AI Won't Fix a Bad Campaign Setup. Here's What to Audit First

What is happening

Businesses are enabling Smart Bidding, Performance Max, and Demand Gen at a faster rate than ever. Google and Meta have made AI-powered campaigns the default option, and most advertisers are accepting that default without changing much else. The promise is that the AI will figure out the best way to spend your budget if you give it enough room to learn.

The problem is that the AI is learning from what you give it. And what most businesses give it is not good enough.

What I learned from this

I have audited a lot of accounts over the past two years where teams enabled AI-powered campaigns and watched performance collapse. In nearly every case, the problem was not the AI. It was the setup.

The most common failures I find: conversion tracking firing on the wrong event (a page view instead of a purchase, a click instead of a form submission). A target ROAS set to what the business wanted rather than what the account data could actually support. Asset groups with three headline variations that all say the same thing. Audience signals built from the full remarketing pool rather than actual buyers.

Feed any of that into an AI and it will optimise it. Just not in the direction you wanted. The algorithm does not know that your target ROAS was aspirational rather than grounded. It will try to hit it, restrict spend chasing a number it cannot reach, and slowly starve the campaign.

The brief - the conversion action, the target, the creative, the audience signals, the landing page - is the campaign. The AI just executes it.

What I recommend for your business

Before you touch the bid strategy, audit these four things.

First, your conversion action. Is it tracking a real customer doing a real thing that correlates with revenue? If it is tracking page views or button clicks that do not lead to purchases, your AI is optimising for the wrong outcome.

Second, your ROAS target. Look at your actual account performance over the last 90 days. Set a target that reflects what the data supports, not what you want it to say. If your account converts at 3x, a 6x target will not make it convert at 6x. It will make the campaign stop spending.

Third, your creative. Give the algorithm at least five genuinely different headlines, not five ways of saying the same thing. Different angles, different benefits, different tones. The AI cannot test what you have not given it.

Fourth, your audience signals. Your actual buyer list, not your full remarketing pool. The AI will learn faster from signals that reflect real customers.

Sort those four things before you change anything about the bidding. The technology works. Most of the time it is not the technology that is broken.