Deal economics
Business context dataset pricing: test the lift a buyer cannot rebuild
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
Business context justifies a higher dataset price only for the measured lift a buyer could not rebuild itself. Test three arms on the buyer's own task: records alone, records plus context the buyer can append, and records plus the supplier's own context. In public CoIL 2000 insurer data, adding the insurer's product records to zip-level context raised buyers found in the top fifth from 120 to 169.
Value drivers
Business context dataset
Key takeaways
- A context premium rests on the gap between supplier-held context and context the buyer can rebuild, not on the gap between bare and enriched records.
- In the public CoIL 2000 data (5,822 customers, 348 caravan policy buyers), zip-level context found 120 buyers in the top fifth, the insurer's own product fields 169, and both together 169.
- Two car policy fields carried 30 of the 49-buyer gain, so the unique value sat in a few fields whose rights and preparation can be checked one at a time.
- Across 20 reshuffled cross-validation runs the gain ranged from 41 to 66 buyers: ask for the range, not a single lift figure.
- Asking prices, public payout ranges and announced deal values do not price context; compare deals only after aligning rights, usable scope, preparation and buyer purpose.
What is a business context premium, and what would prove it?
A business context premium is the part of a dataset's price that pays for information only the supplier holds about each record: the products a customer already owns, how an account was priced, the policy in force when a case was decided, the internal reason code. It is proved by a measured gain on a named buyer task that disappears when the supplier's context is removed and cannot be recovered from context the buyer could add on its own.
Most context claims stop halfway. A seller shows that enriched records beat bare ones and calls the whole difference a premium. Yet a buyer can often append much of that enrichment itself, from a postcode, a calendar, a firmographic file or its own customer base. Paying a supplier for context the buyer could rebuild means paying twice for the same lift.
So the pricing decision has a precise shape. For business context dataset pricing, the question is not whether context helps a model; it nearly always does. The question is how much of the gain survives a substitution test, whether the supplier can license the fields that carry it, and what restoring those fields costs. This page builds that test, runs it on public insurance records, and turns the result into a rights-normalized valuation worksheet.
The substitution test: three arms on one buyer task
Run one model on one set of rows three times and change only the context: arm A holds the records alone, arm B adds context the buyer could append itself, arm C adds the supplier's own context. A premium can rest only on the gap between C and B. The step from A to B is value the buyer already owns.
Three controls keep the arms honest. Rows, outcome, model and evaluation split stay fixed, so a gain cannot come from a friendlier sample. The metric belongs to the buyer and is chosen before the run, for example the number of true cases found at a fixed review capacity. And the spread across repeated splits is reported, because a small positive class makes any single split noisy.
Deciding what goes into arm B is the judgment call, so write it down before any result exists. Context belongs in B when the buyer can produce it from a key it already receives, from public statistics or from its own systems at a cost it accepts. Context belongs in C only when it lives in the supplier's systems and nowhere the buyer can reach. The four arms worth running:
| Arm | Fields included | Who can supply them | What its gain measures |
|---|---|---|---|
| A: records alone | The outcome and the record's own attributes, identifiers removed | Supplier | The floor every other arm is compared with |
| B: plus rebuildable context | Attributes a buyer can append from a key, public statistics or its own data | Buyer or supplier | Value the buyer can create without paying for it |
| C: plus supplier-held context | Account history, products held, pricing, policy terms, internal codes | Supplier only | The only gain a context premium can rest on: C minus B |
| C with top fields removed | Arm C minus the few fields that carry most of the gain | Supplier only | How concentrated the unique value is, field by field |
What restoring an insurer's own context did in the CoIL 2000 data
In the public CoIL 2000 insurance benchmark, adding the insurer's own product records to zip-level context raised the caravan-policy buyers found in the top fifth of customers from 120 to 169, out of 348 buyers among 5,822 customers. Random selection would find about 69.6 of them.
The dataset suits the test because its documentation already separates the two kinds of context. The ISLR package describes 5,822 real customer records with 86 variables: sociodemographic data in variables 1 to 43, derived from zip codes so that everyone in the same zip code shares the same values, and product ownership in variables 44 to 86, where variable 86 records whether the customer bought a caravan insurance policy. The UCI Machine Learning Repository lists the same file as the training set of the CoIL Challenge 2000, donated by Sentient Machine Research.
