Jukebox: day 0 ROAS more than doubled in eight weeks

Where it started

In the month before the work the account spent on 50 different campaigns outside its creative and web tests, up to 34 in a single week with 63 ad sets between them. Almost all were worldwide campaigns optimized for purchase count, with several languages in each. Spread that thin, not one ad set reached 50 purchase events a week, Meta’s own count of purchases and the number it looks for before an ad set stops learning, and Meta sent the money to the countries where a purchase was cheapest. Inside one of those campaigns the best quarter of countries returned more than four times what the worst quarter did on Meta’s 7 day numbers.

Day 0 ROAS on the main campaigns had fallen by about 30 percent since the start of summer, and Meta’s weekly purchase ROAS had halved since spring. Jukebox sells weekly and yearly plans plus extra credits, so a cohort’s return keeps growing for months after install, and every day 0 figure on this page is an early reading.

How we worked

Their team ran the account and the product. I worked from their Meta and Adjust data and sent a list every Monday from week 3 with the ads to stop and the budgets to move, and the following Monday I checked it against what had gone live.

Day 0 ROAS, step by step

The numbers

Across two four week windows, the second starting at the end of the first month, Meta reported 51 percent more purchase value on 13 percent less spend. Purchase ROAS rose 73 percent, iOS 65 percent and Android 83 percent. The number of purchase events stayed about the same, so the gain came from what each purchase was worth.

Step 1: from 50 campaigns to six

In the first week their team built six campaigns to my plan. iOS and Android ran apart, each split into the US, tier 1 markets and the rest of the world, with languages grouped inside. All six bid on purchase value, went live the next day and took over almost all the spend. By week 2 no main campaign bid on purchase count any more. In 13 matched comparisons from the account’s history, the same market and platform with different bidding, value bidding beat purchase count bidding by week 4 every time.

The ad sets came down with the campaigns, from 63 to 11 by week 4, and the typical one went from about 10 purchase events a week to 60 or more, so Meta had enough in each to learn from.

Cost per install roughly tripled, because the cheap installs came from countries that rarely paid. Day 0 ROAS rose by about a quarter before anything else on the account changed, and the two campaigns we left as they were fell over the same weeks.

Step 2: the budget cut

In week 4 the team cut the budgets of five campaigns. I placed the cut campaign by campaign from the cohort data, and day 0 ROAS held. The next day six ads came out, each of them behind the rest of its ad set in every full week on Meta and at day 0 and day 7 in Adjust. Three of the six were the same creative in three ad sets. On this account Meta did not take that money back by itself. Five of the seven biggest losers in the first month still got at least half their spend the following week, so the stops went on the Monday list by hand.

Step 3: the paywall, and four weeks of Monday lists

In week 5 their team moved the annual plan to the front of the paywall. An annual buyer pays for the whole year on the first day, so part of what followed is timing, close to half of the day 0 rise on the cohort curve below.

The paywall was one of four things that landed in those weeks. The rebuilt campaigns finished their learning on value bidding, with 8 of the 11 ad sets past 50 purchase events a week by week 4. Losing ads came out every Monday. Budget moved on 14 day returns against the line for paying back within 90 days, always in a single step. And new concepts went through the gate before they touched a main campaign. Day 0 ROAS stood at 2.2 times the starting level in the first two weeks after the paywall and 2.6 times in the two weeks after that, and Meta’s weekly purchase ROAS climbed three weeks running. The paywall changed once, in one week, and the return kept climbing for three weeks after it, which is what the lists were for.

Return keeps growing for months

A cohort keeps paying long after install. On the cohorts from the spring before the work, return at 7 days was about twice day 0, at 28 days 2.7 times, and at 16 weeks 4.3 times, still rising when the data ends. The annual paywall changed the shape. More of a cohort’s money now arrives on day one, so the cohorts after it stood at 1.5 times day 0 by 14 days, against 2.0 before, and both lines are in the chart.

New creative went through a gate

Between April and early September 58 concept codes spent money on the account, and 15 of them reached the main campaigns without a test. New ads went straight into the main ad sets and took budget from the ads that were already working there.

From week 5 every new concept went through a cheap install test and then a purchase test. From week 6 the ones that passed got an ad set of their own inside each main campaign before they joined the proven ads. New ads reaching the main sets dropped from dozens a week to a handful. Translations needed a rule of their own. About two thirds of the translated versions opened in the first month barely ran, and an ad’s English result said nothing about how its translations would do, so I asked the team to translate an ad only after it had proven itself.

Two good weeks before a concept scales

The passing bar came from the account’s own history. About three in four ads that beat their peers by more than 20 percent in one week stayed ahead of them over their lifetime. Among ads that did it in two separate weeks, 91 percent stayed ahead. The first stage of this kind of test is written out on this site.

What the numbers include

Meta’s purchase event on this account also counts renewals, so its ROAS is best read as Meta’s own score. Both windows fall inside the work, and the lower spend was the team’s decision. The rise from week 5 came in the same weeks as their paywall, the campaigns finishing their learning and the weekly lists, and no test on the account can split the three.

What it says

Their team made every change in the account, from a list I wrote each week out of the Meta, Adjust and paywall data. That split suits a team that already has people in the ad account and needs someone to make the calls.

In the client’s words

We were spending a six-figure monthly ad budget on Jukebox and could not clearly see whether the money was coming back. In the two months we worked with Samet the campaigns were rebuilt by market and platform, and for the first time we could see which country and which campaign paid its money back.

In numbers: Meta reported a purchase ROAS of 0.32 in the four weeks before the new structure and 0.55 in the four weeks after, with 13% less spend. In Adjust, day-0 ROAS on the main campaigns went from 0.25 to 0.46. Part of that came from our own product work, we pushed the annual plan harder in the same period. Samet helped us read the two together, so we could tell what came from the campaigns and what came from the paywall.

He is someone I trust. He has always been the first person we call.

Questions people ask