A dashboard shows you a cohort. One segment, it says, converts at 50%. You feel something tighten and lift at the same time.
You start writing the plan. Double down on that segment. Re-cut the ad copy. Brief the team Monday.
Then you look closer, because some careful part of you always looks closer, and you find the truth under the number. The sample size is two. One person bought. One did not. Fifty percent is not a conversion rate. It is a coin that happened to land once.
The dashboard did not lie, exactly. It just declined to mention the thing that mattered most. It dressed a coin flip in the clothes of a strategy and let you walk out the door wearing it.
Confident numbers are the dangerous ones
We have trained ourselves to trust the clean number.
The big bold percentage. The arrow that points up. The chart that resolves into a tidy slope. Clean reads as true. It is not. Clean is just clean. A number rounded to a confident two decimals tells you nothing about whether it would survive being asked twice.
The most expensive decisions in a startup are not made on missing data. They are made on present data that quietly omitted its own uncertainty. The cohort of two. The lift measured across a week with a holiday in it. The "channel that converts" with an n you never saw because the tool did not think you needed it.
False precision is not a small bug in analytics. It is the main one. It is the difference between a tool that informs you and a tool that flatters you into a wall.
False precision is not a small bug in analytics. It is the main one.
The honest move is the hatched cell
Here is a small idea with large consequences.
When a cohort has fewer than three records, do not color it in. Hatch it. Gray it out. Make the eye skip it. Refuse, on principle, to render false precision as a number a human will act on.
This feels like a downgrade. It is the opposite. A tool that suppresses the cells it cannot stand behind is a tool you can finally trust the colored cells of. Honesty about the weak numbers is what makes the strong numbers worth anything at all.
Every claim should arrive with its own receipt. Not buried in a methods appendix nobody opens. Attached.
- The join we used.
- The window we measured.
- The sample size, stated, not implied.
- And the one thing that could make us wrong, said out loud, before you forward the insight to your team and stake a quarter on it.
"Enterprise FinServ converts at 19.4%, n=22, last 90 days, joined on email plus a 60-day window, and the caveat is two of those came from the same parent company." That sentence is longer than "19.4%." It is also the only one of the two you can actually build on.
Trust is a feature, not a vibe
Founders do not lose faith in analytics because the numbers were wrong once.
They lose faith because the numbers were confident and wrong, and there was no way to have known.
After that, every dashboard gets the side-eye, and the spreadsheet comes back out, and you are doing the join by hand again at 11pm because at least then you can see what you are standing on.
The way back is not a prettier chart. It is a tool that tells you the size of the ground under each number before you step on it. That says two when the sample is two. That hatches the cell it cannot defend. That hands you the receipt without being asked.
More data was never the answer. You already drown in data. What you have been missing is a tool with the discipline to tell you which of it to believe.
Throughline attaches the join, the window, the sample size, and the one thing that could make it wrong to every insight it shows.
Throughline reads across every tool you already pay for and narrates the part that moved. Early access is open.