Skip to content

Deal economics

How to measure data licensing buyer retention when nobody subscribes

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

Measure data licensing buyer retention as a repeat-purchase cohort: group buyers by the date of their first paid license, count a second paid license rather than another delivery, set the window longer than the purchase cycle and license term, and publish every window with its denominator. On 23,570 first-time buyers in the public CDNOW file, the same cohort shows 17.5% repeat at 30 days and 48.3% at 462 days.

Record timeline

Data licensing buyer

This page covers how to measure data licensing buyer retention when nobody subscribes.Illustrative

Key takeaways

  • Subscription retention measures cancellations against a run rate; a business selling one-off or term data licenses has no run rate to cancel, so its retention has to be counted as repeat purchases.
  • The window decides the number: in the public CDNOW cohort, repeat rises from 17.5% at 30 days to 48.3% at 462 days, and 40.2% of returning buyers first came back after day 90.
  • Count licenses, not rows: refresh deliveries, invoice lines and amendments under one license are not repeats, and a buyer whose term has not ended cannot yet renew.
  • Show count and dollars together: in the same cohort, the top tenth of repeat buyers carried 47.9% of repeat dollars within the window.
  • Definitions drift, as Twilio's 2020 change to its expansion cohort shows, so ask for the definition in force for every period on the chart.
  • SourceX publishes no buyer, repeat or revenue figures and shares company and commercial metrics with investors directly.

What should data licensing buyer retention measure when nobody subscribes?

It should measure whether buyers come back and pay again, counted as a second paid license inside a window that fits how often a buyer can buy. That is a repeat-purchase cohort: group buyers by the date of their first paid license, follow every group for the same length of time, and report the share that bought again, with the count beside the percentage.

Subscription metrics do not carry over. A subscription business opens each year with contracts that renew unless someone cancels, so its retention rate mostly measures cancellations. A transactional data licensor sells a dataset, a batch of deliveries or a license for a fixed term, and once that license is paid nothing renews by default. The buyer has to decide to buy again, usually after an evaluation, a budget cycle and another rights review.

SourceX's live terms show the transactional shape on the supply side. It usually pays a one-time fee for a dataset; some agreements set recurring deliveries, each paid as the agreement states; and returning suppliers reuse their Master agreement and sign a new Dataset Deal Confirmation for each new deal. The buyer side of its onward agreements is not published. Anyone judging a licensor built this way needs a retention measure designed for purchases that never renew on their own.

How four public filers define retention, and which definition travels

The definitions worth borrowing count purchases inside a stated window and say which customers they leave out. Four annual reports on Form 10-K show the range. The table paraphrases each definition from search excerpts of the filings on sec.gov; no figure is taken from any of them.

Wayfair counts active customers as those who bought at least once directly from its sites in the preceding twelve months, and its filing notes that a change in that count captures both the inflow of new customers and the outflow of customers who have not bought in the last twelve months, so a rising total can hide falling retention. Twilio's expansion rate follows a fixed cohort of accounts and compares the same quarter a year apart, which suits usage that rises and falls. Its 2021 report also shows a definition moving: large accounts that never signed twelve-month minimum commitments were outside the cohort through 2019 and inside it from the first quarter of 2020.

MSCI's retention rate starts from a subscription run rate and subtracts cancellations, and according to the excerpt of its 2025 report it is not calculated for the part of its run rate tied to asset-based fees. Clarivate describes transactional revenue as deliverables priced per product, data set or project, yet the renewal rate in its 2021 report covers subscription agreements only, even though its 2020 report says transactional buyers are often repeat customers who also subscribe.

