# How to Sell Datasets: Where to Sell and What Buyers Check

> To sell a dataset, check rights, de-identify it, add a datasheet and redacted sample, then license it via a marketplace, AI lab, exchange or your own site.

Source: https://sellmydata.com/guides/how-to-sell-datasets/
Published: 2026-10-28 · Updated: 2026-10-28

## Key takeaways

- To sell a dataset, you usually license it rather than hand over ownership, through one of four channels: a self-serve marketplace listing, a direct deal with an AI lab, a managed exchange, or your own site or API.
- The channels differ in upfront effort, reach and how much control you keep over who buys and on what terms.
- A sellable dataset comes with documentation, a clear schema, a redacted sample, agreed license terms and a clean rights trail.
- Before listing, remove people's names, contact details and identifiers, and confirm your contracts and privacy policy allow the use. This is general information; check specifics with counsel.
- There is no single price list to look up. Price depends on data type, volume, years of history and current demand.

To sell a dataset, inventory what you hold, confirm you have the rights, de-identify it, document it with a datasheet and a redacted sample, then license it through a marketplace, a direct lab deal, a managed exchange or your own site. The sections below cover where to sell datasets, how each channel compares, and what turns a raw export into something a buyer can evaluate.

## Can you sell datasets? (the short answer)

Yes. Companies can sell datasets they own or control, usually by licensing them, as long as their customer contracts, privacy commitments and applicable law allow it. That is general information, not legal advice: whether your company can sell a given dataset depends on your specific agreements, so check with counsel before you list.

**Licensing** means granting a buyer permission to use data under agreed terms while you keep ownership. Compared with an outright sale, a license lets you set the scope of use, decide whether the buyer gets exclusive rights, fix a term and agree the format, all per dataset. For a plain-English primer, see [What Is AI Data Licensing?](/guides/what-is-ai-data-licensing/).

### What counts as a business dataset

This guide covers data a company generates in the course of running itself: support tickets, internal messages, code and code review threads, documents, and operational records. It does not cover consumer [personal data](/glossary/pii/), health records tied to people, or payment card data. We don't buy any of those, and they carry a different set of legal questions.

If you're an individual hoping to sell your own personal data, this guide isn't written for you. Our [can I sell my data?](https://sellmydata.com/sell-my-data/) page explains the difference between individual and company data.

## Where to sell datasets: four channels compared

There are four main places to sell datasets: a self-serve marketplace listing, a direct deal with an AI lab, a managed or reviewed exchange, or your own website or API. The right one depends on how much work you can take on and how much control you want over who buys.

| Channel | Who it fits | Upfront effort | Control over buyers | Typical buyers | Main risk |
|---|---|---|---|---|---|
| Self-serve marketplace listing | Teams with a ready, documented dataset and time to handle sales | High: you write docs, set price, vet buyers, handle support | Depends on the platform's terms | Whoever finds the listing | Thin documentation or rights trail; unclear downstream use |
| Direct deal with an AI lab | Companies with legal and sales capacity | High: outreach, negotiation, contract review | Full: you pick the counterparty | AI labs you contact | Time and legal cost fall on you |
| Managed or reviewed exchange | Companies without a dedicated data team | Lower: you export and approve; the intermediary prepares | Sign-off rights, if the exchange grants them | AI labs matched by the exchange | Commission; you must read the exchange's terms |
| Your own site or API | Companies with recurring data products | Highest: sales, billing, delivery, support | Full | Your own customers | Building and maintaining the infrastructure |

### Self-serve marketplace listings

You upload a dataset, write a description and set a price. Reach can be wide, but everything around the file is your job: documentation, the rights review, deciding which buyers are acceptable, and answering questions after the sale. If you go this route, the packaging steps below matter even more, because the listing has to speak for itself.

### Direct deals with AI labs

You approach a lab, agree terms and deliver. This gives you the most control over who uses your data and how. The trade-off is that you carry the whole process: finding the right contact, negotiating scope and price, and running the contract through your own legal review.

