As an AI answer engine, Perplexity processes your queries (including everything you paste into them), uploaded files and images, your interaction history, and account/usage metadata. Because it synthesizes answers using third-party models, your query content can travel beyond Perplexity's own infrastructure.
That routing matters. A prompt containing a customer's contract terms isn't just sitting in Perplexity's logs — it may be processed by an upstream model provider too, per their arrangements.
This is the crux of Perplexity data privacy, and it's the opposite default from what security teams assume:
Read that top row again. Every employee using consumer Perplexity for work — the overwhelming majority, since it's free and requires no procurement — is contributing their queries to model improvement unless they personally found and flipped a setting. That's not a control you can rely on.
The way people use Perplexity makes this worse than a generic chatbot. Users paste context to get better answers:
Each is a legitimate workflow. Each puts confidential data into a query that, on consumer tiers, is retained and used for training by default.
Three layers, in order of reliability:
Strac does exactly that in the browser across Chrome, Edge, Firefox, and Safari — covering Perplexity alongside ChatGPT, Claude, Gemini, and Copilot — and logs every event as compliance evidence for SOC 2, HIPAA, GDPR, and PCI. See AI DLP.
A settings toggle is not a control. These four are:
On the Free, Pro, and Max plans, yes — AI training is enabled by default, and you must opt out by turning off the AI data retention setting in your account. On Enterprise Pro and Max, your data is never used to train models, and the Sonar API operates with zero data retention. Most employees are on consumer plans, so the default applies to them.
Uploaded files and images are typically deleted after around 30 days on consumer plans and about 7 days for enterprise users, with larger enterprise organizations able to configure custom retention periods. The Sonar API deletes prompts and responses immediately after processing.
Not on consumer tiers as configured out of the box — training is on by default and your query context can be routed to third-party models. Enterprise tightens this considerably. But in every tier, the confidential data you type still leaves your environment, which is why the reliable safeguard is redacting it before submission rather than relying on plan settings.
Enterprise Pro/Max removes the training concern and shortens retention, which is the right foundation. It does not, however, prevent an employee from pasting a patient record or a customer's card number into a query. For regulated data, pair it with a control that inspects and redacts queries at the point of entry.
Technically, not by policy. Detect sensitive data in the query as it's typed and redact, warn, or block based on your rules. This works even when employees are on personal consumer accounts you don't manage — which, for Perplexity, is the majority of real-world usage.
Perplexity data privacy has a default most organizations would never consciously choose: consumer plans train on your queries unless someone opts out. Enterprise fixes that, but neither tier stops the actual leak — the confidential context employees paste in to get a better answer. Put the control in the query path, and Perplexity stays the research tool people love without becoming a data-exfiltration channel.
Book a demo to see sensitive data redacted from Perplexity queries in real time.
Related: AI data privacy · is Perplexity safe? · ChatGPT data privacy · Cursor data privacy · AI DLP
.avif){kind=icon}
.gif)
.webp)