Treating the zip-level block as arm B is an assumption. Area statistics of this kind are what a buyer can usually attach from a postcode, but whether a given buyer can rebuild these exact 43 fields is something to check, not a given. The product block is arm C by construction: only the insurer knows which of its policies each customer holds. Arm A here has no informative fields at all, so its benchmark is random selection.
The result is lopsided. Once the insurer's 42 product fields were in the model, the zip-level block added no buyers at the 20% cut, 169 either way, and only a little ranking quality, an AUC of 0.731 against 0.720. The unique context did the work and the rebuildable context became redundant beside it. Each arm's result, from five-fold cross-validation in file order:
| Arm | Buyers in top 20% (of 348) | Share of buyers found | AUC |
|---|---|---|---|
| Random selection (expected) | 69.6 | 20.0% | 0.500 |
| Zip-level context only (43 fields) | 120 | 34.5% | 0.634 |
| Zip-level plus 40 insurer fields, car policy fields left out | 136 | 39.1% | 0.682 |
| Zip-level plus the 2 car policy fields | 150 | 43.1% | 0.712 |
| Insurer product fields only (42) | 169 | 48.6% | 0.720 |
| Zip-level plus all 42 insurer fields | 169 | 48.6% | 0.731 |
Why the unique value sat in two fields, and what that does to price
Two of the insurer's 42 fields, which the CoIL data dictionary calls contribution car policies and number of car policies, carried 30 of the 49-buyer gain over zip-level context, about 61%. The other 40 insurer fields, without those two, added 16 buyers. The supplier-held value was concentrated, not spread evenly across the file.
Concentration changes a negotiation in three ways. It narrows diligence, because a buyer can test rights, completeness and refresh for a handful of fields instead of every column. It narrows the price base, because columns that add little cannot claim a share of the premium merely by being supplier-held. And it raises the stakes of each field's rights, since losing one licensable field can take most of the premium with it.
It also tells a supplier what to protect. Fields with unique lift tend to sit closest to the core business, such as what a customer bought, how an account was priced or why a case was escalated, and those are the fields most often covered by customer contracts and privacy notices. A premium that depends on a field the supplier cannot license is an exclusion waiting to be discovered.
The gain needs its range beside it. Across 20 reshuffled five-fold runs, with seeds 1 to 20, zip-level context found 113 to 135 buyers, the full context 170 to 186, and the paired gain ran from 41 to 66 buyers. With only 348 positives, one split can move by a dozen buyers, so a single lift figure quoted without its spread presents noise as precision.
The rights-normalized valuation worksheet for a context premium
The worksheet turns a context claim into something comparable: each row strips out one reason two datasets could differ in price that has nothing to do with context, and the last rows keep only the measured, licensable, supplier-held lift. The right-hand column shows how the public CoIL test fills each row; entries there are public data, assumptions marked as such, or a note that the row was not measured.
Fill one worksheet per buyer task. The same records can carry a large context lift for one task and none for another, so a worksheet without a named task is evidence of nothing. The rows:
| Normalization step | What to record | Evidence that settles it | CoIL 2000 reading |
|---|---|---|---|
| Buyer purpose | The named task, its metric and the review capacity | A written task specification agreed before testing | Find caravan buyers within the top 20% of customers |
| Rights to each context field | Which fields may be licensed for that task | Counsel's field-by-field review of contracts and notices | Not assessed: public benchmark data |
| Usable scope | Rows and fields left after exclusions | An exclusion register with counts | All 5,822 rows and 85 input fields |
| Rebuildable share (arm B) | Lift from context the buyer can append | Arm B result plus the buyer's own cost to append | 120 buyers, assuming zip-level fields are rebuildable |
| Supplier-held lift (C minus B) | The gain the buyer cannot create alone, with its range | Matched three-arm test over repeated splits | 49 buyers in file order; 41 to 66 across 20 seeds |
| Field concentration | The fields that carry most of the gain | Arms that drop or add single field groups | 2 car policy fields carry about 61% of the gain |
| Restoration cost | Joins, keys and history needed to deliver the context | The supplier's preparation estimate per field | Not measured: fields arrive already joined |
| Price evidence type | Asking price, indicative offer or signed term | The document the number comes from | No price exists; task utility only |
Is it context, or rights, scope and preparation in disguise?
