Most applied analytics fails at choosing the method, not at running it. This page is about that choice.
That sentence sounds obvious and it is the reason most marketing research is weaker than it looks. A survey records what somebody was willing to tell you, some time after the fact, in the language your question gave them. Behavior records what actually happened and never explains itself.
So we use both, deliberately, and we say which is which. When a client hears “customers want X,” the first question is always whether that came from something they said or something they did — because the two carry very different weight.
The buying decision runs through recognizable stages, and the amount of evidence available at each one is wildly uneven. Analytics sees the transaction clearly and the three stages before it barely at all — which is exactly backward from where the decision was actually made.
This is why an assessment that only reads the analytics account will tell you what happened and almost nothing about why. Search terms, review language, competitor comparison and social conversation fill in the earlier stages, and they have to be gone and got.
Nearly every serious mistake we see in applied analytics is a method-selection error rather than an execution error — a predictive model answering a causal question, a significance test run on data that breaks its assumptions, a segment built from a variable that does not drive anything.
Being able to run a technique is not the hard part and has not been for years. Knowing which one the question calls for, and being willing to say the data cannot support the conclusion somebody wants, is the part we take most seriously.
Analytics accounts, event data, search console, server logs, session and heatmap tools, marketplace search data, and the client’s own order history — usually the most underused dataset in the building.
Structured surveys, on-site and exit-intent polls, interviews, review and ratings analysis, and social listening for the things nobody was prompted to say.
Full-site crawls, metadata and structured data audits, page speed measurement, accessibility checks, and retrievability testing — including whether AI assistants can reach the pages at all.
The same instrument run across a population, so a result means something against a comparable set rather than against an opinion. Public advertising archives. Competitor structure read from what is publicly visible.
Split tests where volume supports them, holdouts and geo tests where it does not, and an honest answer about when a business is simply too small to test its way to certainty.
A confounder log kept from day one — what else changed, whose change it was, and whether attribution survives it. This is the sheet that separates a study from an anecdote.
Findings do not get edited on the way to the client. Independence is the whole reason an outside assessment is worth anything, and if a finding gets softened it stops being useful to anybody.
Experience without method produces confident guesses. Method without experience produces a framework applied by somebody who has never had to make payroll from a website. We think you want both, and we are unusual in having taken them in that order.
The operating years came first — a catalog business run through marketplaces, wholesale and direct, with the analytics read every morning because the rent depended on it. The formal training came afterward, deliberately: statistics, data science, research design, consumer behavior. Not to learn marketing, but to find out which of the things twenty-five years had taught would survive contact with a method.
Most of them did. A few did not — and those are the ones worth talking about.
We’ll give you a free AI analysis and tell you the three things we’d fix first — in writing, for nothing.