This is the mental model I would walk into almost every analysis with: frame the decision, hypothesize the answer, source the data, analyze the signal, translate it into action, and close the loop. The order matters. Most analysts skip the first two phases and jump straight to the data. That is usually where the work goes wrong.
Most bad analysis is not bad because the math was impossible. It is bad because the work started too late in the chain. Someone opened a query window before clarifying the decision, never wrote down a prior, pulled mismatched data, and then reverse-engineered a story after the fact. The better pattern is to treat analysis as a six-phase workflow that keeps the work decision-first from start to finish.
The details change by project. The scaffold should not.
This is where the work actually gets won or lost. Before touching the data, I want a one-sentence problem statement I could repeat back to the stakeholder and get an immediate nod. The goal is not just understanding the question. The goal is understanding the decision the analysis needs to serve.
A fast 90%-confident answer in four hours is often more useful than a 99%-confident answer four days later. That is why framing is such a leverage point.
Before opening the data, commit to a prior. Write down what you expect to find, what the no-story outcome would look like, what would surprise you, and the top alternative explanations if your first guess is wrong. That is what keeps confirmation bias from quietly taking over the minute the numbers load.
This is one of the clearest differences between an analyst and a reporter. A reporter describes what happened. An analyst arrives with a theory, tests it, and is willing to be wrong in a disciplined way.
This phase has three jobs: map, pull, and validate. First map where the answer actually lives: ServiceTitan, CallRail, GA4, ad platforms, warehouse tables, or exported ops data. Then note the grain of each source, because session, lead, booking, job, and customer are not interchangeable. Grain mismatches are one of the fastest ways to get a confident wrong answer.
Next, pull the data in a reproducible way. A script is better than a one-off export because it lets the logic survive. Then validate before analyzing. Check row counts, date ranges, totals, and whether the reporting view reconciles to the source system of record. If the totals do not line up, the analysis has not started yet. It is still a data-quality problem.
This is also where cleaner attribution becomes possible. If the sourcing layer is weak, the later attribution story will be weak too, which is why I think of posts like practical attribution and tracking governance as downstream of disciplined sourcing.
I like to keep this phase in a strict order: explore, diagnose, isolate. First explore the movement with time trends and the natural dimensions of the problem: channel, brand, geography, device, offer, service line, or time-of-day. Then diagnose by testing the original hypothesis and trying to disprove it before you explain it away.
If the prior was "Meta CAC rose because of creative fatigue," the first question is whether CTR actually fell. If it did not, maybe the real issue was higher CPM, worse lead quality, or lower booking conversion. The point is to rule out the obvious alternative explanations before claiming a finding.
Then isolate the lever. Decompose the metric until it stops being vague. CAC is not one thing. It is spend divided by new customers, and new customers are downstream of impressions, CTR, conversion rate, close rate, and the quality of the operational handoff. "CAC rose 23%" is not actionable. "CAC rose because mobile lead quality dropped after a creative change and the lead-to-book rate broke at one funnel stage" is.
If you cannot prove causation, say so. Strong analysis is not false precision. It is clear signal with honest limits.
This is where the analysis stops being interesting and starts being useful. Lead with the conclusion, then the two or three supporting points, then the detail. If the decision-maker reads only the first sentence, they should still know what happened and what to do next.
This is also where operating artifacts like a growth baseline scorecard or a Friday growth brief earn their keep. They are not the analysis itself. They are the delivery mechanisms that make Phase 5 repeatable.
This is the phase many analysts skip, and it is one of the clearest differences between a reporting analyst and a marketing ops analyst. First, track whether the recommendation actually got executed. If it did not happen, that matters as much as the original finding.
Then measure the result against the expected impact. Did the move deliver what you thought it would? If yes, the model gets a little more trustworthy. If no, the model or the assumptions need revision. Finally, update the priors. That is how real intuition develops over time: not mystical instinct, but many calibrated cycles of prediction and feedback.
Take a question like "why did one plumbing brand's leads drop last week?" The framework should immediately change how you approach it. You do not start with the chart. You start with the decision: are we reallocating budget, pausing a vendor, or leaving the media plan alone?
Same problem, six different kinds of thinking in sequence.
That is the shape of good analysis. The details will change. The scaffold should not.