Case Study: Winning With Machine Learning in Marketing

By 2018, machine learning in marketing had quietly stopped being a research topic and become something ordinary businesses were already using, often without thinking of it that way. Google’s Smart Bidding tools and Facebook’s Lookalike Audiences were both powered by machine learning models running behind a simple interface, no in-house data scientist required. The scenario below is an illustrative example built around that reality, not a real named company.

The starting point

Picture a mid-sized online retailer selling outdoor gear, running Google Shopping ads and Facebook ads with a marketing team of two people managing bids largely by feel: raising budgets on campaigns that felt like they were working, cutting ones that felt slow, checking in every few days. It was reasonable given their bandwidth, but it left obvious money on the table, since manual bid adjustments can’t react to the thousands of micro-signals, time of day, device, browsing history, that predict whether a specific person is likely to buy.

What changed

The team switched their Google Shopping campaigns to target ROAS bidding, a Smart Bidding strategy that uses machine learning to set a different bid for every single auction based on the likelihood that particular impression converts, rather than one static bid applied broadly across a whole campcampaign. On Facebook, they built a Lookalike Audience from their existing customer list, letting Facebook’s model find new prospects who statistically resembled buyers who had already converted, rather than guessing at interest-based targeting themselves.

Why it worked better than manual management

Neither tool required the team to understand the underlying model. What it required was feeding it clean data: an accurate conversion pixel, a customer list free of duplicates and test orders and enough historical purchase volume for the algorithm to find real patterns rather than noise. The team’s actual job shifted from setting individual bids to making sure the inputs were trustworthy, a smaller but more important task.

What the results looked like

Within about two months, cost per acquisition on the Shopping campaigns dropped meaningfully as the algorithm learned which specific product and audience combinations converted best, reallocating spend toward them automatically and continuously, something the two-person team could never have done by hand at that granularity. The Lookalike Audience campaign, run alongside their existing interest-based targeting, consistently outperformed it on cost per purchase, a direct result of the model working from actual customer behavior instead of assumptions about who might be interested.

What it didn’t fix

Machine learning bidding didn’t rescue underperforming products or bad creative. Campaigns promoting items with weak margins or unconvincing product photography still struggled, because the algorithm was optimizing toward conversions that depended on factors outside the bidding system’s control. The team learned quickly that automated bidding amplifies what’s already working and just as efficiently amplifies waste on what isn’t, so it doesn’t remove the need for good offers and creative underneath it.

The broader lesson

The retailer’s experience reflects what a lot of businesses found in 2018: the machine learning payoff didn’t come from building anything custom. It came from adopting the automation already built into the ad platforms most businesses were using anyway, and taking seriously the unglamorous work of keeping the data feeding those systems clean. That’s a much lower bar than “hire a data scientist,” and it’s why so many ordinary retailers, not just tech-forward ones, saw real returns from machine learning this year.