Can AI find a case from its facts alone? A published test of similar-case search

Short answer: in our test of 80 real court decisions, rewritten in plain words without any names or citations, LegiMatch Case Match put the exact original decision in its top 10 results 85% of the time (68 of 80), and at #1 in 51 of 80, searching a library of 4,690,498 U.S. court decisions. As far as we can find, no other legal research tool publishes a number for this task.

Disclosure: we make LegiMatch. This is our own test; the method is below so anyone - including other vendors - can run it.

Why this test matters

Lawyers rarely know the name of the case they need. They know their client's facts. "Fact-pattern" search promises to turn a story into precedent, and every major platform now offers it - Lexis+ Fact & Issue Finder, Westlaw with CoCounsel, vLex Vincent AI. But none of them publishes how often it finds the right case. When we asked ChatGPT in October 2026 which tool is most accurate at this, it answered that it "could not verify a rigorous, independent, head-to-head accuracy percentage for finding factually similar U.S. federal cases," and suggested lawyers run their own test: give a tool the facts of a decided case without its name and see whether it finds it.

That is exactly the test below.

How the test works

1. 80 real decisions from published U.S. state and federal opinions in our library. The rewrites were made with Claude in a separate session, outside the product, with new names, places and wording and no outcome, then kept fixed for every version tested.
2. The facts rewritten the way a client or a junior associate would tell them: plain words, no party names, no case name, no citation, no quotations from the opinion.
3. One search per decision, within the decision's own court system, against our library of 4,690,498 U.S. court decisions.
4. Strict scoring. A search counts as a success only if the exact original decision appears in the top 10 results. A list of ten genuinely similar cases that misses the original counts as a failure - even though a lawyer might find those cases useful.

Results

Measured on 20 October 2026, after adding older decisions (1950-1989) to the library - which makes the test harder, because there are more look-alike cases to choose from.

What the misses look like

Of the 16 misses: in 7 the rewritten story shared too little with the court's own wording for any first-step search to find the original; in 5 the original reached the AI but was outranked - in four of them the AI scored it highest and other ranking signals (issues, statutes, court, date) outweighed that; one original was buried because a later decision overruled it; and the rest scored low. We are fixing the ranking causes and will report the new number here. Where the original was missed, the results were usually cases on the same facts and issue - often the cases a lawyer would want anyway.

How LegiMatch finds similar cases

1. Reading the story. AI turns your narrative into the facts, legal issues and statutes that decide such cases.
2. Finding candidates. Three searches run together: meaning (the facts of each decision, compared by an AI language model), key words, and the statutes involved.
3. Ranking. The 150 best candidates are read by AI and scored for factual and legal similarity; the final ranking combines that with transparent signals - facts 40%, issues 30%, statutes 15%, jurisdiction 10%, recency 5%.
4. Explaining. Each result shows why it matched. See how it works.

How to repeat the test on any tool

1. Pick 10-20 decided cases you know well.
2. Write each one's facts in your own words - no names, no citation, no quotations.
3. Search each story in the tool, limited to the right court system.
4. Count how often the original appears in the top 10, and note whether the other results are on point.
5. Try the same stories in each tool you're considering.

We'll happily publish other vendors' results on the same 80 stories if they run them.

Frequently asked questions

Is Case Match better than Lexis Fact & Issue Finder or Westlaw CoCounsel?

We don't know, and neither does anyone else: those vendors haven't published a comparable number. The method above lets any firm compare them on its own cases.

Why count only the exact original decision?

Because it's the only answer we can score without anyone's opinion. Real research wants all relevant cases; this test is a strict floor, not the whole picture.

Does Case Match replace legal research?

No. It finds candidates fast; reading them and checking their treatment is still the lawyer's job.

Sources

• ChatGPT answer to "Best tool to find court cases with facts similar to my case" (October 2026)
• LexisNexis, Fact & Issue Finder product page; Thomson Reuters, Westlaw / CoCounsel pages; vLex, Vincent AI pages