Often a price gap credited to context is really a gap in rights, usable scope or preparation. Two datasets drawn from the same records can sell at different prices because one licenses training and the other only evaluation, because one has outcomes linked and the other does not, or because one arrives joined and documented while the other needs months of work. Crediting the whole difference to context overstates it.
That is why deals should be compared only after the first four worksheet rows are aligned. A deal that bought broad rights to a narrow file and one that bought narrow rights to a rich file can print similar totals for opposite reasons. Until rights, usable scope, preparation and buyer purpose match, a price difference says nothing about context.
The strongest objection is that public numbers already settle the matter. They do not. Asking prices are offers, public payout ranges are conditional ceilings, and announced deal values rarely state which rights or fields they bought, so none is a realized, market-clearing price for a defined bundle. SourceX does not publish its commercial terms either; its process page says pricing stays indicative until a Dataset Deal Confirmation fixes the exact dataset, payout and rights for that deal.
Retrieval systems add a practical caution about what counts as unique. Anthropic has reported that prepending short, chunk-specific context to documents before indexing cut failed retrievals in its own experiments. Some context can be regenerated from the document itself at indexing time, and context a model can reconstruct cheaply belongs in arm B, however much effort it once cost the supplier to record.
What could make a context premium wrong
Five developments can erase a measured context lift between the test and the signed terms. Each can be checked before a price is agreed, and a seller who has not checked them is quoting an untested number:
- The buyer rebuilds more than assumed. A buyer with its own customer base or a richer area file has a larger arm B, and the gap shrinks; in the CoIL test, a buyer able to rebuild the two car policy fields would remove most of the premium.
- Joined context raises identifiability. NIST's report on de-identifying personal information describes quasi-identifiers as attributes that identify no one alone but can be linked with other data to identify people, and notes they often cannot simply be masked because they carry analytic value. Context fields are frequently exactly these attributes.
- The fields turn out not to be licensable. Product holdings, pricing history and internal codes often sit under customer contracts or notices that limit outside use, and a field that counsel excludes takes its share of the lift with it.
- One public dataset is not a market. The CoIL result covers one insurer, one task and one model family; the method transfers, the 49-buyer figure does not.
- Generated context substitutes for recorded context. Where a model can produce useful context from the record itself, as contextual retrieval does for documents, supplier-held context loses its scarcity for that task.
Questions for a data company that claims context value, SourceX included
Ask for the three-arm result on a task you name, with arm B defined in writing before the run. A company that can show only enriched-versus-bare comparisons has measured enrichment, not uniqueness, and should not price it as a premium.
SourceX takes licenses to business records from the companies that create them and grants licenses to AI developers itself, so it runs the chain from sourcing and rights review through contracting, preparation, delivery and payment. Its deal and commercial figures are unpublished and go to investors directly. Its current raise is limited to accredited investors under Rule 506(c), with materials available through sourcex.si/investors. Questions to put to it or to any peer:
- On a buyer task you have tested, which fields carry most of the context lift, and are those fields cleared for that use?
- Who decides which context a buyer could rebuild, and is that list signed off before results exist?
- What is the range of the context gain across repeated splits, and how many positive cases does it rest on?
- What does restoring the context cost per dataset, and who bears that cost?
- Which exclusions would remove a high-lift field, and how often has that already happened?
Methods note: the CoIL 2000 context test
Data: the ISLR package's copy of the CoIL 2000 training file, 5,822 customers and 86 variables with 348 caravan policy buyers, read from csv/ISLR/Caravan.csv in the Rdatasets repository at commit 1dcc2bf (SHA-256 beginning 785be82c). It matches the ISLR 1.4 package's data/Caravan.rda in the CRAN mirror at commit b72e0ce cell for cell. The UCI repository describes the same training file but could not be downloaded from this environment, so a rerun on the UCI download is a release check.
