Home · Blog · Research
Article · Research

AI research report with sources you can check

Run a deep research report in Oppermind, then check each numbered source against the claim it supports and keep a source log you can hand on.

17 NOV 2026 · 8 MIN READ · OPPERMIND PTY LTD
In short. Start a run from the chat with “Deep research:” and a specific question. The report comes back with numbered citations, a Sources list of every page it read and, when something went wrong during the run, a Limits of this report section. A citation number tells you which page the report credits for a sentence, not that the page agrees with it. Open the source behind every claim you plan to repeat, compare the wording, and record what you find in a source log.

Someone will ask where a number came from, usually the day after you send the report. The answer has to be a page you opened yourself. Oppermind’s deep research does the searching, reading and first draft, and shows enough of its working to make the checking quicker.

Ask a question the searches can work with

In the chat composer, type /deep and pick /deep-research, or click the Research chip when an empty chat shows it. Both put “Deep research:” in the box, and you type the question after it. A plain message such as “do deep research on...” starts a run too. Stay in Auto, where the run always happens and the report arrives as written.

Before it reads anything, the run turns your question into several web searches. The words you give it decide what those searches find, so name the place and the period, and say what you’re comparing. Here is a question made up for this post:

“Deep research: How do councils in Victoria and South Australia run kerbside food and garden organics collection, and what do their published reports say about contamination since 2023?”

That gives the searches more to work with than “FOGO bins”. To get journal articles, ask for them. Words such as peer-reviewed, systematic review or literature review switch the run to a mode that ranks scholarly sites higher.

It reads public web pages only. A file attached to the same message isn’t read, and neither are pages behind a login or your organisation’s own figures. Bring those into the same chat afterwards with an ordinary message.

Watch what it reads while it runs

A plan card under your message ticks through planning, searching, reading and writing the report. The Reading count is how many pages it tried to open, and Writing the report from is how many it could read. When the second number is smaller, some pages didn’t load and the report was written without them.

By default a run reads up to eight pages and may do a follow-up round of searches if it judges there are gaps. On each page it works from the opening part of the text. A figure near the back of a long annual report or PDF may never have reached it, which matters later when a claim seems to be missing from a source you know covers the topic.

If the run fails, the reply says what went wrong and no report is written in its place. A search outage is reported as a problem on Oppermind’s side, so wait a few minutes and send the question again. If it read pages but couldn’t write the report, the reply lists the pages it read so you can open them yourself.

Read the end of the report first

The report body puts numbers like [1] and [2] after the sentences they come from. Under it, the app adds two sections of its own.

Limits of this report appears when the run hit trouble. It can say that some web searches failed, that some pages couldn’t be read, that only one source could be read (and to treat the report as preliminary), that the time limit stopped the follow-up searches, or that citation numbers pointed at a page that wasn’t read and were removed. Read this before the body. A report built on one or two sources gives you leads to follow up.

Sources lists every page the run read, numbered to match the body, each linked. The list is built from what was read. If the writing step added a reference list of its own, that list is dropped and replaced.

Go down the Sources list next. A source whose number never appears in the text was read but credited with nothing. And look at who published each one. Five pages from the same council website give you one council’s view of its own program.

Check each claim against its own page

Only check the claims you’ll repeat. Go through the report and mark every sentence that carries a figure, a date, a named organisation, or a word like “most”, “first” or “only”. Those are the sentences someone will quote back at you.

For each one, open the numbered source and find the words. Ctrl+F (Find in page on a phone) for the number or a distinctive term beats reading top to bottom. When you find it, read the paragraph around it and check that the place, the year and what was measured all match your sentence. A contamination rate from a six-month trial in two suburbs may not describe a whole council’s service.

Then leave the page. University of California, Irvine Libraries’ guide to the SIFT method says to investigate the source, find other coverage and trace a claim back to where it started. Search for the publisher to find out who runs it, and for the claim itself in case other sources report it differently. Wineburg and McGrew’s study of how people judge websites, published in Teachers College Record, found professional fact checkers did this, opening new tabs to check a site rather than staying on it. If the page you’re on is quoting someone else’s report, open that report and cite it in your own work.

When the page doesn’t say what the report says

Most mismatches fall into a few patterns. Treat the table as a starting point for your own judgement.