That is the diligence point. In these filings, companies with mixed models publish retention for the recurring part and leave the transactional part without a repeat measure, so a licensor that is transactional by design has to build its own and show how. The four definitions, side by side:

How four public filers define retention, and which definition travels
Filer and metric (Form 10-K)What it countsWindow or cohortExclusions and changesWhat a data licensor can borrow
Wayfair: active customers (2023)Customers who bought at least once directly from its sitesPreceding twelve months at each period endNet change mixes new customers in with lapsed customers out, as the filing notesSeparate inflow from outflow, and count both by entry cohort
Twilio: dollar-based net expansion rate (2021)Revenue from accounts active a year earlier, this quarter over the same quarter a year beforeFixed account cohort, same quarter a year apartAccounts without twelve-month minimums excluded through 2019, included from Q1 2020The fixed cohort, with every scope change disclosed
MSCI: retention rate (2025)Subscription run rate kept after cancellations, over the opening run rateFiscal year; interim periods annualize cancellationsNot calculated for asset-based feesA warning: run-rate retention has nothing to measure in one-off licenses
Clarivate: annual renewal rate (2021, with 2020 and 2023 definitions)Annualized value of subscriptions renewed over value up for renewalAgreements due for renewal in the periodSales priced per data set or project sit outside itThe gap a transactional licensor has to fill

A public test: one cohort of first-time buyers, five windows

One entry cohort followed in a public file shows how far the window alone moves a repeat rate: from 17.5% at 30 days to 48.3% at 462 days for the same 23,570 buyers. No data licensing company publishes buyer-level purchase histories, so the test uses the closest public record of the same measurement problem.

The CDNOW master file, documented by Peter Fader and Bruce Hardie, holds the full purchase history to the end of June 1998 of the 23,570 people whose first purchase at the online music retailer CDNOW came in the first quarter of 1997. It has become a benchmark for buying without contracts: nothing renews, and every repeat is a fresh decision. The copy analyzed here sits in the lifetimes Python library's repository at commit 4f2833f.

Each cohort column below shows only windows that all of its members completed before the file ends; the latest first purchase was on 25 March 1997, so 462 days is the longest window every member finished. The three monthly cohorts are never more than 2 percentage points apart at any window, so the spread down each column comes from the window, not from the cohort:

A public test: one cohort of first-time buyers, five windows
Window after first purchaseJanuary 1997 cohort (7,846)February 1997 cohort (8,476)March 1997 cohort (7,248)All first-quarter buyers (23,570)
30 days17.5%17.6%17.3%17.5%
90 days29.9%28.6%29.2%29.2%
180 days37.4%37.2%37.8%37.4%
365 days45.8%45.0%46.9%45.9%
462 days, the longest window every buyer completed48.3%47.6%49.1%48.3%
Horizontal bars for 23,570 first-time CDNOW buyers from the first quarter of 1997: repeat within 30, 90, 180, 365 and 462 days at 17.5, 29.2, 37.4, 45.9 and 48.3 percent, and two trailing twelve-month readings, active at 35.4 percent and two or more purchase days at 21.7 percent.
CDNOW master file (Fader and Hardie), lifetimes repository at commit 4f2833f; computed for this article. Not SourceX data.

Why the window has to outlast the purchase cycle and the license term

Because a short window books slow buyers as lost. In the CDNOW file, the median buyer who came back did so 57 days after the first purchase, yet 40.2% of all buyers who returned before the file ends made that first return after day 90, and 23.4% after day 180. A 90-day figure would have written off two in five of the people who did come back.

Data licensing stretches the cycle further. A buyer may need to train and evaluate on a dataset before it knows whether more of the same is worth buying, and procurement, security and rights reviews run again for each new license. A term license adds a hard floor: a buyer on a 24-month license cannot renew before month 24, so a renewal rate computed over all buyers mixes people who were never eligible with people who declined.

Run two clocks. The expansion clock, for a second dataset or a wider scope, should run at least as long as the longest evaluation-to-purchase interval the company can document from its own records. The renewal clock should start when a term ends, with a denominator of buyers whose term ended inside the window. Report both by entry cohort, and leave a cell blank until every buyer in that cohort has had the full window.

