Data and measurement: what ads really deliver
Lesson 15 of 16 in the course “Marketing 2027”. Free, with animations and a quiz.
The signal: pixel and Conversions API
Everything we said about automation and agents rests on one thing: the platform has to know who bought. The classic way is the pixel, a bit of code on your site that tells the platform about a purchase from the browser. But browsers lose data: some people use ad blockers, and since iOS 14.5 in 2021 Apple has asked users for permission to track them across apps, and many say no. That’s why there’s a second route, Conversions API. Here it’s not the browser that reports the purchase but your store’s server, and far less gets lost. Many store platforms like Shopify connect it in a couple of clicks. Best practice is to use both together and turn on deduplication so one purchase isn’t counted twice.
Every platform claims the sale
Here’s a trap almost everyone falls into. Someone saw a video on TikTok, then an ad on Instagram, then googled the brand and bought. Each platform sees its own piece of the journey and, honestly by its own rules, records the sale for itself. In the made-up example in the picture, Meta counted 70 sales, Google 60, TikTok 30. That’s 160 in total, but the store actually has 100 orders. Nobody is lying; each platform just has its own counting rules, and they overlap. So your main source of truth is your real orders, not the ad dashboards. A simple rule: once a week, add up the “sales” from all platforms and compare with the store. The bigger the gap, the more carefully you should trust the reports.
eBay: paying for people who’d come anyway
A real experiment, published in 2015 in the academic journal Econometrica. Economists Blake, Nosko and Tadelis worked with eBay to switch off search ads for queries containing the brand name, like “ebay sneakers,” and compared that with what happened with ads on. The result: 99.5% of the traffic stayed. People searching for eBay by name just clicked the regular link a bit lower down. In other words, the company had been paying for years for visitors who would have come anyway. Meanwhile the dashboard showed great returns: the ads “brought in” lots of buyers. The lesson: a high return in a report doesn’t prove the ads added anything. That’s especially true for ads on your own brand and for people who’ve already visited your site.
Uber: cut 2/3 of ads and lost almost nothing
Another real story, told by Kevin Frisch, Uber’s former head of performance marketing, on the Marketing Today podcast. First Uber removed some ads because they were showing on questionable sites, and noticed results didn’t drop. So the team kept checking and eventually switched off about two thirds of online ads from a budget of roughly 150 million dollars. There was no significant drop in new user acquisition. It turned out that fraudulent networks faking clicks were taking credit for some of the “installs,” and many people would have downloaded the app anyway. The lesson is the same as eBay’s: a report on how many sales “came from ads” and the ads’ real contribution are two different things.
Lift test: the only honest answer
So how do you find the real contribution of ads? Repeat what eBay and Uber did, but carefully. It’s called a lift test, or incrementality test. People are randomly split into two groups: one sees the ads, the other, the control group, doesn’t. Then you compare what share of each group bought. In a made-up example: 5% bought with ads, 3% without. So the ads added 2 percentage points, not the full 5%: that 3% would have bought anyway. Key point: these tests are right inside the ad platforms. At Meta it’s Conversion Lift, and as we saw, an agent can even launch one through MCP. Google has similar lift studies. A lift test once a quarter costs more than a regular report, but it’s the only thing that answers “would it have been the same without ads?”
MMM: a model for the whole picture
A lift test answers for one platform or campaign. But how do you see how much each platform contributes to the overall mix and where to add money? That’s what MMM, marketing mix modeling, is for. The idea is simple: take weekly spend for each channel over a year or two, plus sales for the same weeks, and a statistical model works out how much each channel adds, accounting for seasons and holidays. This used to be affordable only for big brands. Now there are free open-source models: Robyn from Meta and Meridian from Google, which became available to everyone in January 2025. MMM doesn’t track individual people, so ad blockers don’t affect it. The catch: it needs a lot of data, so it’s the next step once a business has grown.
Do it yourself: an honesty check for your numbers
Take five steps. One: add up the “sales” from all platform dashboards for last week. Two: compare with real store orders. If the platforms’ total is much higher, note the gap and adjust for it. Three: check whether Conversions API, or its equivalent on other platforms, is on, and that purchases aren’t counted twice. Four: if you pay for ads on your own brand name, run a test: switch them off for part of the time or in some regions and see how much traffic stays. Five: plan one lift test a quarter on your main platform. In a couple of months you’ll know more about your ads than most people who only look at the dashboard.
Check yourself
- How should she read this?
Correct answer: Platforms overlap: rely on the 100 orders. Each platform sees its piece of the journey and credits itself. Real orders are the source of truth. - What should she check?
Correct answer: Test: switch off brand ads and watch traffic. As the eBay experiment showed, people searching your brand by name would often come without ads. A high reported return doesn’t disprove that. - What’s the ads’ real contribution?
Correct answer: 2 percentage points: 4% − 2%. Lift is the difference between groups. 2% would have bought without ads, so the ads added 2 points. - What should she do first?
Correct answer: Turn on Conversions API and remove double counting. Without a reliable signal, platforms learn on partial data and every report and test is skewed. Fix the foundation first.
← Your own media-buying agent Final project: a campaign in 7 days →