Fragmented asset groups in a global puzzle-game launch

App campaigns are automated: the system assembles headline, description, image and video combinations and learns which ones win, per asset group. That learning needs two things this account lacks.

Asset depth. Three headlines, one description, four images and a single video yield very few distinct combinations and only one video format. The system cannot test hooks, formats or aspect ratios, and ad strength 'Poor' reflects exactly that. With one video, any auction requiring a video format gets the same creative every time.

Signal density. Six asset groups split $1,500 per day into roughly $250 each. At a $1.55 CPI that is about 160 installs per group per week - far too little for the system to separate real creative differences from noise, and several groups have fewer than 15 conversions over three weeks.

The 70% spend skew is not a creative verdict. When signal is thin, the system concentrates spend on whichever group wins early auctions and keeps winning. 'Competitive' getting 70% means it got the first clean read, not that it is the best theme.

Language splitting only helps if the creative is truly localised. If the German groups reuse translated English concepts line-for-line, they add cost and fragmentation without adding a distinct message - and they halve the data going to each concept.

Collapse from six asset groups to two - English and German - and treat theme as creative variety inside each group, not as structure. Keep a theme split only if the themes have genuinely different audience targeting attached; at this budget, two groups is already the practical maximum.

Asset plan for the 12 images and 4 videos:

• Fill every slot. Five headlines, five descriptions, and all available image and video slots, across the aspect ratios the campaign supports. Leaving slots empty is the cheapest performance loss to fix.
• Distribute new assets across both language groups: roughly 8–10 images per group (portrait and landscape) and 3–4 videos per group, reusing the new videos in both languages with localised end cards rather than shooting twice.
• Diversify headline angles, not synonyms. One headline on relaxation, one on level progression, one on challenge, one on social proof, one with a clear call to action. Three paraphrases of 'fun puzzle game' waste slots.
• Localise the German set properly: copy, ratings and style cues, not literal translation. Cultural framing of 'competitive' differs between the two markets.
• Keep the strongest existing creative rather than replacing everything; the new assets should extend the pool, not restart it.

This converts six thin groups into two dense ones, roughly tripling the daily budget and conversion volume behind each learning loop.

Migrate in a sequence that never leaves the account with nothing to serve:

1. Build the new asset groups alongside the old ones. Do not pause first - you would lose demand signal with no replacement and restart learning from zero.
2. Hold the tCPI at $1.30 during the rebuild. Changing structure and bid at once makes the outcome unreadable. Accept a 1–2 week learning period and do not judge on day three.
3. Pause old groups one at a time once the new groups accumulate meaningful conversion volume, watching CPI and volume for 3–5 days after each step.
4. Track the right metrics: cost per install, ad strength, asset-level reporting to see which new creative is actually being served, and D1/D7 retention if it is available - a cheaper install that never returns is not progress.
5. Then, and only then, reassess the bid. If CPI still sits above $1.30 after two weeks of full delivery on the consolidated, well-populated structure, that is real evidence the target is below market rather than an artifact of fragmentation. At that point a stepped increase of 10–15% is justified; raising it before the structural fix confuses the cause with the cure.