Algorithm Tax: when the machine buys the wrong customer.

This is the Algorithm Tax, and the first thing to understand is what it is not. It is not a story about a slow website. Your page can score 98 on PageSpeed and you will still pay it in full. It is a story about an automated system that was handed your money and a loose instruction, and filled the gap with whoever was cheapest to reach.

In 2003 a retired NTT infrastructure engineer named Taro Mizushima wrote a line in a private newsletter that defines this concept more precisely than anything written since — and he wrote it about telephone cables, two decades before Performance Max existed.

This is the whole concept. Meta’s Advantage+ does not fail. Google’s Performance Max does not fail. They complete. You told the system to find conversions, or engagement, or clicks, and it found exactly that, in the cheapest population where the signal you gave it could be satisfied. If you never told it that a job-seeker is not a customer, it had no way to know. It is not lazy and it is not broken. It is precise, and it was pointed at the wrong target.

Most people who have heard the term think Algorithm Tax means “the AI optimised for engagement instead of sales.” That is real, but it is only one of four ways the machine can mismatch. Whichever axis you leave open is the one it drains you on.

This is the axis almost nobody guards, and it is the most expensive one for service and product businesses. Search platforms match on language. They do not understand your business. When you bid on a broad or phrase match without a disciplined negative-keyword layer beneath it, you are telling the machine: buy anything that contains these words. It obeys.

Consider three businesses, each with a fast site, sharp creative, and a healthy-looking dashboard:

The third-generation watchmaker repairs heirloom mechanical watches at several hundred dollars a ticket. Without negatives, Google spends his budget on a two-dollar battery swap he doesn’t even offer. The algorithm sees the word “watch,” matches, and bills.

The towing company wants stranded drivers at 2am. Without negatives, it pays for “tow truck for sale” and “how to become a tow operator” — job-seekers and tyre-kickers charged at emergency-service click prices.

The commercial real estate broker lists warehouses to businesses. Without negatives, the budget evaporates on “apartments near me” and “flats to buy” — residential searchers who share the phrase “real estate” and nothing else.

Here is what turns an honest mistake into a structural tax: on both major platforms, the budget-draining settings are switched on by default, and turning them off is deliberately not a single switch.

Meta defaults every campaign into Advantage+ Audience. The detail almost nobody is told: when you “turn it off,” Meta only loosens it partially. The interface shifts to Advantage+ “audience suggestions,” which still lets the system expand beyond your definition. The only way to fully reclaim control is to build the entire audience by hand — define the interests, the exclusions, the geographies, the age bands — so there is no empty space left for the algorithm to fill with its cheapest-engagement segment.

Performance Max is the single largest source of Algorithm Tax on Google. It hands the machine your Search, Display, YouTube, Gmail, and Maps inventory at once, with little visibility into where the money goes and almost no native negative-keyword control. It is a default-on invitation to the tax.

Underneath it sit the demographic defaults that ship set to “include everyone”: no household-income tier, all age bands, all devices, the widest location radius, and an empty negative-keyword list. Each default is individually reasonable and collectively expensive. A high-net-worth tax strategist with no income floor pays for clicks from people who could never afford the engagement. A local emergency service on the default radius pays for clicks two cities away.

The pattern is identical on both platforms. The setting that protects your budget exists. It is simply not the default, and reaching it requires you to do the work the machine was happy to skip on your behalf.

Algorithm Tax does not waste a slice of budget once. It poisons the data the machine learns from. When it spends on the wrong audience and those people don’t convert, it concludes your offer is weak and raises your cost to maintain “efficiency.” When it gets an accidental conversion from a poorly-tracked event, it learns to find more people like that — more of the wrong people. The mismatch trains itself deeper with every cycle.

This is the Training Tax: signal damage that persists even after you fix the settings, because the model already learned the wrong lesson. A loose instruction does not just cost you today’s budget; it teaches an expensive habit that outlives the campaign that taught it.

The fix is not to abandon automation. Used correctly, these systems are powerful amplifiers. The fix is to engineer the constraints before you hand over the budget — to point the machine precisely, so that completion and success finally become the same thing.

Your ads are not failing because the algorithm is broken. They are failing because the algorithm is precise and you under-instructed it. It completed the job you described. The job you described was not the job you needed. The watchmaker did not need battery customers. The towing company did not need job applicants. The broker did not need renters. The machine bought them anyway, because nobody told it not to.

Point it correctly, and the same automation that was draining you becomes the most efficient acquisition engine you have ever run. The constraint is the product.