Rightmark vs. Snowflake Marketplace: a 2026 head-to-head comparison

Compare the result the buyer needs with the work the owner can sustain.

Re-check: Confirm current pricing, fee allocation, billing units, permissions and delivery terms with each provider before signing.

A convenient purchase can still leave the buyer with the wrong deliverable or the owner with work it cannot support. The comparison starts with usefulness and money, then follows the rights, approvals and operating responsibilities through to the evidence needed for a decision.

1. Will the data fit the use you can actually license?

A search result can answer a question brilliantly without giving a training team the collection it needs. A sample can look convincing while leaving out the dates, document types or difficult cases that matter most. Owners face the same problem from the other side: preparing an entire archive makes little sense if the buyer only needs a narrow, regularly updated service.

Consider a hypothetical support-history training project. The team wants examples connecting a customer's problem, the attempted fix and the eventual outcome. This is a planning example, not a Rightmark dataset or customer result. If the sample contains polished answers but loses which fix failed before the successful one, reject it for this task. More answers will not repair the missing sequence. If de-identification breaks the links between turns, ask for a corrected sample before negotiating volume.

The same material could still be useful for answering a narrow support question. A relevant retrieved passage can serve that purpose without supplying the connected training examples above. That is why a convincing question-answer demonstration cannot stand in for the training team's sample test. Neither task is intrinsically more valuable; each needs different evidence.

For retrieval, write down questions whose answers you know should be present, questions spanning several documents, and questions the collection should not answer. Keep the expected source beside each question. Snowflake's provider tutorial includes document identifiers and source URLs in its example, providing a concrete starting point for inspecting what a returned passage represents. A fluent response alone is a poor acceptance test.

Bottom line: For the support-history training task, choose the offer that supplies usable, connected examples with training permission. Reject a collection that loses the sequence, even if its individual answers look excellent. For a retrieval task, accurate answers traceable to the right source can be sufficient; acquiring a full archive would add work the project may not need.

Ask both: “Show us what we would actually receive, the use we may make of it, and how we would reject a delivery that misses the agreed scope.” “AI-ready” without those details leaves the decision unfinished.

For your own acceptance record, note the agreed test, the sample result, the delivery result, who decides whether it passes, and the correction deadline. These are proposed planning fields; agree the test and any remedy with the supplier before delivery.

2. What will the buyer spend, and what will the owner keep?

An owner's payout and a buyer's invoice answer different questions. Moving a fee to the buyer can preserve the owner's agreed amount while increasing the buyer's total. Compare the same collection, permitted use, term and delivered form, then give each side its own calculation.

For a paid data listing combining a monthly fee and per-query charge, calculate listing charge = min(monthly cap, applicable monthly fee + per-query price × billable queries). Zero queries mean no usage-plan monthly fee; the first query is charged before any free allowance. Term subscriptions charge upfront. These are selected data/application plan rules, not CKE pricing. Snowflake service charges remain separate.

The owner's operating budget deserves equal attention. Snowflake's Cortex Search cost guide says a running search service incurs costs without queries, alongside warehouse, embedding and storage costs. It also describes suspending serving during idle periods. An owner should budget a quiet month with the service running and compare it with a suspension plan. The buyer, meanwhile, must decide whether interruptions fit the application. A lower operating bill is no advantage if the product cannot answer when needed.

Use actual quoted inputs in a blank comparison:

• Owner receives: contracted payment base ___; deductions ___; net payment ___; separately funded operating work ___.
• Buyer spends: license or access charge ___; intermediary fee ___; preparation and transfer ___; platform and ingestion costs ___; taxes and other agreed charges ___; total ___.
• Both confirm: currency ___; permitted use ___; term ___; delivered form ___; renewal ___; rejection/remedy terms ___.

For a Rightmark standard license, begin with the agreed owner payout and buyer fee, then add only the preparation or transfer costs allocated to the buyer in writing. Count a bundled charge once. Confirm the contractual percentage base for premium and bespoke work and any other route-specific deductions before applying the percentage. This blank aid is a planning worksheet, not a quote.