### Managed or reviewed exchanges

An intermediary helps clean, de-identify and rights-check the data, then matches it with buyers, usually in return for a commission. As one example of this model, SellMyData charges nothing up front and takes a commission only when a sale closes. The seller signs off on the buyer, the terms and the price; buyers see only a redacted sample before purchase; and the licensed lab receives the [de-identified dataset](https://sellmydata.com/de-identification/), never raw data. Sellers can unlist and delete their data at any time, the target is payment within 7 days of a sale closing, and company exports are [deleted within 30 days](https://sellmydata.com/security/) of the de-identified dataset being approved.

### Your own website or API

Selling directly from your own site suits companies that already run, or want to run, a recurring data product. You'll need the machinery to support it: a way to take payment, deliver files or API access, enforce license terms and support buyers.

For a wider look at turning data into revenue beyond AI training, see [How to Monetize Data](/guides/how-to-monetize-data/).

## Step by step: selling a dataset to an AI company

The short version is: size it, price it, check rights, export, de-identify, document, approve the deal, deliver. The full process is in [How to Sell Your Company's Data to AI Labs](/guides/how-to-sell-company-data-to-ai-labs/).

1. **Inventory what you hold.** List tools, date ranges and rough counts. Our [inventory prompt](https://sellmydata.com/inventory/) can help.
2. **Get an estimate before sharing any data.** You only need counts and dates at this stage, not the data itself. On our [homepage](https://sellmydata.com/) you answer a few questions to get an estimated price range; every request is reviewed by a person, who replies by email.
3. **Run the rights check,** with counsel where needed.
4. **Export only the tools and date ranges you choose.**
5. **Clean and de-identify the data,** then review and approve the result yourself.
6. **Prepare documentation and a redacted sample.**
7. **Approve the buyer, terms and price.**
8. **Deliver the de-identified data and get paid.**

## What makes a dataset sellable to AI companies

A sellable dataset is real-world business content with meaningful volume and history, a stable structure and the noise removed. Our [pricing drivers](https://sellmydata.com/pricing/#how-we-estimate) are data type, volume, years of history and current lab demand, so those are the dimensions worth measuring first.

**Real-world, hard-to-get content.** Support conversations, internal documents, code review threads and operational records reflect how work actually happens. For context, Epoch AI estimates the total effective stock of human-generated public text at roughly 300 trillion tokens (a token is a chunk of text, roughly a word or part of one), with a 90% confidence interval of 100 to 1,000 trillion. Its 80% confidence interval is that this public stock will be fully used at some point between 2026 and 2032 ([Epoch AI](https://epoch.ai/publications/will-we-run-out-of-data-limits-of-llm-scaling-based-on-human-generated-data)). That estimate is about public text only; it is consistent with interest in private business data but doesn't prove it.

**Volume and years of history.** Count what you have by tool and date range. Years of history is one of the price drivers listed above, so an archive that runs back several years is worth documenting precisely.

**Consistency and structure.** Keep fields stable, timestamps intact and threads linked so a reply stays attached to the message it answers. Flattening a conversation into loose text removes context a buyer may want.

**Cleanliness.** Remove duplicates, spam and auto-generated noise such as bot alerts. Researchers found that deduplicating [training data](/glossary/ai-training-data/) produced language models that emitted memorized text ten times less often and needed fewer training steps to reach the same or better accuracy ([Lee et al., ACL 2022](https://aclanthology.org/2022.acl-long.577/)).

To size what you hold, use our volume guides for [Slack](/guides/how-to-check-slack-data-volume/), [Google Drive](/guides/how-to-check-google-drive-data-volume/), [Microsoft Teams](/guides/how-to-check-microsoft-teams-data-volume/) and [OneDrive and SharePoint](/guides/how-to-check-onedrive-sharepoint-data-volume/).

## How to package a dataset for sale

Packaging turns a raw export into something a buyer can evaluate: a datasheet, a schema, a redacted sample and clear license terms. Without these, a buyer has to guess what they're paying for.

