Does an AI Visibility Vendor Use Customer Prompts to Train Its Models?
Whether an AI visibility vendor uses your prompts to train its models depends on its terms, and you need to check two layers. The first is the vendor itself. Some use customer prompts and results only to provide the service. Check whether the terms also allow other uses, such as building benchmarks, prompt-volume databases or the vendor's own classification models. The second layer is the AI providers the vendor queries. Business API terms from major model providers generally exclude training on customer data by default. Consumer apps can be different, depending on account settings, and some vendors collect answers through those apps. Prompts can reveal competitive strategy, product plans and target markets, so ask the vendor directly and get the answer in the contract.
What are the two layers?
| Layer | What can happen to your prompts | What to check |
|---|---|---|
| The AI visibility vendor | Used to run your tracking; possibly also for aggregated benchmarks, prompt databases, product improvement or training its own models | Terms of service, privacy policy, DPA, any "aggregated data" clause |
| AI providers queried by the vendor | Processed to generate answers; retention and training depend on the provider's terms and whether business APIs or consumer apps are used | Which providers and access methods the vendor uses; those providers' data policies |
How do AI providers treat data sent through their APIs?
Major model providers' business API terms generally state that customer inputs and outputs are not used to train their models by default. Retention periods vary, for example for abuse monitoring. Some providers treat free tiers differently from paid ones. Google, for instance, distinguishes between unpaid and paid use of the Gemini API in how data may be used to improve its products. Consumer chat apps are governed by consumer terms and account settings, which can allow model improvement unless the user opts out. Terms change, so check each provider's current policy.
This matters because vendors collect answers in different ways. A vendor using business APIs under standard terms is in a different position from one automating consumer accounts. Ask which method is used for each engine.
Why can prompts be sensitive?
Your tracked questions can reveal:
- which competitors you are targeting;
- new products, markets or features before launch;
- positioning you are testing;
- weaknesses you are worried about, such as questions about complaints or recalls;
- named individuals, which raises data protection questions.
A prompt database built from many customers' questions can also reveal market-level intelligence. It is reasonable to want to know whether your questions contribute to it.
What should you ask the vendor?
- Do you use customer prompts, answers or results for anything other than providing our service?
- Do you train or fine-tune any models, including classifiers, on customer data?
- Do you publish benchmarks or sell prompt or market data that includes our data, even aggregated or anonymised?
- Which AI providers do you query, and through business APIs or consumer apps?
- What are those providers' retention and training terms for the method you use?
- How long do you retain our prompts and answers, and can we delete them?
- Can we opt out of any secondary use?
A vendor that answers these clearly, and puts the answers in writing, is easier to trust than one that points to a general privacy policy.
What contract terms should you look for?
- A clause limiting use of customer data to providing the service.
- An explicit statement on whether aggregated or anonymised data may be used, and an opt-out if so.
- A list of sub-processors, including AI providers.
- Retention periods and deletion on request or at contract end.
- Confidentiality covering prompts, answers and reports.
- Notice of material changes to data use.
Where personal data is involved, the Data Processing Agreement should cover these points too; see whether an AI visibility platform needs a Data Processing Agreement for EU customers.
How do you reduce exposure in practice?
- Avoid confidential product names in tracked prompts until launch, or track generic category questions instead.
- Avoid naming private individuals unless necessary.
- Separate sensitive tracking, such as crisis or legal questions, into a tool or process with stricter terms.
- Review prompt sets before uploading them, as you would any shared document.
How to choose which prompts to track in the first place is covered in how to choose the right prompts to track for AI visibility.
What do people ask about vendors and prompt data?
Is it normal for vendors to use aggregated data?
Many software vendors use aggregated, anonymised data for benchmarks and product improvement. What matters is that it is disclosed and that you can opt out if needed.
Can a model provider learn our strategy from API calls?
Under standard business API terms, data is generally not used to train models by default. Check retention terms and the access method the vendor uses.
Does this apply to answers as well as prompts?
Yes. Collected answers and reports are your data too, and the same questions apply.
What if the vendor will not answer?
Treat that as a risk signal, and limit what you track with them until you get clear terms.
Published by The Kovmere Team · updated Sep 29, 2026 · More resources