A percentage with no source behind it is a multiple of five 41.7% of the time. One you measured yourself, 24.6%. We got those two rates by sorting 2,473 percentage claims across our corpus by where each number came from, then reading nothing but the last digit. That gap is a big part of why unsourced stats are round numbers. It doesn't prove any one round figure wrong, because a real, measured 50% happens all the time. What roundness does is tilt the odds: the rounder a figure, the farther it sits from an actual count.
Call this the round-number tell. You can read a number's precision before you open its source, and that precision carries something the citation does not. The tell sits one layer under the state of content decay, which asks whether a published number has gone old. This one asks something earlier. It asks whether anyone counted the number at all.
Why Unsourced Numbers Run Rounder
Made-up statistics cluster on round numbers because estimating a figure, half-remembering it, or inventing it all pull the number toward a multiple of five, while actually counting something tends to land on an uneven digit. That is why made-up statistics come out round more often than measured ones do. Across 2,473 percentage claims in our corpus, a multiple of five turned up 33.2% of the time, with a 95% confidence interval of 31.4 to 35.1. Chance alone would put that near 20.8%, since 21 of the 101 integers from 0 to 100 are multiples of five. The measured rate runs half again above chance.
Roundness is a reason to check where a number came from. It doesn't settle anything by itself. A survey really can come back at 40%. A share really can sit at exactly half. Every corpus holds genuine round values that no amount of counting would have nudged off the mark. So are round numbers less reliable? On average, a little, and the useful move is to treat a clean round figure the way a careful editor treats a number with no citation: as the first place to look.
Every claim in that corpus already carries a provenance label, assigned for reasons that have nothing to do with its digits. That is what let us line the last digits up against where each number actually came from.
Why Guessing Rounds To Five
Reach for a number you didn't measure and your mind hands you a round one. Ask a publisher what share of their readers open an email and they say "about a third" or "half," almost never "37%." The round number is the cognitive default, the value memory grabs when it has nothing exact to hold. Estimation slides toward zero and five the way water finds the drain.
Counting works the other way. Tally a thing and it lands where it lands, on 37 or 62 or 43, an integer with no reason to be tidy. The specific figure is the receipt. It records that someone did the work and did not round the result away.
So the first-party signature is specificity: a number resting on an unremarkable integer that estimation would never have picked. It has nothing to do with whether the number carries a decimal point.
Set two side by side. "35% of marketers" and "33.7% of respondents." The second is harder to invent and harder to misremember, because it reports an actual count instead of a mental round-off. The first might be real too. It just has to survive the question its round shape invites, and the specific number never faces that question.
The pull toward round numbers is well documented, and in one field it is treated as forensic evidence. Integer percentages turn up as a fingerprint of electoral fraud: Kobak, Shpilkin and Pshenichnikov, writing in the Annals of Applied Statistics in 2016, found that manipulated election results carry a surplus of round-percentage turnout and vote shares. The mechanism is the same one at work in a blog post. As they put it, "if election results are manipulated or forged, then, due to the well-known human attraction to round numbers, the frequency of reported round percentages can be increased." Different stakes, same tell.
Less Sourced Numbers Get Rounder
Sort those 2,473 claims into three provenance tiers and the tell sharpens into a gradient. Numbers with no attribution at all are a multiple of five 41.7% of the time (95% confidence interval 37.7 to 45.7). Numbers attributed to a named source drop to 36.7% (33.7 to 39.9). Numbers a workspace measured itself fall to 24.6% (22.0 to 27.5). The rate slides in one direction as sourcing weakens, and the intervals at the two ends never touch.
Unattributed 41.7% (95% CI 37.7–45.7), Sourced 36.7% (33.7–39.9), First-party 24.6% (22.0–27.5). Overall 33.2% versus a 20.8% chance baseline (multiples of 5 among integers 0–100 ≈ 21/101). N = 2,473 in-range percentage claims; 140 growth/out-of-range exempted. Wilson 95% intervals; deterministic, zero-LLM.
Those tiers map onto Claim Origin, the provenance label LiquiChart's living content infrastructure assigns to every claim it extracts. A number is Original or First-party when the workspace produced it, Sourced when it names an external citation, and Unattributed when it stands with no source at all. The classifier never looks at roundness. The gradient falls out of a labeling done for entirely separate reasons, which is what makes it hard to wave away. Precision predicts provenance, and you can read it off the last digit.
Read only that digit and the same shape shows up. Among claims that land on a whole-number percentage, 37.9% end in a zero or a five (35.9 to 40.0). Split those by origin and the unattributed ones end in zero or five 48.5% of the time, the first-party ones 28.9%. That is a different denominator, integer claims rather than all percentage claims, and it tells the same story from the other end of the number.
What The Test Excludes
A multiples-of-five test breaks the moment you feed it the wrong numbers, so most of the work was deciding what not to count. Years are round by nature; 2024 is not a rounded estimate. Counts, sample sizes, currency in round lots, denominators, single-digit small-N values, a genuine 50/50 split, a flat 100%: each lands on a five or a zero for reasons that have nothing to do with whether anyone measured it. All of them are held out and disclosed, because the exemptions are part of the finding. The test runs on percentage claims inside the 0 to 100 range, which is why 140 growth and out-of-range figures drop and 2,473 remain.
The result survives its own stress test. Drop the three values most likely to be genuinely round, zero, 50, and 100, and the gradient barely moves: 37.8% to 34.6% to 23.0%, still sliding one way, still separated at the ends. Every figure here comes from one committed script, scripts/staleness-study-v5/number-shape-analysis.ts, running the production Claim Origin classifier with Wilson 95% intervals and no LLM or network call anywhere in the path. A post about false precision does not get to hand-wave its own.