Rows, deliveries and licenses: what counts as a second purchase

Only a second paid license counts; another delivery, invoice line or amendment under the same license does not. The CDNOW file shows the mechanical version of the mistake. Of its 69,659 records, 2,068 are extra rows on a day when the same customer already had one, and counting any second row as a repeat lifts the any-time repeat share from 48.9% to 49.5%: 146 customers had extra rows only on their first purchase day, and one more had nothing later but a zero-dollar row.

The data licensing version is larger. A license with monthly refresh deliveries produces twelve delivery events a year from one purchase decision, and an amendment that adds a field is not a new decision either. Make the signed license or order form the unit, and log refreshes, partial acceptances and amendments as attributes of that license.

Calendar windows answer a different question. Counted the way Wayfair counts active customers, over the twelve months ending on the file's last day, 35.4% of the 1997 cohort were active, and requiring purchases on two or more dates in that window leaves 21.7%. Those shares describe who was still buying late in the period, not how much of the cohort came back, which was 45.9% within 365 days.

Resellers and related buyers need their own line. Fader and Hardie's own notes base their exploratory analysis of the file on 23,560 customers rather than all 23,570, setting a few buyers aside before studying the rest. For a data licensor, an affiliate, a reseller, or a buyer that is also an investor or a supplier can buy again for reasons unrelated to the data, and should be shown with and without.

The repeat-purchase cohort specification for a data licensor

The specification turns the public test into a reporting standard a licensor can fill from its own contracts and invoices. Each rule names its unit and its denominator, and the last column shows the form in which the CDNOW test applied it:

The repeat-purchase cohort specification for a data licensor
RuleData licensing definitionDenominator or treatmentHow the CDNOW test applied it
Unit of repeatA second signed and paid license or order: a new dataset, a wider scope or a renewed termDeliveries, refreshes, invoice lines and amendments under one license are attributes, not repeatsOne purchase per customer per day; extra same-day rows never count
Entry cohortBuyers grouped by the date their first license was paidEach buyer sits in one cohort, set by its first paid dateThree monthly cohorts, January to March 1997
Expansion windowAt least the longest documented evaluation-to-purchase intervalA cell stays blank until every cohort member has had the full windowWindows of 30 to 462 days; 462 is the shortest follow-up of any buyer
Renewal windowStarts when a term license endsOnly buyers whose term ended inside the windowNot applicable: CDNOW sales carried no terms
Paid testA collected or contractually due amount above zeroPilots, credits and zero-value licenses on their own lineZero-dollar occasions never count as repeats: 80 of them, 72 on a first purchase day
Related and reselling buyersAffiliates, resellers, investors and suppliers that also buyReported with and without themWithout the ten largest spenders the 462-day rate stays 48.3%
Dollar viewRepeat revenue and its concentration, next to the countTop decile's share of repeat revenue, plus the top buyer's shareThe top tenth of repeat buyers held 47.9% of repeat dollars within 462 days

Who carries the repeat revenue? Counts and dollars tell different stories

Report dollars beside the count, because repeat revenue concentrates even when repeat buyers are many. In the CDNOW cohort, purchases after each buyer's first day supplied 68.6% of all dollars spent to June 1998, and the top tenth of buyers who came back within 462 days, 1,139 of 11,387, accounted for 47.9% of repeat dollars in that window.

A data licensor will have far fewer buyers than a retailer, so concentration is sharper and one account can decide the dollar figure. Let the count lead, the share of buyers who bought again, and print the dollar measure and the top buyer's share of repeat revenue next to it. A dollar-weighted retention ratio above 100% can come from one buyer expanding while most never return.

Small cohorts need their uncertainty on the page. In an illustrative cohort of twelve buyers where six return, the repeat rate is 50%, and one buyer more or fewer moves it by more than eight points. Print the count with every percentage, and pool quarters into half-years before computing rates when cohorts are that small.