Rules as the script applies them: zip-level context is the 43 columns MOSTYPE to MKOOPKLA, insurer context the 42 columns PWAPART to ABYSTAND, and Purchase equal to Yes the outcome. MOSTYPE and MOSHOOFD are one-hot encoded with the levels seen in the training fold, so a level absent from training gets all zeros; every other column is a number standardized with the training fold's mean and standard deviation, and a column with no variation in a training fold is dropped for that fold. The model is logistic regression minimizing the summed log-loss plus 10 times the sum of squared weights, intercept unpenalized, fitted by Newton's method for up to 25 iterations or until the largest step falls below 1e-8.
Headline folds are row position modulo 5 in file order. Out-of-fold scores are pooled, the top 20% is the 1,164 highest scores with ties broken by row order, and AUC counts ties as half. The car policy arms use the fields PPERSAUT and APERSAUT, whose meanings come from the data dictionary printed in the kernlab package's ticdata documentation. The range comes from 20 repeats of five-fold cross-validation: for repeat s, from 1 to 20, perm = numpy.random.default_rng(s).permutation(5822) and row perm[i] goes to fold i modulo 5. With penalties of 1 and 100 instead of 10, zip-level context found 117 and 123 buyers and the full context 168 and 181, so the ordering holds.
Not measured: any price, the cost of restoring context, the rights status of any field, a real buyer's ability to rebuild the zip-level fields, other tasks or other model families. The analysis was run on 2026-10-28 by a script that only reads the data file; no SourceX data is involved.
Frequently asked questions
Does business context always raise a dataset's price?
No. Context raises value only for tasks that use it, and only by the part a buyer cannot rebuild. In the CoIL test the zip-level context stopped adding buyers once the insurer's product records were present, so paying again for that layer would have bought nothing extra.
How do you measure the value of business context in a dataset?
Run one model on the same rows three times: records alone, records plus context the buyer can append, and records plus the supplier's own context. Score each on the buyer's task metric over repeated splits. The value attributable to unique context is the third arm minus the second, with its range.
What counts as context a buyer can rebuild?
Anything the buyer can produce from a key it already receives, from public statistics or from its own systems at an acceptable cost, such as area demographics from a postcode or calendar effects from a date. Write that list down before testing so the result cannot reshape it.
Can announced deal values show what business context is worth?
Rarely. Announced values seldom state which rights, fields or services they bought, and asking prices or public payout ranges are not realized prices. Compare deals only after aligning rights, usable scope, preparation and buyer purpose, then attribute what remains to measured context lift.
Does SourceX publish how it prices context?
No. SourceX does not publish its commercial terms; pricing stays indicative until a Dataset Deal Confirmation fixes the exact dataset, payout and rights for a deal. Investors can request materials, including company and commercial metrics, at sourcex.si/investors.
Sources
- The Rdatasets copy of the ISLR Caravan data: 5,822 customer rows with 86 variables and 348 caravan policy buyers; the file every figure in the context test is computed from. Source
- Rdatasets documentation of the ISLR Caravan data: 5,822 real customer records; sociodemographic variables 1-43 derived from zip codes and shared by everyone in a zip code; product ownership variables 44-86; supplied by Sentient Machine Research for the CoIL Challenge 2000. Source
- The ISLR package's own documentation of Caravan (CRAN mirror, version 1.4) with the same description; its data file matches the Rdatasets copy cell for cell. Source
- The UCI Machine Learning Repository lists the Insurance Company Benchmark (COIL 2000) training file, with sociodemographic attributes derived from zip codes and product ownership attributes, donated by Sentient Machine Research. Source
- Anthropic reported that prepending chunk-specific context before indexing (contextual embeddings and contextual BM25) reduced failed retrievals in its experiments. Source
- NISTIR 8053 describes quasi-identifiers as attributes that do not identify a person alone but can be linked with other data sets to identify data subjects, and that often cannot simply be masked because they carry analytic value. Source
- The kernlab package's documentation of the CoIL 2000 data prints its data dictionary: PPERSAUT is contribution car policies, APERSAUT is number of car policies, and variables starting with M are zip code variables. Source
- SourceX's how-it-works page says pricing stays indicative until a Dataset Deal Confirmation is signed, that the Confirmation fixes the exact dataset, payout and rights, and that SourceX does not publish its commercial terms. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.