Bad question = bad answer. How to ask your data.
Most questions people ask their data are bad. They don't get an answer not because the data lacks it, but because they aren't looking for it the right way. Here are four patterns that lift the odds of a useful answer — they work equally whether you're asking AI or stitching it together in Excel.
1. Concrete time window
Bad: "How did Meta do?"
Better: "How did Meta do over the past 14 days vs the 14 days before?"
Why: without a reference window, "doing well" is subjective. Specify the window and you get a comparison, not an impression.
2. Define what "good" means
Bad: "Which campaign is the best?"
Better: "Which campaign has the lowest CPA among campaigns spending at least $1,000?"
Why: the system has to pick what "best" means — it will default to ROAS, but the best ROAS on a campaign that spent $50 is statistical noise, not a result. Add the criterion plus a minimum-weight floor and you get a realistic answer.
3. Ask for the cause, not the number
Bad: "What's my ROAS?"
Better: "Why did my ROAS drop 30% last week?"
Why: a number without context doesn't drive action. "ROAS is 2.8" is information, not instruction. "ROAS dropped because the main campaign spent its budget by Wednesday and didn't run through the weekend" is an answer you can act on.
4. Name the segments
Bad: "Who's buying from me?"
Better: "What's the difference between customers from brand campaigns and prospecting — average order, repeat purchase?"
Why: averages mislead. If average order value is around $60, you might have half your customers at $20 and half at $100 — and each group wants completely different marketing.
These patterns don't depend on the tool — they work equally with GPT, Lupli, a dashboard, or Excel. The key is the question, not the software. A bad question gets you one of two things: a mediocre answer, or an answer that sounds right but isn't true. Data, like people, gives quality conversation only when you bring a quality question.