The Signal We Expected
The first version of this finding was wrong, and the wrong version was prettier, which is why it had to go. We expected measured numbers to carry more decimals, so that a non-integer value would be a second, independent receipt of first-party work. I ran about 30 cross-tabs against the v5 artifacts looking for that decimal signal. It is not there. The share of non-integer values comes out at 14.1% for unattributed claims and 14.8% for first-party ones, with sourced claims lowest at 9.2%, a U-shape whose confidence intervals overlap instead of climbing. Decimals say nothing reliable about where a number came from. Roundness does, so the roundness gradient stayed and the decimal story went in the bin.
Numbers Round As They Travel
Look again at the middle tier. A number with a working citation is still round 36.7% of the time, well above the 24.6% of numbers a workspace measured itself. Attribution helps, and it leaves most of the gap open, because a citation only tells you that someone else measured this. It says nothing about whether they measured it or borrowed it from a fourth party who rounded it first.
Every hand a number passes through is a chance to round it. The original 33.7% becomes "about a third" in the retelling, then "a third" becomes 33% for tidiness, then someone writes 35% because it reads cleaner. Four citations deep, the specific measurement is gone, and every link in the chain still resolves.
We have measured where those chains end. Just 17.2% of borrowed citations reach a primary source at all, the headline of our citation provenance study, and the tier that reaches fewest is the same tier of unsourced statistics that rounds hardest. Roundness and distance-from-source are two readings of one fact: the number has drifted from whoever counted it. What sets this reading apart is that you can take it before you open a source.
Read The Digits On Your Draft
The fastest way to spot a fake statistic is to run this tell on other people's numbers, because you read those cold, with no memory of where they came from. Your own draft is harder. You remember writing the sentence, and you don't remember whether you measured it. So start with the mechanical part, the part that needs no memory: open your last few posts and mark every percentage that ends in a zero or a five.
Treat each mark as a question. Did anyone count this, or did it arrive pre-rounded from a source you no longer remember? Some of those numbers were genuinely measured and will survive the question. The rest are candidates to re-measure before one of them becomes the stat someone quotes back at you.
When I read a number now, my own or anyone else's, I check the round ones first. A clean 40% is where I start looking. The number sitting on an arbitrary integer, the 37 or the 43, is the one I check last, because a figure that specific is hard to reach without counting.
This tell is the reading half of a pair. Its sibling, a vague quantifier is a poll question in disguise, works on the number that is missing, the "many" or "most" where a figure should sit, and shows how to run a poll that produces the real one. This tell works on the number that is already there and asks what its digits confess. The fix underneath both is the same: when a number is thin, the cheapest way to make it solid is to measure it yourself, which is also how to publish original research without a budget.
Before you audit a whole back catalog, answer it for one number. Where did the last percentage you published actually come from?
Whichever bucket you landed in, you now have a provenance label attached to that number, and the rounder it was, the more likely it belongs in the last two.
Answer Engines Read Precision Too
A round unsourced number is built to travel. It is easy to remember, easy to repeat, and it reads as confident, so it moves through human hands faster than the awkward specific figure it may have started as. That is its advantage on an internet of people sharing links.
The same roundness works against it in front of an answer engine. When a language model assembles an answer and walks the reference trail behind a claim, it is hunting for the place the number was actually produced. It settles on the specific first-party figure at the end of a chain, the 33.7% with a dataset under it. The laundered 35% that four posts assert and none can source gives it nowhere to land. The number that travels furthest through people is the same one an answer engine won't cite.
This is much of why AI cites third-party sources as unevenly as it does, and why the work of getting cited by AI search comes down to being the place a number can be traced to instead of one more stop on its way through the corpus. Your most shareable statistic and your most citable statistic are often different numbers, and the digits tell you which one you are holding.
Where The Scanner Goes Further
Everything so far is a heuristic you run by eye. It is fast, it is free, and it gets you most of the way, because the trailing digit really does carry provenance. It also stops where your eye stops. You can't scan a three-year back catalog by hand, and roundness alone can't tell you whether the source behind an attributed number is still alive or ever held the figure.
That systematic half is the job of the Content Health Scanner inside LiquiChart's living content infrastructure. The scanner classifies provenance: it extracts every claim on a page, assigns each one a Claim Origin, verifies that cited sources still resolve and still carry the number, and rolls the result into an Originality Score, the share of a page's or a workspace's claims backed by original data. It never reads the digits, and the round-number tell is the naked-eye stand-in for what the scanner measures directly. The scanner runs that read across everything you have published and checks the sources your eye can't.
Two halves of one question. Reading the digits tells you which numbers deserve a second look. The scan tells you where each one came from and whether its source still stands, across every post at once instead of the handful your eye can hold.
Your Most Quotable Number May Be Least Measured
The numbers you are proudest to quote are the clean ones. A flat 40%, a tidy third, a round half. They fit in a headline and they survive the retelling. That is the exact property that should make you check them, because the same tidiness that makes a number quotable is what it picks up as it drifts from whoever counted it.
Nothing in a normal publishing workflow flags the difference. A measured 40% and a guessed 40% render the same on the page. Same font, same confidence. They look identical. They are not the same claim.
The only tell you get for free is the shape of the digits. The roundest, most repeatable, most quotable number in your best post may be the one nobody ever measured. Until you read the last digit, it will keep reading as your strongest.