The Ecomm Analyst

I have had ThoughtMetric connected to Claude for a while now. ThoughtMetric sponsors this blog, and I want to be specific rather than enthusiastic, so this is the actual list of questions I ask it, the ones that were not worth the trouble, and where I still open the dashboard instead.

The short version is that it replaced a category of small lookups I used to either do slowly or skip entirely. It replaced none of the analysis.

The Monday questions

Most weeks start with the same four. Spend, orders, and revenue for the week just closed, compared against the week before. ROAS by channel for the same window. New customer orders as a share of total. Anything where week over week moved more than about twenty percent.

That last one is the useful one. Asking which channels moved most rather than reading a table of all of them means the outlier finds me instead of the other way round. It takes two minutes rather than fifteen, and I was never reliably good at spotting it manually, particularly for smaller channels where a large percentage swing is a small absolute number.

The questions that turned out to matter more

The new versus returning split is the thing I now check far more often than I used to, purely because it became easy. New customer ROAS and returning customer ROAS are separate metrics, so asking for both alongside blended ROAS by channel takes one question.

That gap is diagnostic. When blended ROAS holds steady but new customer ROAS is sliding, delivery has drifted toward warm audiences and the account is quietly becoming a retargeting operation. I would have caught that eventually. I catch it about a month earlier now, and a month of misallocated prospecting budget is real money.

The other one is cost per new customer order by campaign. Blended CAC across an account tells you very little once campaigns have different jobs. Per campaign, against a threshold you have already decided on, it is directly actionable.

Ad hoc slices I would never have built a report for

This is where it earns the most, honestly. ROAS by product category for one campaign over a specific fortnight. Revenue by discount code for a promotion that ran across two calendar months. New customer orders by state during a regional test. Conversion rate by UTM source for one landing page.

Every one of those is a legitimate question I would previously have answered by exporting to a spreadsheet, or more often by deciding it was not worth twenty minutes and going with my instinct. The threshold for checking a hunch dropped, and I check more hunches. A reasonable number of them turn out to be wrong, which is the point.

Where I still open the dashboard

Anything I need to look at rather than read. Trends over months have a shape, and a shape is not something a sentence conveys well. If I want to know when a curve bent, I want the chart.

Anything requiring row-level data. The connector returns aggregates grouped by dimensions, so questions about individual customers or specific orders are outside what it can do. Cohort work in particular still means an export.

And anything going in front of other people. If a number is going into a board deck or a client report, I pull it from the dashboard, because I want to have seen it in context with its own axis labels before I stake anything on it.

The habit that keeps it honest

I verify anything surprising before repeating it. This is not scepticism about the tool specifically. It is that natural language makes it easy to ask a subtly different question than the one in your head, and prose reads more confidently than a chart does. A dashboard shows you the date range and the filters. A sentence just gives you a number.

I also name the store and give explicit dates every time, having been caught once by a question that resolved a month boundary differently than I assumed. Two seconds of extra typing, and it removes the entire category of quiet disagreement with the dashboard.

What it did not do

It did not improve the underlying attribution. The numbers are the same numbers, with the same modelling assumptions and the same limitations, and asking them in plain language does not make them more correct. If you do not trust your attribution, a faster way to query it is not the fix.

It also did not make decisions, and I would be cautious about anyone selling that. Knowing that Meta ROAS fell from 3.1 to 2.6 is not the same as knowing whether to cut spend, and the second question involves margin, inventory, seasonality, and what you are trying to do this quarter. That is still the job.

What changed is friction. The distance between having a question and having a number got much shorter, and over a few months that has meant asking more questions rather than fewer. That is worth the five minutes it takes to set up. The rest of what I look at each morning, most of which has nothing to do with this, is in my morning attribution routine.