### A datasheet or data card

The established framework is *Datasheets for Datasets* by Gebru et al., published in [Communications of the ACM](https://www.microsoft.com/en-us/research/publication/datasheets-for-datasets/) in December 2021. It proposes that every dataset come with a datasheet documenting its motivation, composition, collection process, recommended uses and other information ([CACM](https://cacm.acm.org/research/datasheets-for-datasets/)). The idea is borrowed from electronics, where every component comes with a datasheet describing its operating characteristics, test results and recommended usage. The authors note that, at the time of publication, there were no industry standards for documenting machine-learning datasets ([CACM](https://cacm.acm.org/research/datasheets-for-datasets/)), and argue that datasheets will facilitate better communication between dataset creators and dataset consumers ([arXiv](https://arxiv.org/pdf/1803.09010)).

In practice, cover: what the data is, how and when it was collected, which teams or systems it covers, known gaps, and intended uses.

### A schema and format description

A **schema** is a description of the fields in each record, their types and how records relate to each other. Add the file format, file and record counts, and the exact date ranges included.

### A redacted sample

Give buyers a small sample with identifying details removed so they can check quality and structure before committing. The full dataset should only move after terms are signed.

### License terms to decide up front

- **Scope of use:** what the buyer may do with the data, for example model training only.
- **Format:** how the data will be delivered and structured.
- **[Exclusivity](/glossary/exclusive-license/):** whether you can license the same data to anyone else.
- **Term:** how long the license lasts.
- **Deletion and unlisting:** what happens when the deal ends, and whether you can withdraw the listing.
- **Other terms counsel may suggest,** such as territory.

How these terms are drafted is a legal matter that depends on your agreements, so have counsel review them.

## Rights, privacy and de-identification checks before you list

Before any dataset is listed, confirm you have the right to license it and remove details that identify people. This section is general information, not legal advice; the answer for your company depends on your contracts and the law that applies, so work through it with counsel.

### The rights questions

Work through four areas: customer contracts (do they restrict how you use data customers send you?), your privacy policy (what did you tell people?), employee communications (what were staff told about internal messages?), and third-party content (does the data include material others own?). Our [rights checklist](https://sellmydata.com/rights/) walks through each.

Privacy commitments carry weight. The FTC has said a business that collects user data under one set of privacy commitments cannot unilaterally renege on them afterwards, and that quietly and retroactively changing terms to allow uses such as AI training may be unfair or deceptive ([FTC, February 2024](https://www.ftc.gov/policy/advocacy-research/tech-at-ftc/2024/02/ai-other-companies-quietly-changing-your-terms-service-could-be-unfair-or-deceptive)). Ask counsel how this applies to you.

### De-identification: more than deleting a name column

[De-identification](/glossary/de-identification/), as NIST defines it, is a general term for any process of removing the association between a set of identifying data and the data subject ([NIST SP 800-188](https://csrc.nist.gov/pubs/sp/800/188/final)). In practice that means finding names, email addresses, phone numbers, account IDs and similar identifiers wherever they appear, including inside free text, and replacing them with consistent placeholders so conversations still make sense. NIST's guidance, written for government agencies, also suggests measuring de-identification performance and running re-identification studies to gauge risk.

At SellMyData, de-identified means people's names, contact details and identifiers are removed and replaced with consistent placeholders before anything is listed ([how it works and its limits](https://sellmydata.com/de-identification/)). That is a plain description of the process, not a claim that the result is anonymous in the legal sense. For methods and their limits, see [How to Anonymize Data for AI](/guides/how-to-anonymize-data-for-ai/).

### How regulators view pseudonymised data (any dataset)

This is a general legal point about any dataset, wherever it comes from. The UK ICO says pseudonymous information, data where identifiers are replaced but could be linked back, is still personal data and the law applies to it ([ICO](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-sharing/anonymisation/introduction-to-anonymisation/)). Data counts as effectively anonymised only if no reasonably available means could at any point re-identify the people it refers to ([ICO](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/personal-information-what-is-it/what-is-personal-data/what-is-personal-data/)); see our glossary entry on [anonymization](/glossary/anonymization/). According to one law firm, 'de-identified data' is not defined in UK law, and the ICO discourages the term because it causes confusion ([Taylor Wessing](https://www.taylorwessing.com/en/global-data-hub/2025/spotlight-on-the-uk-data-landscape/gdh---new-ico-guidance-on-anonymisation-and-pseudonymisation)). Ask any provider exactly what their process removes, and ask counsel how your jurisdiction classifies the result.