What you find on the pageWhat it usually meansWhat to do
The figure is there, for the same place, year and measureThe claim is supportedLog it as checked and cite the page
The figure is there, but for a different year, area or groupThe report widened itRewrite your sentence to match the page
The page says “may”, “in one trial” or “early results”The hedge was lost in the summaryPut the hedge back in your version
You can’t find the claim anywhere on the pageWrong number, a part of the page it never read, or no support at allCheck the other sources, then cut it if nothing supports it
Two sources give different figuresDifferent methods or dates, or a real disagreementGive both with their dates, or use the primary source and say why
The link is dead or the page has changedMoved, paywalled or updated since the runFind the current version, or drop the claim

You can ask in the same chat which source supports a sentence. That follow-up gets an ordinary chat reply, so treat the answer as a pointer and still open the page. Keep the original report as it came back, and write corrections into your own copy.

Papers, DOIs and retractions

When a page it read, or the report itself, carries a DOI, the run looks up to eight of them in Crossref’s public metadata and checks whether any update has been registered against them. What it finds goes into Limits of this report: a DOI that appears in the report but not in any cited page, a DOI it couldn’t confirm (which, the note says, doesn’t prove the DOI is invalid), a title that differs from the registered one, or a registered update such as a retraction or correction, with a prompt to read the notice.

Crossref says it acquired the Retraction Watch database in September 2023, and since January 2025 those retractions have been in its public API. Retractions are the most complete update type in that data, by Crossref’s own account, while corrections and expressions of concern are less complete. No flag on a paper can mean nothing was ever recorded. For any paper your argument leans on, open its page on the publisher’s site, look for a notice and read the part your sentence relies on, since a DOI that resolves only shows the paper exists.

Keep a source log

A source log is one row per claim you keep. Make it in the Sheet editor with these columns:

Paste the report’s Sources list in as a starting column, or ask in the same chat for a spreadsheet with one row per source. Then fill in the claim columns yourself as you check.

The log is what you hand on. A colleague can recheck the changed and cut rows instead of the whole report, and anyone asking “where’s that from” gets a URL and a date. Turning the checked report into a long document with footnotes and a bibliography is a separate job, done after the log is complete.

Run the report. Check every source.

Frequently asked questions

Can I make it use only government or academic sources?

There’s no setting for it. Say what you want in the question (“from state government and council sources”), then check the Sources list to see whether the run followed it. Scholarly wording in the question does make it rank journal sites higher.

Why does the Sources list have pages no sentence cites?

The run read them, and the writing step credited nothing to them. They can still be worth opening, since a page it read but didn’t cite sometimes holds the detail a sentence left out.

Will the same question give the same report next week?

Not necessarily. Each run searches again and pages change, so put the run date in your source log. Two dated logs side by side show what moved.

Does it work on the phone?

Yes, from the same composer. The finished card reads Task Complete. If a report stops short with no Sources list, reload the chat to see the whole reply.

Related reading

Oppermind is not affiliated with, endorsed by, or sponsored by Crossref, Retraction Watch or the Center for Scientific Integrity, the University of California, Irvine, Teachers College Record, the Digital Inquiry Group (formerly the Stanford History Education Group), or the authors cited below.

Sources

  1. Retraction Watch. Crossref documentation (maintainer Martyn Rittman). Last updated 19 January 2025. https://www.crossref.org/documentation/retrieve-metadata/retraction-watch/. Accessed 2026-11-04. Used for: acquired the database from the Center for Scientific Integrity in September 2023, available through the REST API and as a CSV, updated every working day, retractions most complete, corrections and expressions of concern less complete.
  2. Retraction Watch retractions now in the Crossref API. Crossref blog, Martyn Rittman. 29 January 2025. https://www.crossref.org/blog/retraction-watch-retractions-now-in-the-crossref-api/. Accessed 2026-11-04. Used for: retractions and corrections from the Retraction Watch database added to the REST API (January 2025), the source field, the same retraction can appear more than once.
  3. How do I evaluate information? (Misinformation - Get the Facts). University of California, Irvine Libraries. Last updated 6 May 2026. https://guides.lib.uci.edu/Misinfo/EvaluateInformation. Accessed 2026-11-04. Used for: the SIFT moves (stop, investigate the source, find better coverage, trace claims to the original), lateral reading, credited to Mike Caulfield.
  4. Lateral Reading and the Nature of Expertise: Reading Less and Learning More When Evaluating Digital Information. Sam Wineburg and Sarah McGrew. Teachers College Record 121(11), 2019 (record page dated 1 July 2018). Record page: https://cor.inquirygroup.org/research/lateral-reading-and-the-nature-of-expertise. Accessed 2026-11-04. Used for: fact checkers compared with historians and students, fact checkers leaving a site to judge it. DOI: 10.1177/016146811912101102.