Imagine if I gave you a stack of flyers and asked you to stand outside a department store, handing them out to as many people as possible. I tell you that your goal is to figure out who reads the flyers and who throws them away. Over time, I need you to find a way to spot the readers before you hand them a flyer, so you don’t waste any on people who will just trash them.
This is a simplified version of what happens when you launch a digital ad campaign. Like a person would, the platform goes through what’s called the “learning phase.” Its ultimate objective is to show your ads to people who are most likely to take the action you selected as your campaign objective. In technical terms, this process is called machine learning. And it’s incredibly powerful because machines can process data at a much higher speed and larger scale than people can, recognising patterns in the data via algorithms.
Now, what if only five people read the flyer? It’s probably going to be a lot harder to recognise a pattern in those people than if a hundred people read it. With a larger sample, you’re more likely to find trends in the “data” – similarities among the people who actually read the flyers, and how they differ from the people who don’t. Ad platforms learn in a similar way. The more data they have, the better they can understand who to target and how to deliver your ads successfully.
This is why you’ll see platforms mention the need for a minimum number of conversions per week to exit the learning phase. In the background, the algorithm builds a model of who to target, testing its accuracy until it’s confident it knows who will respond most efficiently. Much like you’d form hypotheses and test them while handing out flyers.
So what does this mean for advertisers? All things equal, ad platforms tend to favour campaigns with larger budgets. With more budget, your ads reach more people, giving the algorithm a larger sample of data to work with during the learning phase. But there’s nuance. Lower acquisition costs matter just as much.
For example, if I spend $1,000 and my cost per sale is $5, then the ad platform has 200 customer data points to learn from. A campaign that spends $2,000 but has a cost per acquisition of $200 generates only 10. So while budget matters, what really counts is the relationship between budget and acquisition costs, because that relationship determines how much data the algorithm actually has to work with.
This is an important concept for advertisers to understand, because it reveals that not all advertisers are playing the same game when they run campaigns online. Some business models make it inherently harder to benefit from ad tech – particularly those with smaller budgets or higher acquisition costs.
It’s also important to know because it adds context to a lot of the information you’ll find online about how algorithms and ad platforms work. Most of what I see written about algorithms assumes campaigns with large budgets and high conversion volumes. For example, much of the discussion around Meta’s Andromeda update revolves around testing large volumes of creatives. That approach simply isn’t feasible for small businesses with limited budgets or for B2B industries where budgets can’t be scaled to a point where they’re significantly higher than the cost per acquisition.
Finally, it means that as an advertiser, it’s not always in your best interest to “follow the rules.” Many of the guidelines presented as universal truths are context specific, and recognising this distinction will save you from a lot of frustration and misdiagnoses. For example, ad platforms often discourage targeting and encourage broad reach, urging you to remove limitations and let the algorithm find your audience. While this may work well for some e-commerce businesses operating at scale, companies in niche markets will benefit from applying targeting and exclusions that make sense for their industry. Just as you wouldn’t target the whole world and wait for the algorithm to figure out you only operate in one country, excluding irrelevant segments from the start is prudent, and helps the algorithm shorten its exploration time during the learning phase.
