Push Creative Testing With a Stopping Rule, Not a Grid

Most push creative tests are decided by noise. At a click rate near half a percent and a landing conversion near two percent, a variant needs tens of thousands of impressions before its cost per conversion carries information, and a sixty-cell grid at that threshold costs more than the improvement it is looking for. We run six to eight concepts, kill them against a rule written before the money moves, and put the survivor against one new challenger every fortnight. The icon does the work: Networks accept an icon at 192 by 192, a main image at 360 by 240, a title near thirty characters and a description near forty-five. On a large share of Android builds the main image stays collapsed until the user pulls the notification down, which most never do. So the icon and the first four words of the title are the advertisement, and the image is the part that absorbs the design budget. Our concepts vary the icon first and the image last. Macros, used narrowly: City and brand macros raise click rate reliably and often lower conversion, because they buy curiosity from people whose city merely appeared in a line of text. We use a geo macro where the offer is genuinely local: a departure airport, a utility service region, a branch. We refuse it as decoration. Browser and battery macros we do not use at all, since their only function is to suggest the notification came from the device. The stopping rule, agreed before launch: No variant is judged before it has spent three times target cost per acquisition. At that point variants running above twice target are killed, variants between are held, and the leader continues only if its interval separates from the runner-up. This is written down in advance because a test that can be reinterpreted afterwards is not a test, and push data invites reinterpretation more than most. Burnout is scheduled, not discovered: A winning creative on a fixed source list typically loses around half its click rate within ten to fourteen days, faster on a small subscriber base where the same devices see it repeatedly. The rotation calendar comes from your own decay measurement in the first cycle rather than from a rule of thumb, which is why this is standing work rather than a delivery of files. Refused without discussion: Icons imitating a messaging application, a bank, a courier, an operating system dialog or a lock screen. Titles claiming a message, a delivery, an infection or a payment that does not exist. Countdowns that reset. Shock imagery. Competitor marks. These beat our creative in head-to-head tests and we will show you by roughly how much, from campaigns already run, before declining to make them. Not for: Teams who want a creative library delivered with the testing done in-house. The value here sits in the kill decisions, and handing over files without them is selling the wrapper. Also not for offers with no server-side conversion signal, where the only measurable outcome is the click and the exercise collapses into picking whichever image misleads most efficiently. What arrives: Production files at each network's specification. The test log with spend, clicks, conversions and interval per variant. Every decision with its date and its reason. The rotation calendar for the following quarter. The kill list, including what each dead concept was testing, so the next round does not repeat it. Written in English or Mandarin, whichever your team works in.