Already checking references with AI?
Claude can check your references. RefRunner checks every one, every time.
A reference list with one retracted paper or one invented citation is a problem for the author, the reviewers, and the journal. Asking a chatbot feels like a check. So we tested it: the same 85 references, with 20 problems planted in them, given to Claude and to RefRunner.
No account: 20 references a day. Free account: 50 a day. Paid plan: any paper, any size, 1,000 references a month. See plans
📱 On a phone? Create a free account now and your reference lists will be waiting when you start working on your paper.
What we found
Without web search, a chat can't see retractions
Six papers in the list were retracted this year. Plain Claude passed all six and told us it knew of no retractions. It answers from what it learned in training, and a retraction notice from last month isn't there. It spotted the invented references well; it had no way to see the retracted ones.
With search, it's good, but not every time
Turn search on and Claude caught every retraction and every invented reference. But it looked some entries up and took the rest from memory, and in two of three runs a wrong title or a mistyped DOI got through. So we told it to search every entry. Those runs took longer, cost two to three times as much, and did worse: twice it marked a retracted paper "Confirmed."
Its corrected list is one more thing to check
Asked for one, Claude rewrites the list, and nothing shows what it changed or why. RefRunner looks up every entry against Crossref and OpenAlex, flags what doesn't exist or was retracted, and shows each correction beside your original, so you accept or reject it. Then it gives you the list in your journal's style, as Word, BibTeX or RIS.
The test, side by side
Claude, plain chat
- Problems found
- 12 of 20
- Retractions
- 0 of 6
- Time
- 48 sec
Said it knew of no retractions.
Claude with web search
- Problems found
- 18–20 of 20
- Retractions
- 6 of 6
- Time
- 3½ min
Three runs: 19, 20 and 18. Each missed different entries.
Claude, told to search every entry
- Problems found
- 17–20 of 20
- Retractions
- 4–6 of 6
- Time
- 5–6 min
Three runs: 20, 18 and 17. Twice it called a retracted paper “Confirmed.”
RefRunner
- Problems found
- 20 of 20
- Retractions
- 6 of 6
- Time
- 49 sec
Every entry looked up.
October 2026, Claude Opus 5.5: one plain-chat run and three runs of each search setup. The first two prompts asked, as an editor would, for a verdict on each reference; the third added "use web search to look up every reference, do not rely on memory" and asked for a corrected APA list. At API prices a search run comes to about $1–3.50; on a subscription it counts against your plan's usage limits. The list: 65 accurate references (three published after Claude's training data), 6 retracted in 2026, 6 invented, and 8 real papers with a wrong title, authors, year, journal, pages or DOI. In fairness to Claude: with search it noticed that one paper's Crossref record and the publisher's website give different pages (the printed PDF has a third range), which RefRunner can't see, and it named the real paper behind one invented citation. The lists, answer keys, every answer and the grading scripts are on GitHub, so you can check our grading or run it yourself, and the smaller 16-reference list opens in RefRunner with one click.
Check every reference, not most of them
Drop a manuscript or paste a reference list. Your first 20 references need no account, and you get back a list you can send.