A Note on This Column
The Auditor's Digest publishes cases. Real situations, anonymised, diagnosed.
The Auditor's Diaries is different. These are personal notes. Things I have been reading, thinking about, returning to. I have kept notes like these since around 2004 when I first understood that the engineering problems I was being paid to solve were almost never the problems the person paying me thought they were. I am putting them here now because they belong here and because some of them are old enough that I cannot keep calling them notes.
This is the first entry.
What Mizushima Wrote
Taro Mizushima worked at NTT. Nippon Telegraph and Telephone. from 1984 until he retired in 2002. He was an infrastructure engineer in the era when infrastructure meant physical things: cable routes, switching stations, the topology of a network that was being built in real time without a complete map of where it was going. He was not famous. He did not write books. After retiring he circulated a private newsletter called Keiro. which translates roughly as pathway or course. among former colleagues and a small number of engineers he had corresponded with internationally. The newsletter ran from 2003 to 2009 when he stopped, without announcement, in the way people of his generation stopped things.
I found a translated excerpt in an engineering forum in 2019. Someone had posted three paragraphs from what they described as the eighth issue, translated by a colleague who had studied at Kyoto and later worked in Germany. I have no way to verify the translation is precise. I have read it enough times that the question no longer seems important.
The excerpt, as I have it:
"The machine does not waste. The machine completes. It receives an instruction, it follows the instruction, it reports completion. The waste belongs to whoever pointed it. We say the machine failed when what we mean is that we described success incorrectly and the machine was too precise to know the difference."
I have thought about this paragraph for six years.
Why This Matters in 2026
Mizushima was writing about physical network infrastructure in Japan in 2003. He was not writing about Google Ads. He did not know Google Ads existed. The sentence applies with an accuracy that should embarrass everyone who has ever described a campaign as underperforming when what they meant was that they had pointed it incorrectly.
Performance Max does not fail. It completes. It receives an objective. maximise conversions. and it pursues that objective with the full weight of Google's machine learning infrastructure applied to your specific budget. It finds conversions. The conversions it finds are the ones it can see, measured the way the tracking was set up to measure them, in the populations where the data supports the signal it was given.
If the tracking is broken, it optimises toward a ghost signal. If the landing page is slow, it learns that certain audiences do not convert and moves away from them. If the conversion event is a form submission when the actual client calls directly, it learns that calling clients do not exist. It completes. The instruction was wrong. The machine was too precise to know the difference.
Mizushima's sentence is the cleanest definition of the Algorithm Tax I have encountered and he wrote it twenty-three years ago about telephone cables.
The Question He Did Not Answer
The excerpt ends there. Three paragraphs, one central observation, no solution offered. This is characteristic of the Keiro excerpts I have found. Mizushima diagnoses, he does not prescribe. He was an infrastructure engineer, not a consultant. He had spent his career building systems that worked. He seemed to regard explaining how to build systems that worked as someone else's problem.
The question his paragraph opens but does not answer is this: if the machine completes and the waste belongs to whoever pointed it, what does pointing correctly require?
My answer, built from fifteen years of auditing campaigns and sites and tracking setups that were wrong in precisely the ways Mizushima described without knowing it: pointing correctly requires knowing what you are measuring, why you are measuring it, and whether the measurement corresponds to the thing you actually need. This is not a technical question. It is a prior question. The technical implementation follows from it.
Most campaigns skip the prior question. They go directly to the implementation and discover, three months and several thousand dollars later, that they built a precise instrument pointed at the wrong thing.
Mizushima called this pointing incorrectly. I call it the Algorithm Tax. The machine would call it completion.
The Auditor's Diaries is a personal column on vsourcecode.com. These are reading notes and engineering observations accumulated over fifteen years of practice. The cases in the Auditor's Digest are anonymised real situations. The Diaries are how I think about them.
Ready to Fix This on Your Site?
Run your site through PageSpeed Insights and send me the score. Free diagnosis. No pitch. Just the numbers.