Bottom line: For a standard Rightmark license, the owner can compare the agreed payout directly while the buyer budgets the added fee. For an eligible Snowflake buyer, usable capacity funds can influence procurement. Choose on the resulting buyer total and owner net for the same job. A favorable fee split cannot compensate an owner for unbudgeted work or a buyer for the wrong deliverable.

Ask both: “For this collection and use, give us the buyer's full cost and the owner's net, including what changes with usage, renewal or a rejected delivery.” A percentage without its base is not a complete answer.

3. Which rights and exposures does each side take on?

Keeping ownership and granting a license can coexist. Keeping ownership and promising exclusivity can also coexist, which makes a simple “you keep your data” answer incomplete. The owner needs to understand which future uses remain possible; the buyer needs permission for the work it actually intends to do.

Snowflake's Standard Agreement applies to products listed as subject to it. Do not read one contract and assume it governs the whole marketplace. Equally, do not assume that a queryable product forbids every external use, or that receiving a batch permits every AI use. Put the intended activity beside the clause that authorizes it.

For example, separate retrieval, evaluation, model training and redistribution in the request. Then name the artifacts the project will retain: source records, transformed records, derived datasets and trained models. Ask counsel to resolve any gap between that list and the proposed grant. This makes the discussion specific enough for an owner to identify material it must exclude, and for a buyer to stop a project whose essential use has no permission.

Rightmark's license summary, still labeled “Draft for counsel review,” says some code routes can permit a licensing partner to sublicense, take an own-use license or build derived datasets, disclosed before signing. That is a reason to read the selected route carefully, rather than infer a blanket resale restriction from the Rightmark name.

Bottom line: An owner who wants to license the same code again should examine Rightmark's non-exclusive route; an archive owner considering an exclusive lump sum must weigh the payment against the uses reserved for that term. For the hypothetical training project, hold the purchase if the grant omits training or the artifacts the team needs to retain. Technical access cannot supply that missing permission.

Ask both: “Which clauses cover our intended use, derived artifacts, external transfer and onward access, and what restrictions remain on the owner?” A general assurance that the data is licensed does not answer those questions.

Before granting the license, ask which recipients will be disclosed, whether any recipients can be excluded, and what the agreement permits them to retain as trained models or derived datasets.

4. What can you approve, restrict and withdraw?

Signing a license, disconnecting a system and ending an existing buyer's rights are separate events. An owner needs to know what its initial approval authorizes later. A buyer needs enough clarity to plan for an expired term or interrupted access without confusing a technical change with the contract's consequences.

Rightmark's connection control stops access into the owner's systems; it is not a recall of delivered batches. Snowflake's listing modification guide describes removing the consumer from the share. Unpublishing alone only hides the listing; existing access continues. These controls act at different points in delivery. Neither action, by itself, answers what the license allows a buyer to retain.

Build a simple exit record before signing. Name the person who can revoke access, the notice that must be sent, the rights that end, the materials to delete and any certification required. Then run through two scenarios: an owner stops contributing new material, and an existing license expires. If the answers are identical without checking the agreement, the discussion is probably skipping a relevant distinction.

Bottom line: If an owner needs to approve every later code sale, Rightmark's described signing model does not promise that control. For a Snowflake share, an owner who wants to stop current access must remove the consumer, not merely unpublish. A buyer should separately settle what happens to material already used. Do not build a termination plan around a button that changes the wrong permission.

Ask both: “What do we approve now, what can happen under that approval later, and what exactly stops or survives when access or the term ends?”

For the selected route, ask whether the owner can approve recipients and which retention and deletion duties survive expiry.

5. How will the data reach the buyer's working environment?

A correct license can still buy the wrong form of access. The team that needs to run its own evaluation over complete records has a different requirement from a team that needs relevant passages returned to an application. Agree the delivery test early enough that the owner knows what it must prepare.

Snowflake's listing creation guide also allows a provider to attach a notebook containing completed SQL or Python examples and their results. Such notebooks are view-only. They can help a buyer see how the provider expects the product to be used. Ask for a demonstration against your own requirement as well: a provider's successful example does not establish that your joins, filters or retrieval questions will work.

For a batch requirement, agree an ingestible format and the metadata needed to interpret it. A file that opens successfully may still lack the identifiers needed to connect related records or the date information needed to exclude a period. For an ongoing service, identify who will warn the application owner about a changed field, unavailable source or paused feed. These are delivery questions worth resolving even when the first demonstration succeeds.

