The Ecomm Analyst

For years my analysis workflow had a step in the middle that I never thought about. Question comes up. Open the dashboard. Realize the view I need does not exist. Export to CSV. Open the sheet. Pivot. Get the answer. Close the sheet. Never open it again.

I did that hundreds of times without noticing it was a workflow. It just felt like what analysis was.

The exports were not analysis

When I finally looked at what I was exporting, almost none of it was a real project. Real projects are worth building properly, and I still build those properly. What filled the folder was one-question detours. Which campaigns did we spend more than a thousand on last month that produced fewer than five new customers. What does new-customer ROAS look like by channel if I exclude brand search. Did that discount code actually pull in first-time buyers or just give margin away to people who were buying anyway.

Every one of those is a two-minute question with a five-minute retrieval cost. Which means in practice I asked maybe a third of them and let the rest go. Not because they did not matter, but because the answer was not worth the interruption in the moment. That is a bad trade to make repeatedly, and I made it constantly without registering it as a decision.

What actually changed

The change was connecting my analytics platform to an assistant so I could ask the question in a sentence and get the number back. ThoughtMetric, which sponsors this blog, publishes a connector that exposes its metric and dimension catalog, so I can ask for cost per new customer order grouped by campaign and it resolves against a defined metric rather than me rebuilding the calculation in a sheet. Several other platforms in this category publish something equivalent now.

The mechanical benefit is obvious and not that interesting. The behavioral one is the part worth writing about. I ask more questions now. Not better questions, necessarily. More of them. The half-formed hunch that used to die because checking it required a detour now gets checked, and maybe one in six turns into something.

That is the real return. Not time saved on exports I was doing anyway, but questions asked that I previously swallowed.

What I still export

Plenty, and I want to be clear about that because the pitch for this category tends toward never touch a spreadsheet again, which is nonsense.

Anything I need to hand to somebody else still gets exported. Anything going into a board deck or a client report gets exported, because a conversation transcript is not a deliverable. Anything involving data the platform does not hold gets exported, since the connector only knows what has been ingested. And anything where I need to see the full row-level detail rather than an aggregate still goes to a sheet, because scanning two thousand rows for the thing that looks wrong is a genuinely visual task and chat is a bad interface for it.

Roughly, the ad hoc questions moved and the deliverables did not.

The habit that keeps this honest

An assistant will give you a wrong answer in exactly the same confident register as a right one. I have watched one silently use the wrong date range and hand back a number with no hedging whatsoever. It was not close, and nothing about the response suggested anything was off.

So the rule I run is simple. If a number is going to change what I do, I verify it in the dashboard before I act. If it is just informing my sense of things, I let it stand. That sounds like it would negate the speed advantage, but it does not, because most questions fall into the second bucket. The verification tax only applies to the small share that actually drive a decision, and those deserved a second look regardless of where the number came from.

The other habit is asking for the breakdown rather than the summary. If I ask which channel performed best, I get an answer shaped by whatever the assistant decided best means. If I ask for new-customer ROAS by channel for the last six weeks, I get numbers and I do the judging. The second is almost always what I want. Keep the assistant on retrieval and keep the interpretation yourself.

None of this replaced the routine I run every morning, which is still the same set of checks in the same order. It replaced the interruptions in between. I wrote about the morning version of that in this post on my attribution routine, and that part has not changed at all.