Why I Publish Claims With Numbers in Them

01An adjective is an opinion. A number is a fact you can carry.

I write every claim that matters twice. The first version is how I would say it out loud, and it always comes out as some flavour of good, fast, popular or effective. The second version has a number in it. The first version reads better and has never once come back to me from somewhere else. The second version turns up in places I did not put it.

A sentence with a number in it gets quoted. An adjective gets skipped.

What a model is doing when it quotes you

It is not admiring the prose. It is looking for something it can hand to a stranger and stand behind if the stranger pushes back. A figure is portable in a way an opinion is not, because it survives being lifted off your page and dropped into an answer written by somebody else. That is the whole mechanism, and it is why models take sentences rather than paragraphs. The unit of citation is one sentence standing on its own feet, with no help from the two around it.

The adjective problem

Adjectives feel safe precisely because nobody can check them, and that is the flaw rather than the feature. A claim that cannot be verified is also a claim that cannot be relied on, so it carries no weight and gets left where it sits. Popular cannot be measured, so it cannot be trusted, so it never travels. Nothing that could never be wrong is worth repeating. The moment you attach a figure you take on the risk of being contradicted, and that risk is the entire reason the sentence is worth anything at all.

The other half of it is whose number it is. A statistic borrowed from a study is already sitting on a hundred other pages, so repeating it costs a model nothing and earns you nothing. A count from your own operation exists in one place. How many things you have built, how long the job actually takes, what the fix cost, how many times you have run the process. Those figures are unremarkable to you and unavailable to everybody else, which is the exact combination that gets picked up.

02Publish the floor, never the peak

Here is the rule that makes all of this survivable, and most people have it backwards. I do not publish the real number. I publish the floor underneath it, the figure I could still defend on the worst month of the year, and then I leave it alone for a long time.

Why a floor beats the truth

An exact figure is a hostage. It is correct on the morning it goes up and quietly wrong by the next quarter, and every stale number left standing teaches a reader to discount the next one you publish. A floor does not decay. If the real figure doubles, the floor is still true. If a bad month arrives, the floor is still true. The exact number is true today. The floor is true for a year. Round down until the claim is almost boring, then publish the boring version, because the boring version is the one still standing when somebody finally checks it.

There is a way to overdo this. Round far enough down and the claim stops saying anything, because “more than none” is unfalsifiable in precisely the way an adjective is. The floor still has to be a number a reader would find worth knowing. If rounding has made it comfortable, you have talked your way back into the safe version with digits bolted on.

03Where to put the number so it gets found

Having the figure is half the work. The other half is putting it somewhere a machine reading at speed will actually pick it up, and two habits decide that more than anything else on the page.

Put the number in the sentence, not in the chart

A chart is a picture of a number, and a picture is not a sentence. I have watched pages carry a beautiful graph and then get quoted for a throwaway line in the second paragraph, because the line was readable and the graph was an image with a caption under it. A number locked inside an image is a number nobody can quote. Draw the chart if it helps a person, then write the finding underneath in one flat sentence with the figure spelled out. That sentence is the part that travels, which is the same reason a small free tool can out cite a large company.

Date it, or watch it rot

Put the window next to the figure. Not a marketing year, an actual span: last month, the fortnight to the twelfth, the quarter just closed. An undated number asks to be believed forever, which is why it gets believed for nothing. Dating it costs four words, and it is the whole difference between a figure that ages honestly and one that simply gets old. It also makes you the kind of source worth coming back to, which is most of how a model decides who to recommend.

Tomorrow: your analytics say nobody read this. Your server log disagrees, and it has been disagreeing for months.