Bottom line: In the hypothetical training project, require dates that let the team enforce its chosen cutoff and identifiers that survive ingestion. If the delivery strips those fields, an otherwise readable sample can fail the actual pipeline. Test the intended format before agreeing the full delivery. For retrieval, perform the corresponding check in the application where users will ask questions, not only in the provider's demonstration.

Ask both: “Demonstrate the delivered form in our intended environment, identify who operates changes, and confirm the location, transfer and use conditions.”

Before choosing a route, confirm the supported locations and who performs ingestion, then put any refresh commitment in the scope.

6. Who prepares, operates and supports the arrangement?

A marketplace gives a product somewhere to be offered; an intermediary gives a private requirement somewhere to be scoped. Neither description assigns every operational task. Before choosing on convenience, find out who does the work when a scan fails, a field changes or a buyer disputes what arrived.

Snowflake's CKE tutorial loads and chunks documents, creates search, then tests from a consumer account. For the hypothetical support archive, identify who investigates a wrong resolution: missing source material, a bad link and unsuccessful search need different fixes. The buyer needs someone who can investigate, rather than repeat the demonstration.

For shared data, Snowflake explicitly lists adding or removing columns, renaming objects and removing objects as examples of potentially breaking changes that require consumer notification. That is an operating commitment to consider before offering a changing dataset. Buyers should identify the person who receives those notices and decides whether a change requires another acceptance test.

Use the blank buyer request brief to describe the common specification. It is a planning aid, not a quote or confirmation of availability. Then assign each task:

This is a planning worksheet, not a service commitment. Leave a field unresolved until someone accepts it; putting a company name in every row is not the same as assigning an accountable person and an agreed result.

Bottom line: An owner offering a changing Snowflake share must be able to notify consumers when changes could break their work. An owner considering a defined Rightmark batch should negotiate the delivery cadence explicitly. In the support-history example, a one-time archive can fit a fixed training exercise; a service used for current support answers needs someone accountable for keeping it useful. Choose the obligation the owner can actually sustain, and give the buyer a named person to contact when it fails.

Ask both: “Who is accountable for rights, preparation, delivery, refresh, acceptance and exceptions, and which tasks stay with our team?”

7. Which evidence would justify choosing this route?

The strongest demonstration is one that can fail in a way your team recognizes. A polished answer or a large record count can hide the defect that matters to the job. Before the demonstration, write down what would make you buy, reject or ask for a corrected sample. Owners benefit too: this prevents an endless evaluation whose requirements change after every delivery.

A Snowflake listing notebook can show an executed example's tables and visualizations. Use that demonstration to select tests, then run the tests the actual project requires. In the support-history example, inspect a case where the first fix failed. Can the buyer reconstruct that failure and the eventual resolution without re-identifying the person? If the sample omits the unsuccessful step, the apparent quality of the final answer masks the missing training signal.

For a retrieval application, test a question the collection should answer and one it should not. Inspect the returned source rather than only the wording of the answer. A plausible answer tied to the wrong document is a failed evidence test. If the underlying records are correct but the retrieval fails, ask for a corrected demonstration; buying more source material may not solve that problem. These are proposed tests for the hypothetical project, not measured results from either provider.

An owner should judge a comparable sale by the work and obligations behind its proceeds. A one-time archive payment is not a useful earnings benchmark for a service that needs continuing maintenance. A buyer should distinguish an evaluation result from a training result: success at finding a document does not demonstrate that a model learned from a licensed collection. Ask for the test conditions and the relevant costs before treating someone else's result as a forecast.

Bottom line: Turn the successful sample test into a delivery test. If the received records lose the required sequence or the retrieval result points to the wrong source, document that failure and use the agreed correction process. An earlier demonstration does not make a failed delivery acceptable. This gives the owner a stable target and the buyer a reason to accept or reject beyond how impressive the demonstration looked.

Ask both: “Show the test that would expose a poor fit for our project, the result it produced, and what happens if the delivered product fails that same test.”

These five details explain why a buyer or owner might choose to discuss a private collection with Rightmark. They are reasons to evaluate the service, rather than claims that no other provider can offer them.

