Most placement advice is someone else's anecdote wearing a confident tone. Experiments are how you replace anecdote with your own numbers — and AdSense ships a native A/B tool that splits traffic automatically. Most publishers never open it. The ones who do stop arguing with themselves.
What's worth testing first
Rank experiments by expected information: Auto Ads on/off (the biggest swing), anchor ads on/off, unit count on long articles (2 vs 3), the end-of-content unit present/absent, and vignette aggressiveness on mobile. Placement micro-moves — a unit 50 pixels lower — produce micro-results; test the decisions that change the architecture, not the decorations.
Running one properly
One variable at a time. A 50/50 split unless traffic is thin (then 80/20 toward the incumbent). A fixed window of two to four weeks regardless of early noise — the first week's numbers lie routinely, especially with weekly traffic rhythms. Decide on RPM as the primary metric, with bounce rate and pages/session as the guardrails. A placement that lifts RPM 15% while cratering session depth 30% isn't a win; it's a loan against future traffic.
The traps
Seasonality crossing your test window (don't run through Black Friday unless testing seasonality). Declaring victory at 90% confidence on a Tuesday and reversed by Thursday. Changing the template mid-test. And testing so many things at once that the result explains nothing — the multi-variable experiment produces data, not decisions.
The compounding habit: one clean experiment per month, documented in a running file — date, variable, window, result, decision. Within a year you own something no consultant can sell you: your site's actual response surface. Every later decision gets cheaper.
