What you would need to know

The single most useful question you can ask about any analysis, yours or somebody else's, is what does one row represent. An order. A customer. A customer-week. A session. A support ticket. It takes four words to ask and it is the commonest structural error in the whole field, because it is invisible once the numbers come out.

Here is why it matters at Daybreak, concretely. Marketing's question is about orders, so the ideal table has one row per order and the conversion rate is a count of converted orders over all orders. The founder's question is about people, so the ideal table has one row per customer. Those are two different tables, and answering the founder's question on the order table does something specific and silent: a customer who ordered forty times contributes forty rows, a customer who ordered once contributes one, and the answer you get describes the habits of heavy buyers wearing the label "our customers". Nobody sees this happen. The number looks completely ordinary.

So write the grain on the first line of the sketch, in words, before any column name. "One row per order, for orders placed between 15 September and 8 December." If you cannot write that sentence, you are not ready to ask anyone for data yet, and five more minutes with the pen is cheaper than the two days that follow.

The ideal table is not invented from nothing. Each of the four slots from session 1 demands a specific kind of column, and if you work through them in order you will have most of the sketch without any creative effort at all.

The decision slot is the one people leave off the sketch, and leaving it off is how you end up with a single global number that nobody can act on. If the spring choice is per channel, an answer that does not carry a channel column is not an answer to the question that was asked, however correct it is.

Two more columns earn their place on most sketches even though no slot demands them: an identifier, so you can check whether a row is duplicated, and a date of record, so you can tell when a value was written rather than only what it says.

Never collected. Nothing anywhere holds this value, because no system was ever asked to write it. The reply goes to whoever owns the system that would write it, it asks for collection to begin, and it names the date from which you need it. The cost of this gap is time: the answer arrives one collection period from now, not today, and saying that plainly on day one is what lets somebody plan around it.

Collected but not kept. The value existed and was overwritten, truncated or aged out of a retention window. This is the gap most worth chasing hard and fastest, because a copy is often sitting in an export on somebody's laptop, in a vendor dashboard that keeps ninety days, or in a monthly report nobody deleted. The reply asks who has a copy and how far back it goes, and it is urgent in a way the other two are not, because the window is still closing while you write.

Exists but not joinable. Another team has the column, under an identifier yours does not share. This is a plumbing question, not a data question, and the reply is one line: what is the shared key, and has anybody mapped these two before? The answer is frequently yes, and frequently from someone who tried it last year and can save you a week.

What none of these replies says is "the data is bad", which is the sentence that ends conversations. Every gap message has the same three parts: what is missing, what it would let you answer, and who can change it. Three parts, one line each, and the third part is the one that turns a complaint into a request.

The order is not a formality. Open the spreadsheet first and something specific happens within about ninety seconds: you see which columns are there, your sense of what is askable rearranges itself around them, and the question you end up answering becomes the question that file supports. The original question is not rejected. It is quietly replaced, and the replacement feels like the original, because nothing was ever written down to compare it against.

Writing the ideal table first is the only way that substitution becomes visible. The sketch is a record of what you wanted made before you knew what you could have, and holding the real table against it is the moment the difference shows up as a list rather than as a feeling. That list is the gap list, and it exists only because the sketch came first.

Be straight about the standing of this, because it is craft rather than a research finding. No study shows that teams who sketch the ideal table produce better analyses than teams who do not, and nobody appears to have run one. What can be said is that every process model in the field independently orders the work this way: CRISP-DM, published by a cross-industry consortium in 2000, puts business understanding ahead of data understanding among its six phases, which encodes exactly this rule. Agreement between process models is not evidence of effectiveness, and it should not be presented as if it were.

When you cannot measure whether customers stayed, you measure whether they ordered again within twelve weeks. That is a reasonable thing to do. It is also, strictly, an answer to a different question, and the whole of the honesty here lives in one habit: you name the substitution in the same sentence as the number, every time, and never in a footnote.

Compare the two sentences. "Sixty-one percent of promotion customers stayed" is a claim about retention that nobody measured. "Sixty-one percent of promotion customers ordered again within twelve weeks, which is what we are using in place of whether they stayed" is the same number with the substitution attached, and the second version costs fourteen extra words. Those fourteen words are the entire difference between an honest proxy and a misleading one.

The reason it has to be the same sentence is mechanical rather than moral. Numbers travel and their footnotes do not. Your caveat survives one slide, maybe two, and then somebody pastes the figure into an email and the qualification is gone forever. If the substitution is inside the sentence, it travels with the number, and by the time it reaches the board the claim is still the claim you made.

When the decision is expensive and hard to reverse. A proxy is a loan against the truth, and the interest is paid by whoever acts on it. For a reversible weekly choice, borrow freely. For hiring a team, signing a year of spend, or telling a board that a channel works, the gap between the proxy and the real outcome is being converted into money that cannot come back.

When the proxy and the real outcome are known to diverge for the group you care about. This is the failure that looks most like success. Email opens are a defensible proxy for engagement across a whole customer base, and a terrible one for the subset who pay by standing order and never open anything. If the divergence lands precisely on the population in the question, the proxy is not weaker than the real thing, it points the other way.

When the substitution would not survive being said out loud to the person who asked. This is the most practical test of the three, and you can run it in your head in five seconds. Say the sentence with the substitution in it, to the founder, in your imagination. If it comes out as "we measured whether they opened an email, and we are calling that staying", you already have your answer, and no amount of careful wording fixes it.