### Why provenance matters

**Provenance** means where data came from, who made it, and under what licence. A large audit of more than 1,800 text datasets, published in Nature Machine Intelligence in August 2024, found that more than 70% of licences for popular datasets on popular public dataset-sharing platforms were 'unspecified', and described systemic problems with sparse, ambiguous or incorrect licence documentation ([Nature Machine Intelligence](https://www.nature.com/articles/s42256-024-00878-8)).

Copyright is part of the same debate. On May 9, 2025, the US Copyright Office released a pre-publication version of Part 3 of its AI report, on generative AI training and fair use ([Sidley](https://datamatters.sidley.com/2025/05/28/generative-ai-meets-copyright-scrutiny-highlights-from-the-copyright-offices-part-iii-report/)). One law firm's summary says the report treats each use as context-specific rather than endorsing or rejecting fair use outright ([Wiley](https://www.wiley.law/alert-Copyright-Office-Issues-Key-Guidance-on-Fair-Use-in-Generative-AI-Training)), and a legal publisher notes it rejects the view that training is inherently transformative, while recognizing that many uses may be transformative ([Practical Law](https://uk.practicallaw.thomsonreuters.com/w-046-9134?transitionType=Default&contextData=%28sc.Default%29)). The report also considers licensing options ([Policy Commons](https://policycommons.net/artifacts/33446392/copyright-and-artificial-intelligence-part-3/34346062/)). A dataset with a clear, documented rights trail is easier for a buyer to evaluate. None of this is legal advice; counsel can tell you how it affects your data.

## How much can you sell a dataset for?

There's no single price list to look up. We don't publish one: price depends on data type, volume, years of history and current lab demand. Our homepage calculator gives an indicative payout range based on headcount today and at peak, years of history, data types and industry, calibrated to reported deals and client quotes as of October 2026 ([how we estimate](https://sellmydata.com/pricing/#how-we-estimate)). That range assumes you keep your full history and license it exclusively, and it is not an offer.

Two levers are worth thinking about before any negotiation. First, history: years of history is one of the four price drivers, so know exactly how far back your archive runs and how complete it is. Second, exclusivity and term: these are rights you give up, so decide what they're worth to you before agreeing to them.

## Common mistakes when selling datasets

The mistakes below come from skipping a check or a document. Watch for these:

- **Listing before checking contracts or the privacy policy.** Your earlier commitments still apply, as the FTC guidance above makes clear. Review them with counsel first.
- **Assuming that deleting a name column equals de-identification.** Identifiers hide in free text, signatures and email threads, and the ICO is clear that data which can still be re-identified hasn't been anonymised.
- **Giving away exclusivity or perpetual terms without pricing them in.** Decide what exclusivity and term are worth to you before you agree to them, and have counsel review the terms.
- **Shipping without documentation.** A buyer who can't see what's in a dataset, how it was collected or what's missing has little to base a price on.
- **Losing the ability to unlist or delete.** Whichever channel you use, confirm in writing that you can withdraw the listing and have your data deleted.

## Frequently asked questions

### How do I sell datasets?

Inventory what you hold, confirm you have the rights, de-identify the data, document it with a datasheet and redacted sample, then license it through a marketplace, a direct lab deal, a managed exchange or your own site. Agree scope, format, exclusivity and term before anything is delivered.

### Where can I sell datasets?

The four main channels are self-serve marketplace listings, direct deals with AI labs, managed or reviewed exchanges, and your own website or API. Pick based on how much work you can take on yourself and how much control you want over buyers.

### Can you sell datasets online?

Yes, companies can license business datasets online if their contracts, privacy commitments and applicable law allow it. This is general information, not legal advice, so check your situation with counsel.

### How do I sell my dataset to an AI company?

Prepare a de-identified, documented dataset with a redacted sample, then either approach labs directly or work through an exchange that matches you with them. Our full walkthrough is at /guides/how-to-sell-company-data-to-ai-labs.

### Is selling data profitable?

It can be, but there is no published price list: price depends on data type, volume, years of history and current lab demand. On SellMyData, sellers pay nothing up front and a commission is taken only when a sale closes.