What could make repeat-purchase evidence mislead

Five things can make a repeat-purchase cohort look stronger or weaker than the business behind it. Each is checkable in documents a licensor should already hold:

  • Definitions drift. Twilio moved large accounts without twelve-month minimums into its expansion cohort from the first quarter of 2020, and a licensor can widen what counts as a repeat just as quietly. Ask for the definition in force for every period on the chart.
  • The consumer analog flatters speed. CDNOW buyers could order again within days; a data buyer may need a full evaluation and procurement cycle. The public test shows how windows and units move a rate, not what any data licensor's rate is or should be.
  • Repeat demand needs fresh supply. A buyer can only buy again when there is new data it wants, so a licensor's repeat rate depends on its ability to source the next dataset as much as on how the first one performed.
  • Budgets and substitutes end relationships that went well. A buyer that pauses data spending, or finds synthetic or public data that pass its acceptance test, stops repeating regardless of quality, and with few buyers one pause can halve a cohort's rate.
  • Survivorship in what gets shown. A company may present only cohorts old enough to look good, or drop a cohort whose buyers left. Ask for every cohort since the first paid license, blank cells included.

Questions to ask any data licensor about repeat buyers, SourceX included

SourceX publishes no buyer, repeat, revenue or pipeline figures; its investor page says it shares company and commercial metrics with investors directly. For readers weighing the company itself, its current offering is limited to accredited investors under Rule 506(c), and any of these questions can accompany a request for materials at sourcex.si/investors.

Ask for the cohort table itself, built from signed licenses and invoices, with the definition that applied in each period. The checklist works for any transactional licensor:

Checklist · 7 items

  1. Which event starts a buyer's cohort: signature, first delivery, acceptance or first payment, and is it the same in every period shown?
  2. What counts as a repeat, and are refreshes, amendments and extra deliveries under one license excluded?
  3. How long is each window, and which documented evaluation-to-purchase interval justifies it?
  4. For term licenses, how many buyers in each cohort reached the end of a term, and how many of those renewed?
  5. Which cells are blank because buyers have not yet completed the window, and is any cohort missing?
  6. What share of repeat revenue comes from the top buyer and the top three, and are affiliates, resellers, investors or suppliers among them?
  7. Has the definition of a buyer or a repeat changed, when, and what would earlier periods show under today's rule?

0 of 7 ticked. Tick each item as you go.

Methods note: the CDNOW count, the filings and what was not measured

Data: CDNOW_master.txt in github.com/CamDavidsonPilon/lifetimes at commit 4f2833f (committed 28 June 2024), read on 28 October 2026 as a data file by a script that imports and runs no repository code. The file holds 69,659 records for 23,570 customers. The script strips any byte-order mark, requires four whitespace-separated fields on every line after the header, parses each date strictly so an impossible date stops the run, and stops if any customer's first record falls outside January to March 1997. A purchase occasion is one customer on one date, with that day's records merged and their dollars summed; an occasion is paid when its total is above zero. A customer's entry date is the date of the first record of any value, and a repeat is a paid occasion on a later date.

A repeat falls within W days when the first repeat comes at most W days after entry, by exact day difference. A cohort cell appears only when every member had at least W days of follow-up before the file's last date, 30 June 1998. The cohort gap is the largest minus the smallest of the three monthly cohorts' exact shares at each window. Trailing measures count paid occasions from 1 July 1997 to 30 June 1998: active means at least one, repeat buyer means two or more distinct dates. The row-count reading counts any customer with two or more records, and the script splits those without a paid repeat into customers whose extra rows all fall on the entry day and customers whose only later rows are zero-dollar. The time to first repeat covers every customer with a repeat before the file ends. The top decile is the first 1,139 of the 11,387 buyers who repeated within 462 days, one tenth rounded up, ranked by repeat dollars inside that window with ties broken by customer id; the reseller check drops the ten customers with the largest total dollars, ties again broken by customer id. Percentages are exact fractions rounded once, half up, to one decimal, and a median of an even count would be the mean of the two middle values, rounded half up. Nothing is random. The script, xs0119_cdnow_cohorts.py, and its outputs, which hold only numeric customer ids and dates, are kept for recount.

