Your AI will answer every question you ask it about your business. That is the problem. It answers the ones it should refuse, in the same confident voice, and nothing on the screen tells you which is which.
These twenty-five checks tell you instead. Each one takes a minute. Tick what you can prove today and watch your score move.
Start scoringA check passes only if you can show it. Not if you believe it.
"We know our costs" is not a pass. Opening the file where the costs live, with dates on them, is a pass. The difference matters because half-done is precisely what produces a confident wrong answer, and a confident wrong answer is worse than no answer at all. You act on it.
So score honestly. One point for each check you could demonstrate to somebody this afternoon. Zero for anything you would have to build, find, or go and ask about first.
Nothing you tick is sent anywhere. The score lives in this browser and nowhere else.
Ask three people for last month's revenue and you should get one number. Often you get three, and all three are defensible.
Ten metrics, one sentence and one formula each: revenue, new customers, repeat rate, ad spend, return on ad spend, contribution margin, units, average order value, refunds, inventory on hand.
Orders placed, money settled, or store gross sales. Named on the number itself, not in a footnote nobody opens.
The date each source stops changing. For example: a month is closed ten days after it ends.
One time zone per channel, written down, including the marketplace clock your sales are stamped in. One currency, with the conversion rule if you sell in more than one.
A definition that changed in June says so, so a year-on-year comparison can be trusted or discarded on purpose.
One bottle of shampoo wears three codes: a marketplace product code, a store variant, a barcode on the shelf. Nothing in your data says they are the same bottle, so your best seller arrives as three modest products.
Every sellable code maps to the product family a customer would name. One file, one owner.
A 3-pack is one order and three units. Both numbers are knowable, on every row.
Every order belongs to somebody, and every somebody has a start date, even where the marketplace hides the person behind an alias.
Each campaign points at a product or family and a destination, so spend can be scored against the sales that actually happened.
Or you can state plainly which sales are missing from every roll-up, and everyone knows it.
A number with no label is a trap, and the trap is sprung by whoever trusts it most.
Which system produced it, not which dashboard displayed it.
Visibly, wherever it appears, including inside the answer an AI gives you.
The ad pixel that cannot see marketplace checkouts. The buyer identity that arrives late. The report that leaves a channel out.
A group of customers too young to have a repeat rate shows nothing. Not a small number. Nothing.
Ad platforms keep crediting sales after a week ends. Marketplace settlements arrive later and correct what you already reported.
The most important numbers in your business are not in any system you can connect. They are in a founder's head, a supplier email, and a spreadsheet somebody renamed.
What it costs to make it, ship it, clear customs and prepare it, per unit, and from what date.
Costs, launch dates, targets, which ads point where. Each with a person's name and a date attached.
Lifetime value and return on ad spend are recalculated every time they are asked for. Never saved.
The number you consider healthy, set by you, not an industry average from an article.
A named person, a review date, and the habit of adding new products the week they launch.
The first twenty checks make good answers possible. These five are what stop a bad one reaching a decision.
The system that recorded the money, and it ties to the bank and the books.
With a written threshold, so two tools disagreeing becomes a message instead of a silent compromise.
The number in the sentence is checked against the system that produced it, before anyone reads the sentence.
A report that cannot compute something says so, rather than printing a zero or a blank that reads as a real result.
Not a person remembering. A job that runs on its own and complains.
Score honestly. The list is only useful if the gaps it finds are real ones.
An AI will answer every question you ask and you cannot tell which answers are wrong.
Safe for description. What happened, roughly. Not safe for decisions about money.
Safe for most decisions, with named gaps you can say out loud in the meeting.
An answer can be acted on the same day it is given.
Copy this standard, adapt it, publish your score. No permission needed, no attribution required, though we would like the link.
So here is ours, honestly. We wrote the list; we do not clear all of it.
Shipped by default. Definitions shared by every answer, labels on every number, disagreements flagged rather than averaged, brand-by-brand isolation, and a weekly job that complains when an assumption stops holding.
Yours to decide, ours to hold. Your product family map, your costs, your targets, your close rule. We give them one home and apply them everywhere, but you set them. Nobody can outsource what a product costs.
Not yet. Some relationships in the model cannot be explored yet. Facts captured automatically from a session skip the approval queue. Our implausible-figure gate covers reports, not file exports. And a wrong answer does not yet leave a permanent automated test behind.
We publish this because the fourth column is the roadmap, and because a standard nobody fails is not a standard. If you find a claim of ours that overstates what we do, tell us and we will change the page.
Pick five questions you already know the answer to, where the answer is surprising. A product that sells well and loses money. A channel that looks bad and is not. Then ask an AI connected to your data. If it only gets the easy ones right, the twenty-five above tell you which piece is missing.
Bring one of those questions to a call