2. Summary
A misfit says where to look, not that anything is wrong. Fixed fees, price lists, caps and recurring bills fail these tests honestly.
3. First-two-digits test
The most informative test. Bars are what your data shows; the line is Benford's expectation. Orange bars are over-represented with |z| above 1.96. A spike here (say at 49 or 99) often means amounts held just under a limit or a recurring invented figure. With 90 bins, four or five will cross 1.96 by chance alone even in clean data, so read a lone small spike as noise.
4. First- and second-digit tests
5. Just under an approval limit
For each limit, the amounts within 10% below it against those within 10% above. With smooth, Benford-shaped data a little over half should fall below (about 52.5%); many more than that is the pattern left by split purchases and limit-shaving. Flagged when p < 0.01 with at least 10 amounts in the window.
6. By vendor, employee or card
First-digit test for each group with enough amounts, ranked. Excess MAD subtracts the MAD sampling noise alone would give at that group's size, so small groups are not punished for being small. q is the p-value corrected for testing many groups at once (Benjamini-Hochberg); q < 0.05 is flagged.
7. Repeated amounts, round numbers and last two digits
8. Summation test
Sums the amounts in each first-two-digits bin. Each bin should hold about 1.1% of the total; a bin with twice that or more is carried by a few large amounts, listed here.
9. Records to look at
Every record caught by at least one test, most-caught first. Row numbers match your spreadsheet (the header is row 1). Showing the first 200; the download has them all.
Common questions
Does a failed Benford test mean fraud?
No. A digit test says where the numbers differ from what Benford's law predicts, not why. Price lists, fixed fees, caps, a few vendors that bill the same amount every month, or a narrow range of amounts all fail the tests with nothing wrong. Use the results to choose which records to look at, then look at them.
How many records do I need?
About 300 amounts for the first-digit test and about 1,000 for the first-two-digits test, which is the most useful one. The page says so when a file is smaller, and the per-vendor ranking only includes vendors with at least 50 amounts by default.
What is MAD and why is it the headline?
MAD is the mean absolute deviation between the observed and expected digit proportions. Mark Nigrini proposed it as the conformity measure for Benford tests because chi-square rejects almost any large file for trivial differences. The verdict bands (close, acceptable, marginally acceptable, nonconformity) are his, from Benford's Law (Wiley, 2012). Chi-square and its p-value are shown too.
Which data fits Benford's law?
Amounts that arise from many independent transactions and span several orders of magnitude: vendor payments, expense claims, invoice lines, journal entries, sales. Not assigned numbers (invoice or check numbers, ZIP codes), not amounts bounded by a rule (a $50 meal cap), and not a price list. The applicability notes under the summary check the range and skew of your file.
Is my data uploaded?
No. The CSV is read in your browser and never leaves your device. The optional usage counter stores only that an audit ran, in your own browser, and contains no amounts or names.
Who this is for, a worked example and when to use something else: Benford's-law testing for internal auditors, fraud examiners and bookkeepers. Looking for the same invoice paid twice rather than unusual digits? The duplicate payment audit pairs up likely double vendor payments, also in your browser.