One Amazon seller's notebook.

What this is

Every report Amazon gives you is a shadow of the account: ACOS on one wall, AMC on another, Search Query Performance on the floor. For years I made decisions from one shadow at a time, like most sellers. The answers kept contradicting each other.

What changed is that AI now lets one operator do what used to need an analytics team: pull the reports together, line up their definitions and time windows, and test a claim from several angles at once. I still decide what to test and what the result means. That part doesn't delegate.

Finite Samples is where I put the results in the open, with the premises showing.

What changed when I could finally test my hunches

For most of seven years, a lot of what I believed about Amazon advertising was a hunch: a pattern I kept seeing, with no way to check it. Then two things arrived. Amazon Marketing Cloud let me see who was behind the numbers, new or returning, and where they met my ads. AI let one person connect that with every other report, week after week, without a team.

So I went back and tested what I thought I knew. Some of it held up. Some of it didn't. And some of it turned out to be the wrong question entirely.

• Confirmed: where a shopper stands in their journey is theirs, not mine. My bid decides how hard my ads reach for the people I can be matched to. It doesn't move anyone closer to buying.
• Confirmed: the same keyword on the same product reads differently depending on who is searching it that week. The number alone is flat; the mix of people is the missing dimension.
• Overturned: I used to explain the auction as second-price with a quality score. Measured against what I actually paid, that model doesn't hold.
• New angles I didn't have before: new versus returning customers by placement and by position within a placement; searches read as intent groups instead of single keywords; whether ads reached new people or just met the same ones again.

Every one of those is in the public ledger, with its verdict and the date it was last checked. The method is the part worth taking: the specific results are mine, and yours will differ.

The attitude

I don't hand out answers. What I offer is a view that has been tested hard and could still be overturned tomorrow. Four words carry it: discussion, data, logic, verification.

• Premises come before conclusions, marked as holding, partly holding or untested.
• Every claim is labelled Experience (measured on my own accounts) or Hypothesis (framework formed, still being tested).
• Every claim names the result that would prove it wrong.
• When I'm wrong, the reversal goes into the public ledger and stays there.
• Mechanisms such as eligibility, the auction, attribution and report definitions work the same everywhere. Category and account size change how visible an effect is and how big it gets, not whether the mechanism exists.

Boundaries

• No face, no real name, no location. The work should stand without them.
• No client, brand, product or category names, and no absolute revenue or cost figures. Ratios and indices only.
• No naming of other educators. Popular claims are judged as claims.
• No tools for sale, no affiliate links, no promised results.
• I'm not here to sell consulting, and I don't solicit it.

Where AI fits in my analysis

I use AI the way an analyst uses a good assistant: to put the reports side by side, line up their definitions and time windows, and run the first cross-checks, so the analysis covers more angles than one person could check by hand. The questions, the verdicts and the disproof conditions stay mine. How to bring AI into your own operating analysis this way is part of what the notes and the course teach. Screens use de-identified or demonstration data.

Talk to me

The most useful thing you can send is data that disagrees with me. Share a case, or reply to any email from the list.