Filings: the Wayfair, Twilio, MSCI and Clarivate definitions were verified through search excerpts of their 10-K filings on sec.gov on 28 October 2026, because sec.gov could not be opened from this environment; several excerpts were paraphrases, so the article paraphrases them and takes no figures from them. Fader and Hardie's notes were verified the same way on brucehardie.com. SourceX's terms come from its live how-it-works and investors pages, read the same day. The specification and the checklist were drawn up for this article.

Not measured: any data licensing company's repeat rate, SourceX's included; typical evaluation-to-purchase intervals for data buyers; how license terms change renewal behavior; and how closely consumer repeat buying resembles business licensing. Each would need company records that have not been published.

Frequently asked questions

What is buyer retention for a data licensing business?

It is the share of buyers who pay for another license after their first, measured by entry cohort inside a stated window. Because data licenses are often one-off or fixed-term purchases, retention has to be counted as repeat purchases rather than as subscriptions that were not cancelled.

Is net revenue retention useful for one-time data licenses?

Only as a supporting line. Net revenue retention compares revenue from the same customers in two periods, which assumes the revenue recurs; with one-off licenses the ratio swings with the timing of a few large deals. Lead with the count of buyers who bought again, and show repeat dollars and the top buyer's share next to it.

How long should a repeat-purchase window be for data buyers?

At least as long as the longest evaluation-to-purchase interval the company can document, and for renewals long enough to run past the end of the license term. In the public CDNOW test, 40.2% of buyers who returned did so after day 90, so a 90-day window would have missed them.

Does a license renewal count the same as buying a new dataset?

Count both as repeats but report them on separate lines. A renewal shows the buyer still values the same data; a new dataset or wider scope shows the relationship growing. Each needs its own denominator, and renewals belong only to buyers whose term has ended.

Why test a data licensing metric on a music retailer's purchase file?

No data licensing company publishes buyer-level purchase histories, and the CDNOW file is a documented public record of buying without contracts, where every repeat is a fresh decision. It shows how much the window and the counting unit move a repeat rate; it does not show what a data licensor's rate is or should be.

Sources

  • Wayfair's annual report on Form 10-K defines active customers as individual customers who purchased at least once directly from its sites during the preceding twelve-month period, and notes that the change in active customers captures both new customers and the outflow of customers who have not purchased in the last twelve months. Source
  • Twilio's 2021 annual report on Form 10-K defines its dollar-based net expansion rate on a cohort of accounts active a year earlier and says periods through 2019 excluded large accounts that never entered 12-month minimum revenue commitment contracts, which were included from 2020. Source
  • MSCI's annual report on Form 10-K defines the full-year retention rate as the subscription run rate at the start of the year minus actual cancellations, over the opening run rate; the excerpt of the 2025 report says it is not calculated for run rate tied to asset-based fees. Source
  • Clarivate's 2021 annual report on Form 10-K defines its annual revenue renewal rate as the annualized value of subscription agreements renewed over the annualized value up for renewal. Source
  • Clarivate's 2023 annual report on Form 10-K describes transactional revenues as earned on specific deliverables, usually priced per product, data set or project. Source
  • Clarivate's 2020 annual report on Form 10-K notes that transactional contracts often involve repeat customers who also buy subscriptions. Source
  • The CDNOW master file in the lifetimes repository at commit 4f2833f: 69,659 purchase records for 23,570 customers who first bought in the first quarter of 1997, counted for this article. Source
  • Fader and Hardie's notes describe the CDNOW master file as the entire purchase history to the end of June 1998 of the 23,570 people whose first-ever CDNOW purchase came in the first quarter of 1997, and base an exploratory analysis on 23,560 customers. Source
  • SourceX's how-it-works page says it usually pays a one-time fee for a dataset, some agreements set recurring deliveries paid as each agreement states, and returning suppliers reuse the Master agreement and sign a new Dataset Deal Confirmation. Source
  • SourceX's investor page says it shares company and commercial metrics with investors directly and offers investor materials on request. 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