The machine hits the target you give it

In October 2025, Meta told investors that more than $60 billion a year of its ad revenue runs through campaigns its AI builds and manages end to end.

In those campaigns the machine picks the audience, the placement and the bid. One human choice is left that outweighs all the others: what the machine is told to count as success.

Give it a target and it will hit it, whether or not that target is what the business needs.

So which conversion should ads be optimised for? The deepest event in the sale that still happens often enough, and soon enough, for the platform to learn from it. Google's own guide for lead generation says the same: "Optimize to deeper funnel conversion actions where possible."

The machine does exactly what it is told

Google puts it without decoration: "Your inputs, especially accurate conversion data, control the output of AI-powered campaigns."

Read that as a warning. The system has no idea what a customer is. It knows which events it was asked to produce more of, and it is very good at producing them.

The wrong target can cost an estimated third more

Four researchers measured it in a randomized experiment on more than 70,000 advertisers on Facebook and Instagram. Same campaigns, two targets: purchases, or clicks.

Figure 1

Aim at clicks: an estimated 31% more per customer

Estimated cost of one added customer

Aimed at purchases$38

Aimed at clicks$50

Aimed at purchases, the median advertiser paid about $38 for each added customer. Aimed at clicks, the authors estimate, the same campaigns would pay about $50.

Nothing else changed. The machine was pointed at an easier event, and it delivered that event at the expense of the one that pays.

The right target is rare, and it arrives late

Here is why so many accounts chase the easy event anyway.

Figure 2

A thousand leads, twelve customers

B2B software, paid search

Leads1,000

Marketing-qualified360

Sales-qualified94

Opportunities36

Customers12

At the stage rates First Page Sage published for B2B software in 2025, a thousand leads from paid search become about twelve customers. The machine learns from the thousand. The business lives on the twelve.

And the twelve take their time. Dreamdata measured an average of 219 days from a first conversion to revenue for LinkedIn advertising. Google recommends a signal that arrives in less than seven days.

So the platform gets the signal that is plentiful and fast, and learns to find people who fill in forms.

The fix exists. It takes work nobody has time for

An advertiser's own records hold what the platform cannot see: which leads were qualified, which became customers, what each was worth. Every large platform has a way to receive them.

The randomized evidence is the experiment above, read the other way: campaigns that kept the deeper signal paid about $38 for a customer where the others would pay $50. LinkedIn reports, for its own product, a 20% lower cost per action across 829 campaigns that sent their conversion records back.

When it goes undone, the reason is rarely strategy. It is labour. Every lead has to be tagged at the click, followed through the sales system, and returned to each platform in that platform's own format, on time, forever. It is a job with no owner, on a team with no spare hours.

That is the ceiling from the first article, in its plainest form. Quality is capped by hours.

What AI changes

Run acquisition as one loop and the signal stops being a side project. The outcome travels back as part of the job, because the same system that launched the ad is the one that hears what happened to the lead.

The hours come back too, and quality has two more uses for them: research on the buyer before the brief is written, and a different message for each kind of buyer in place of one message for everyone.

A machine that is told the truth about customers, and given ads worth running, has something worth scaling. That is what quality means in the formula.

What to do on Monday

1. Write down what each campaign is told to count. Open the settings and list the conversion event for each one.
2. Follow that event to a customer. Take last quarter's count of that event and the number of customers it produced. The ratio is your own version of the funnel chart.
3. Move the target one step deeper. Choose the deepest event that still happens at least 15 times a month and within days of the click, and send it back to the platform. A qualified lead is usually within reach when a closed sale is not.

A good signal tells the machine what a winner looks like. It still has to be shown enough candidates to find one. That is velocity, and the next article.