### Do I sell the dataset outright or license it?

Licensing lets you keep ownership while granting a buyer permission to use the data under set terms such as scope, format, exclusivity and term. Which structure suits you depends on your agreements, so review the terms with counsel.

### Can I sell a dataset that contains customer or employee information?

Only after checking customer contracts, your privacy policy and employee communications, and removing names, contact details and identifiers. Whether it is permitted depends on your commitments and the law that applies, so confirm with counsel first.

### What makes a dataset sellable to AI companies?

A sellable dataset is real-world business content with meaningful volume and history, a stable structure and the noise removed. SellMyData's pricing drivers are data type, volume, years of history and current lab demand, so those are the dimensions worth measuring first.

### How do I package a dataset for sale?

Packaging turns a raw export into something a buyer can evaluate: a datasheet, a schema, a redacted sample and clear license terms. Without these, a buyer has to guess what they're paying for.

### What should a dataset datasheet include?

In practice, a datasheet covers what the data is, how and when it was collected, which teams or systems it covers, known gaps, and intended uses. The established framework is Datasheets for Datasets by Gebru et al.

## Sources

- [Datasheets for Datasets](https://cacm.acm.org/research/datasheets-for-datasets/)
- [Datasheets for Datasets (publication page)](https://www.microsoft.com/en-us/research/publication/datasheets-for-datasets/)
- [Datasheets for Datasets (arXiv 1803.09010)](https://arxiv.org/pdf/1803.09010)
- [Generative AI Meets Copyright Scrutiny: Highlights from the Copyright Office's Part III Report](https://datamatters.sidley.com/2025/05/28/generative-ai-meets-copyright-scrutiny-highlights-from-the-copyright-offices-part-iii-report/)
- [Copyright Office Issues Key Guidance on Fair Use in Generative AI Training](https://www.wiley.law/alert-Copyright-Office-Issues-Key-Guidance-on-Fair-Use-in-Generative-AI-Training)
- [Copyright Office Part 3 report summary (Practical Law)](https://uk.practicallaw.thomsonreuters.com/w-046-9134?transitionType=Default&contextData=%28sc.Default%29)
- [Copyright and Artificial Intelligence, Part 3](https://policycommons.net/artifacts/33446392/copyright-and-artificial-intelligence-part-3/34346062/)
- [A large-scale audit of dataset licensing and attribution in AI](https://www.nature.com/articles/s42256-024-00878-8)
- [AI (and other) Companies: Quietly Changing Your Terms of Service Could Be Unfair or Deceptive](https://www.ftc.gov/policy/advocacy-research/tech-at-ftc/2024/02/ai-other-companies-quietly-changing-your-terms-service-could-be-unfair-or-deceptive)
- [NIST SP 800-188: De-Identifying Government Datasets: Techniques and Governance](https://csrc.nist.gov/pubs/sp/800/188/final)
- [Introduction to anonymisation](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-sharing/anonymisation/introduction-to-anonymisation/)
- [What is personal data?](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/personal-information-what-is-it/what-is-personal-data/what-is-personal-data/)
- [New ICO guidance on anonymisation and pseudonymisation](https://www.taylorwessing.com/en/global-data-hub/2025/spotlight-on-the-uk-data-landscape/gdh---new-ico-guidance-on-anonymisation-and-pseudonymisation)
- [Deduplicating Training Data Makes Language Models Better](https://aclanthology.org/2022.acl-long.577/)
- [Will we run out of data? Limits of LLM scaling based on human-generated data](https://epoch.ai/publications/will-we-run-out-of-data-limits-of-llm-scaling-based-on-human-generated-data)
- [What is checked before data is listed](https://sellmydata.com/rights/)
- [De-identification](https://sellmydata.com/de-identification/)
- [How we estimate](https://sellmydata.com/pricing/#how-we-estimate)
- [Security](https://sellmydata.com/security/)
- [Can I sell my data?](https://sellmydata.com/sell-my-data/)
- [Inventory prompt](https://sellmydata.com/inventory/)
- [SellMyData homepage and estimate](https://sellmydata.com/)