Read the denominator before the headline.
In this invented example, an initial audience sees the thumbnail ten thousand times and clicks at eight percent: eight hundred clicks. A broader audience sees it a hundred thousand times and clicks at three percent: three thousand clicks. The rate fell. The count rose.
This arithmetic does not prove why a real video expanded, or which recommendation signals mattered. It does show why “CTR went down” is not a complete diagnosis. YouTube’s impressions and CTR FAQ, checked on 9 September 2026, explains that traffic sources and wider distribution can change CTR.
What the number can tell you.
Impressions multiplied by CTR gives the views attributable to those counted impressions. It is not necessarily all views on the video: some traffic does not enter the same impression count. Use matching reporting periods and sources. Treat this lab as a clean example of denominator effects.
Inspect in a useful order.
- Exposure: inspect impression counts by traffic source. Very little exposure gives you little evidence about how an audience responds.
- The click: compare CTR within similar traffic sources and audience conditions. A blended number can conceal a changing mix.
- The watch: examine average view duration and where people leave. Compare similar videos on your own channel.
- The hypothesis: write one possible explanation and the observation that would challenge it before changing the video.
A diagnosis is still a hypothesis.
“Not seen,” “seen but not clicked,” and “clicked but left” are useful questions from the book’s autopsy framework. They are not an externally validated classifier. A dashboard cannot reveal every reason a viewer acted, and a single upload cannot establish a universal rule.
If you test packaging, write down what changes and how you will interpret the outcome before starting. The goal is to learn from the test, including an inconclusive result. Treat each dated result as one observation in a growing record, rather than a new law of the platform.
Adapted from Nobody Pulled a Lever, chapters 4–9 and Appendix B, the flop autopsy. Platform guidance checked 9 September 2026. All audience counts and CTRs in the lab are fictional teaching examples.