• Start with the missing collection. The buyer workflow describes sourcing owners who have not listed their data, checking title and consent, and delivering graded batches to a specification. That is worth exploring when an available product does not cover the job. Supply and timing still need to be established for the request.
• Make the rights packet inspectable. The ownership pack identifies the licensor, signer, scope, rights and privacy evidence, and technical checks without opening the raw archive. It helps an owner assemble the material a buyer needs to review; it does not clear the collection legally.
• Connect the sample to written delivery requirements. Rightmark describes a written scope, representative scrubbed evaluation samples under NDA, graded batches with metadata and another scan on receipt. Agree the acceptance condition and exception procedure for your purchase.
• Separate the standard owner payout from the buyer fee. Rightmark takes nothing from the owner's payout on standard licenses and adds its fee to the buyer's side. Premium and bespoke deals use a separately agreed owner share of 75% to 85%. Compare actual totals before treating either structure as more favorable.
• Let an owner start without files. The appraisal asks about the asset and rights before requiring files or system access. Use its indicative result to decide what to investigate, not as evidence that a buyer will make an offer.

Choose Snowflake when the required database objects are available through a suitable share. CKEs offer licensed retrieval in a Snowflake account without delivering the whole corpus. That can suit question answering better than preparing an archive. Test whether the answers serve the application before choosing.

For an owner, changes to shared data carry consumer-notification duties. A running search service also needs a budget during quiet periods. Choose that obligation only if the owner can fund operation over the agreed term. A buyer relying on changing information should value that commitment explicitly, rather than compare it with a one-time archive on price alone.

The Marketplace Capacity Drawdown program gives eligible, opted-in customers another reason to evaluate a matching offer: qualifying partner purchases can draw on committed Snowflake funds. Treat that as a procurement advantage to investigate, not a discount or a reason to overlook an unsuitable license.

Choose a Rightmark discussion when the buyer needs a private collection sourced to its specification. Families, filters, formats, date coverage, de-identification, volume and delivery cadence belong in the written scope; training or evaluation permission belongs in the license. A clear requirement helps an owner decide whether it has something appropriate to license before work begins.

For an owner with private code or a company archive, the first useful preparation is an honest account of what it controls and what must be excluded. Code is non-exclusive by default, while company-archive lump sums are usually exclusive for a term, with non-exclusive revenue share as an alternative. Code signing can authorize later non-exclusive sales. Those differences should influence the owner's choice as much as the proposed payment.

Rightmark's sample review and specified batches give the buyer a way to inspect the collection before accepting delivery. For the hypothetical training job, that means checking linked problem, fix and outcome records rather than accepting an answer-only demonstration. An owner can use the same requirement to exclude unsuitable material before preparation begins. Agree who corrects a rejected batch and whether the job ends after acceptance or includes future updates.

Confirm the minimum usable volume and readiness requirements for this request, then agree any correction and update duties before preparation begins.

Consider a third option if the rights holder is already known

A direct license deserves comparison when the buyer knows the owner of the exact collection and both can organize the contract, preparation, delivery and support themselves. Use the same scope and acceptance questions. Include the time and work that would otherwise sit with a platform or intermediary before comparing cost.

Start with the authorized signer, relevant rights evidence, permitted use and delivered form. If the evidence cannot support the use or the collection cannot pass the agreed test, narrow or hold the request. Changing the route does not cure that gap.

Relevant reading: Rightmark vs. Defined.ai compares another licensing route. Keep the same collection and intended use when extending the shortlist.

For the support-history training project, the winning offer supplies connected examples, permission to train and a delivery the team can actually ingest. If those conditions fail, a convenient purchase or a larger collection does not rescue the job. For question answering, correct passages tied to the right source may be the entire deliverable. Do not buy a different project by accident.

Owners should make the corresponding choice about work: prepare a defined collection, or support something buyers will keep relying on. Put the net payment beside that obligation. Buyers should put the full cost beside the acceptance test. Then the choice turns on a useful result each party can recognize, rather than the platform name.

Have a private-data requirement to scope?

Send your data specification to Rightmark to discuss what can be sourced, in what volume and on what timing. Scope and sample timing are confirmed in writing; representative samples are for evaluation only.

Request data