Sources and Citations in Chat

An answer built on your own material should be checkable in one click. That is the whole idea behind citations in Scholarly chat: you should never have to take the assistant's word for something, and you should never have to go hunting through a 200-page PDF to find out where a claim came from.

So every answer that draws on your sources shows you what it used, and every pointer it gives you is a link to the exact place.

The Sources button

When an answer cites anything, a Sources button appears beneath it with a count — the number of distinct sources behind that answer.

Open it and you get a list: each source by name, with the pages it was actually cited from. Click one and Scholarly takes you straight there.

The count is worth reading on its own. An answer citing one source is a close reading of one document; an answer citing six is a synthesis, and probably deserves a skim of the list before you forward it to anyone.

Inline citations

Citations also appear inline, right next to the claim they support, so you can check a single sentence without evaluating the whole answer. What a citation does when you click it depends on what it points at:

For anything with a timeline — a recorded customer call, a training video, a podcast episode — this is the fastest way to confirm a quote: the citation drops you at the second it was said, not at the top of a 40-minute recording.

How citations are checked

This is the part that makes the citations worth trusting.

Every citation is verified before the answer is saved. Scholarly checks each one against the sources the conversation can actually see: does that source exist in this conversation, and does the page it names exist in that document? A citation that fails either test is removed from the answer rather than shown to you.

The practical consequence is worth knowing: an answer sometimes cites fewer places than you expected. That is the check working, not the assistant being lazy. A pointer that would have led you to a document the chat cannot open, or to page 90 of an 80-page file, is taken out instead of left in to waste your time. What remains is the set of citations that actually resolve.

The assistant also will not invent a citation to fill a gap. If your sources do not answer a question, it says so rather than manufacturing a reference.

What the assistant can see

An answer can only cite what the conversation has in front of it. That is:

• Sources you linked with @ — type @ in the chat input to pull in a PDF, page, flashcard deck, podcast, recording, research session, or whole folder.
• Files you attached to the conversation.
• The content page you opened chat from. Start a chat from a PDF, a deck, a recording, or a video and that item is already in front of the assistant — you do not have to link it.
• Everything in the project, when the conversation is scoped to one. See below.
• Google Drive files you approved, if you have connected Drive. Only the files you picked through the Drive picker are visible — see Connections.
• Web results, when the assistant searches the web for something your sources do not cover.

Anything outside that list is not visible to the conversation, and so cannot be cited.

Scoping an answer to a project

A project keeps the sources, creations, and chats for one piece of work together. Chatting inside a project scopes the conversation to it: the assistant can draw on everything filed there without you linking items one at a time, and its citations point back into the project's own material.

This is the setting to use when a question spans a body of work rather than one file — a policy library, a quarter of customer calls, a certification syllabus. See Projects.

Rating an answer

Every assistant answer carries thumbs up and thumbs down at the left of the message actions.

• Your rating is saved with the message, so it stays attached to that answer when you come back to the conversation.
• Tapping the one you already picked clears it. Ratings are easy to change and easy to undo.

Ratings tell us which answers are landing and which are not. They are the fastest signal you can send about a bad citation or a weak answer.

When you disagree with an answer

You have better options than arguing with it:

1. Open the citation. Click through to the page or timestamp. Most disagreements resolve here — either the source says what the answer claimed, or you now know exactly where it went wrong.
2. Check the Sources list. An answer that cited the wrong document, or a thinner set of sources than the question needed, explains itself the moment you see the list.
3. Narrow what it can see. Link the specific document you mean with @, or move to the project that holds the right material. A tighter source set produces a tighter answer.
4. Retry on a different model. Hover the answer and use Retry to run it again on another model. Both versions stay in the conversation and you can flip between them — see AI Chat and Voice Mode.
5. Rate it down, so we see it.

Citations elsewhere in Scholarly

Chat is not the only place your material is shown its receipts:

• Deep Research reports carry their evidence — the sources behind each finding, so a report you circulate can be audited by whoever reads it. See Deep Research.
• Podcast chapters carry citations back to the material each chapter was built from. See AI Podcasts.
• Exam answer keys do too. Why this answer shows the cited source passage the key came from, or the original flashcard answer, labelled as such. If you think the key is wrong, Challenge this answer — once per attempt — re-reads that passage and tells you whether the key is upheld, corrected, or simply not settled by the source. A correction that makes your answer right raises your score. See Quizzes and Exams.

Related

• AI Chat and Voice Mode — linking content, models, retries, and tools.
• Projects — scoping a conversation to one body of work.
• Connections — approving the Google Drive files chat can cite.
• Deep Research — cited reports you can circulate.
• Quizzes and Exams — where Why this answer and Challenge this answer live.