← ResourcesCampaigns6 min read
From exploration to traction: a campaign's first eight weeks
A new campaign does not start at its best. It starts by learning. Here is what happens before launch and across the first four to eight weeks, and why it follows that order.
Why a campaign starts expensive
A new campaign has no history of its own. The first impressions are bought to learn: which apps, placements, formats, and moments produce the users you want. That is exploration, and it is why cost per install starts high. It is the cost of learning, and the job is to keep it short.
Before launch: automatic setup
The work starts before the first impression is bought. Our models run a pre-analysis on your non-attributed data (NAD): the install and event data from your app that is not attributed to our campaigns, meaning your organic users and the users other channels bring in. They segment those users, find the patterns that separate your valuable users from the rest, and use them to point the campaign in the right direction from the start. All of this is automatic.
Weeks 1–2: early CPA optimization
The models start at the top of the funnel. Installs and the first in-app actions arrive quickly and in volume, so they are the first events with enough data to optimize toward. CPI falls fastest in these two weeks as bidding moves away from supply that does not convert.
Weeks 3–4: mid-funnel optimization
With enough early events in hand, the models move up the event stream to actions that say more about user quality: the ones that sit between an install and a purchase. Around the third week CPI flattens out. From here the goal is no longer cheaper installs, it is better ones, and ROAS begins to move.
Week 5 onward: traction and ROAS optimization
Revenue events are the rarest and the slowest to arrive. By about week five there are enough of them for the models to optimize toward ROAS directly. CPI stays roughly flat while ROAS climbs progressively, week over week. That is traction: the campaign is past learning and into compounding.
Why the order matters
Each stage produces the data the next one needs. Optimizing toward ROAS on day one would mean learning from a handful of purchases, which is noise, not signal. So the models optimize toward earlier events first and move through and up the event stream as the data allows. The exact timing depends on budget and on how quickly your events arrive, which is why the range is four to eight weeks.
How non-attributed data shortens exploration
NAD does two jobs. Before launch it drives the setup, so the campaign does not start from zero. After launch we keep feeding it into the models alongside campaign data, so they learn from all of your users and not only the ones we bought. Both shorten exploration and bring the move up the funnel forward.
What you see along the way
All of it happens in the same console our team uses. You can follow CPI and ROAS week by week, and the weekly review covers which stage the campaign is in and what moves it to the next one.