Knowing that Notion AI "searches your workspace" doesn't tell you much. Knowing how it searches lets you decide whether to trust the answer. Here's the real mechanism, as Notion documents it.
Step one: your pages become embeddings
Before it can answer anything, Notion AI needs a way to figure out which pages are relevant to a given question. To do that, it creates an "embedding" for every page: a numerical representation of what the page means, not just the words it contains. Pages covering similar topics end up with similar numerical representations, which lets search work by meaning rather than exact keyword matching.
These embeddings are generated through a zero-retention embeddings API from OpenAI, then stored in a vector database (Notion uses Turbopuffer, among others) built specifically for fast similarity lookups. That vector database has been reviewed by Notion's own security team and by an outside auditor, and holds a SOC 2 Type 2 certification.
Worth stating plainly: embeddings get the same contractual protection as the rest of your data. Whatever Notion's customer contracts say about handling your data applies to embeddings too.
Step two: how an answer actually gets generated
Once embeddings exist, here's what happens when you ask Notion AI something:
- Your request comes in.
- If it's simple enough to not need a workspace search, an LLM answers directly.
- If a search is needed, an LLM first turns your question into a search query.
- That query goes to the vector database, which returns a list of pages ranked by relevance.
- Those pages go back to the LLM, which reviews and re-ranks them by how well each one actually answers your question.
- The LLM generates an answer from the refined, ranked list of pages.
- Notion formats the result and shows it to you.
Two things stand out. First, this is a real search process, not a black box: the AI finds actual pages in your workspace and writes from them, it doesn't invent an answer from general training knowledge. Second, several rounds of ranking and refining happen before you see anything, which is why answer quality depends so heavily on whether the right page existed in the first place.
Who's actually "thinking"
Notion names its AI providers directly rather than hiding behind a line like "our proprietary AI": it uses large language models hosted by Notion itself, plus models from Anthropic and OpenAI, and says it continuously evaluates these providers to keep quality high.
Every third party that processes your data is listed on Notion's public subprocessor list, and you can subscribe by email to be notified the moment a new subprocessor is added, before it starts processing your data.
The fact that matters most: permissions are respected
Here's the detail worth remembering above everything else in this guide: Notion AI respects your existing permissions. The models generating an answer for a given user can't see or use any information that user doesn't already have access to in Notion.
That means if someone on your workspace can't open your HR folder, Notion AI won't surface HR content to them, not even indirectly, even if the answer to their question technically lives on a page they can't open. This isn't a setting you have to configure or a promise you have to take on faith: it's built into how the search step itself works.
Data separation and encryption
Two mechanical facts worth keeping in mind:
- Your data never gets mixed with any other customer's during AI processing. Each customer's accounts stay separated within Notion's production environment.
- Data sent to third parties is encrypted in transit, using TLS 1.2 or higher, the same baseline standard described in Notion's general security practices.
Who owns what gets generated
Notion claims no rights or license over your content just because you used AI on it. What you type in (your prompt) and what comes out (the AI's response) are both treated as your own "customer data" under Notion's terms, exactly like anything else you type on a Notion page.
Article 2 of 4 · See all articles in this guide →