What an AI deep research tool reads, and how it checks its citations

What does an AI deep research tool actually read?

Real pages, opened and read: the open web, the pages you point it at and, with no web search, what is already on your computer. A report that cites a page the tool never read is a guess with a link on it. In Minda Super Intelligence, the personal AI agent app for your computer with a built-in browser, Research mode reads real pages and lists only those it read. Start it from the Research switch in the chat box’s Modes menu ("A planned report with sources"), by typing /research and your question, or from "Research a topic in depth" in an empty chat.

• The open web. Each sub-question is searched in several wordings, and the best results are opened and read.
• Pages you point it at. Type @ in the chat and pick tabs, or All tabs in the workspace (the 20 most recent). Your picks are numbered first, and when you name the sources to use, it uses only those.
• Pages behind your sign-ins. It reads in the window’s workspace, with that workspace’s sign-ins, so a site you are signed in to there can be read.
• Videos. The transcript of a YouTube video, or any video with captions, is added, up to 20,000 characters.
• What is already on your computer. With no web search, it researches your tabs, saved pages, highlights, notes, past chats, history and bookmarks instead.

How does it pick which pages to read?

1. It plans first: your question becomes 3 to 6 sub-questions, so a broad topic is covered from several sides instead of one search.
2. It searches each sub-question in different wordings, so one phrasing does not decide what it finds.
3. It reads in batches of up to 10 pages at a time: at least 5 sources for a narrow question and 10 to 20 for a broad one, and at most 80 pages in one research reply.
4. It prefers primary sources (the table below) and labels what comes from forums and reviews.
5. It cross-checks every number, price, date and name in at least two independent sources it read. When they disagree, it keeps both and says which says what.

How does it read a page without opening a tab?

Pages are read in the background: no tab opens and your tab strip is untouched, but the read happens in the window’s workspace session, so Hide my IP, the ad blocker and the request gate apply as they would in a tab.

• An article is read from its document. Any other page loads in a lean hidden page with images off, at most 30 seconds each.
• The passages that match your question are picked on your computer, ranked by BM25 with no AI call: about 1,200 characters a page when ten pages are read at once, up to 4,000 when three are.
• Footnote numbers such as [12] on a wiki page are dropped, so the only numbers in brackets the model sees are Minda’s own citation numbers.
• A line aimed at an AI, such as a footer’s note to AI assistants, is kept even when it falls outside the chosen passages, so the SI agent can name it and you hear about it instead of it steering the report unseen.
• Sites you blocked in Settings → AI agent → Access, and a task’s "Only these sites", are honoured: a blocked site is not read.

How does it check that every citation was read?

The Sources list is not the model’s own claim. Before the reply is shown, Minda rebuilds it from what the reply actually read: the pages it cited, or, when it cited nothing, every page it read, up to 50.

• A source it never read is removed, with a note under the list: "Removed a listed source this reply did not read."
• When no page was read at all, the Sources say so: "None: no page was read for this reply."
• Each [n] in the chat is a link to the page it stands for, so checking a claim is one click.
• A model with no tools cannot search or read more pages: it writes the report from what you shared and says in the TL;DR that it rests on those sources only.

What does the report look like?

Five parts, always in this order, so you can go straight to the part you need.

• TL;DR: three lines that answer the question.
• Findings: the answer, each claim with its citation numbers.
• A comparison table when the question compares things.
• Disagreements and open questions: where the sources differ, and what they did not settle.
• Sources: one line each, the number, the page’s title and its address.

What happens when it cannot search the web?

With Keep SI on this computer, Minda refuses every cloud model and web search is off. The run does not stop and does not ask you to change the setting: it researches what is already on your computer (your tabs, saved pages, highlights, notes, past chats, history and bookmarks) and says in the TL;DR that the report rests on your own pages and notes, not on a web search.

That makes a private literature review possible: everything you clipped over a semester, answered with citations to your own saved pages. Private SI chat and running SI on your computer cover the models that run there.

Research mode or a research skill: which one?

Research mode answers any question. The gallery’s research skills run the same research with a set shape for a common job; add them once in Settings → Skills → Gallery. Research can take longer and use more model calls than a quick answer, so use it when the sources matter.