognized at a point in time, a license that includes supporting or maintaining the IP over time, and never before a copy is available and the license period has begun. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.