Notion AI · Practical walkthrough

Ask Notion AI a workspace question and check where the answer came from

Someone asks what the agreed refund window is for trade customers. You know it's written down somewhere in Notion, probably twice, probably differently.

Task by Task14 min read + practice

What you'll make

a checked answer to one workspace question, with every cited page opened and judged.

What you'll need

a Notion workspace on the Business or Enterprise plan, where Notion AI and Enterprise Search are available, one question that has a written answer somewhere, and about ten minutes. Answers respect your existing permissions, so you’ll only see content you could already open.

Example result.

Refund window, trade customers: 30 days from delivery.
Governing page: Trade terms (current), updated July (changed from 14 days).
Also cited: Trade onboarding checklist says "refunds within 30 days" (no start point).
Action: point the checklist at Trade terms instead of restating the rule.

1. Pick a question that has a right answer

The method works on questions where a correct answer exists in writing and your problem is finding it. “What did we agree the refund window is for trade customers?” is a good question. “How should we think about refunds?” is not: it asks for advice, so a cited statement of policy may not answer it.

Before you type anything, write down what a usable answer contains. For a policy question, that’s usually the rule, the date it was agreed or last changed, and the page meant to be authoritative.

Why this works: you now have something to check the answer against, instead of judging it on how confident it sounds.

What you should see: one specific question and a three-item list of what the answer must include.

2. Open the search and set the scope

Open the Home tab and type your question into the Enterprise Search window. With the Notion desktop app installed, Shift + Cmd/Ctrl + J opens Notion AI from anywhere. Desktop also has Command Search for asking from outside the app; change its shortcut at Settings › Preferences › Keyboard shortcut for Command Search.

By default the search looks at your workspace, any connected apps and the web. For your own agreed policy, the web is noise at best and contamination at worst. Narrow it:

Select All sources to change the scope.
Turn Web search off.
To search one source only — your workspace, say, or only Slack — select its name in the menu. Hover over the workspace name to narrow further to one page or teamspace.
Use Add context, or @-mention pages, teamspaces and people, to point it at particular material.

There’s a model picker set to Auto. Notion warns that some models may look only at the web and not at your workspace or connected apps, so leave it on Auto unless you have a reason to change it.

Why this works: a plausible industry-standard answer can’t arrive dressed as your policy if the web isn’t in scope.

What you should see: the scope showing only your workspace (or one teamspace), with web search off.

3. Ask it in a way that makes a gap visible

Paste this, with your own question in place of the example:

What did we agree the refund window is for trade customers?

Answer only from my Notion workspace. Give the figure or rule, the page it comes from, and the effective or approval date if stated. Include the last-updated date only if available; do not invent metadata.

If more than one page states a different answer, list each one separately with its page and date rather than reconciling them. If no page states this directly, say so and list the closest pages you found instead of inferring an answer.

Why this works: asking for each conflicting page separately stops the model smoothing two contradictory pages into one tidy sentence that matches neither. Allowing “no page states this” gives it a way out other than inventing something. It doesn’t guarantee a true answer; it makes a false one easier to catch.

What you should see: a rule, one or more citations, and either separate entries for conflicting pages or a plain statement that nothing was found.

4. Open every citation

Notion says Enterprise Search always cites its sources when it answers from your workspace or a connected app. That’s the whole reason this is worth doing: an answer without a source is a rumour with good grammar. Click into each cited page and ask three questions.

Does the passage support the claim? Check the whole condition: who it applies to, when the clock starts, any exception.
Is this the page meant to govern? A workspace usually holds the policy page, a draft of it, a project note quoting it and somebody’s meeting notes disagreeing with it. All four are honest sources. Only one of them governs.
When was it last updated? Notion AI can cite a page from two years ago with the same composure it brings to yesterday’s.

Why this works: citations prove the answer came from somewhere, not that it came from the right place or kept the right conditions.

What you should see: for each citation, a yes or no on support, authority and age, and the one page you’d quote.

Check the result

The rule in the answer matches the governing page word for word on its conditions (who, when the clock starts, exceptions).
Every conflicting page appears separately, with its date.
Nothing in the answer comes from a page you didn’t open.

The usual failure: a confident single answer that drops a condition, because the model merged a governing page with a page that only repeats it. You spot it when the cited pages don’t say quite the same thing, or when one citation is a checklist or meeting note rather than the policy itself.

Ignore summary pages, checklists and meeting notes. Answer only from pages in the [teamspace] teamspace, quote the exact sentence that states the rule, and tell me which of those pages was most recently updated. If none of them states the rule, say that plainly.

[teamspace] is the teamspace that owns the policy, such as Policies or Finance.

Worked example: two pages, one answer

Fictional example. The workspace contains these two pages:

Trade terms (current) — last updated July
"Trade customers may request a refund within 30 days from delivery. Changed in July from 14 days."

Trade onboarding checklist — last updated August
"Remind new trade accounts: refunds within 30 days."

The first answer read: “Trade customers have a 30-day refund window. [Trade terms (current)] [Trade onboarding checklist]”

The failure. Two citations, one figure, no visible problem until you open them. “30 days” and “30 days from delivery” produce different outcomes for anything sitting in a courier’s warehouse over a bank holiday. The answer dropped the condition, and the newer checklist made it look settled.
The repair. With the repair prompt and [teamspace] set to Policies, the answer quoted the governing sentence, named Trade terms (current) as the source and noted the July change from 14 days.
The extra finding. The checklist repeats a rule it doesn’t own, which is how workspaces drift. It now needs a link to Trade terms (current) instead of its own number.

Task kit for this job: Download (ZIP) · Reusable instructions

Go deeper

Optional training: Connect another app, and understand what that changes

Notion AI Connectors extend the search to apps outside Notion, including Slack, Microsoft Teams, Google Drive, SharePoint and OneDrive, Jira (beta), GitHub, Linear, Gmail, Outlook and Google Calendar. Connecting third-party apps needs the Business or Enterprise plan, and setting one up needs a workspace owner with admin rights in the other app. Owners manage them at Settings › Notion AI.

Permissions are mapped, not bypassed. But once Slack is connected, “where did we agree this” starts returning messages, and a message may be a proposal, an approval or a pointer to the governing page. Check which role it plays.

Coverage has limits: a connector generally reaches back one year from setup, new content can take up to three hours to be indexed and the first ingestion up to 36 hours. Yesterday’s decision in Slack may not be visible yet. If you disconnect an app, its content becomes unsearchable (up to an hour for some connectors) and is deleted within a day.

Optional training: The admin settings that quietly shape every answer

Workspace owners can adjust these at Settings › Notion AI:

Share data to improve Notion AI — whether workspace data is shared with Notion to improve the product.
AI Web Search — Enable web search for workspace turns web search on or off for everyone; Require confirmation for web requests makes Notion AI ask before visiting external sites.
Model controls — which models your team can use. Premium models stay off until an owner turns them on, because they spend Notion credits.
Usage — how many credits each member can spend.

Notion uses AI subprocessors to provide these features; its security and privacy page explains how. If colleagues get different answers to the same question, compare their sources, permissions and wording before treating either as the policy.

Optional training: Know which questions belong in a database instead

Notion says connectors are meant for finding and summarising information, not running complex calculations, and that Notion AI may struggle to aggregate lots of information from one or several apps. “What is our refund window?” is a good question. “How many refunds did we approve last quarter, and what was the average value?” belongs in a database view with filters and a total.

It also can’t tell you something was never written down. It can tell you it didn’t find it, which is a different claim. Repeat it that way when you pass the answer on: “Notion AI found no page stating this” rather than “we have no policy”.