Skip to content

t1k:marketing:monetization:max-floor-performance

FieldValue
Modulemonetization
Version1.49.21
Effortlow
Tools—
/t1k:marketing:monetization:max-floor-performance
[--weeks N] [--requests] [--exclude PKG]

Read-only report on the apps running the personalized bid floor. It never writes to MAX (Report Key only, no Management API). To change floors use t1k-marketing-monetization-max-bid-floor.

IntentPath
“How are the floor apps doing now / this week”Weekly report
“Is the floor actually earning more, with a number”Uplift
“Are my 7 rungs earning their slots / any dead rung”analysis/rung_uplift_monitor.py --days 7 (never posts without --post)
“Does the game load 2 rungs or all”--requests flag below
Terminal window
python3 -I .claude/skills/t1k-marketing-monetization-max-floor-performance/scripts/weekly_report.py --requests

Needs APPLOVIN_REPORT_KEY (env, or --env-file, or APPLOVIN_REPORT_ENV_FILE; this studio keeps it in MarketingApplovinMonet/.env). Never print it. One MAX call, 6 weekly windows ending yesterday (API max range 45 days). Output groups:

GroupMeaning
floorrungs 2+ got impressions in the latest week
inerthas a floor config in the max-bid-floor skill’s apps/*.json but no rung 2+ impressions (client not live)
rampinglaunched (first day with 100+ impressions) under 28 days ago; listed per app, never summed
controleverything else, no floor config

Pass --exclude <package> for an app whose own decline would skew a group (e.g. an unexplained drop); say so in the report. Then read it per the rules below and write the answer in plain words: eCPM and revenue per group, latest week vs the week before and vs week 1, the weak apps, and what is NOT proven.

Reading it (the mistakes this skill exists to stop)

Section titled “Reading it (the mistakes this skill exists to stop)”
  1. The game does not load all rungs. It loads the auto unit, reads its revealed price, then requests only the 1-3 rungs just above it (client ThirdPartyServices FloorGrid.Resolve; the serving-model reference of the max-bid-floor skill). With --requests, rung 1 should dominate and high rungs be small; ~1x on every rung means a different client.
  2. Never estimate uplift as rung eCPM minus auto average eCPM. The real alternative is the auto unit’s revealed price on that request, below the rung’s floor; the MAX report cannot give it. Rung 1 counts as part of the floor. Report a range, not a point.
  3. Revenue moves with traffic. A +100% revenue week on an app with +100% impressions says nothing about the floor. Compare eCPM first, then say what happened to impressions.
  4. Sample size. A group or app under ~1,000 impressions a week is noise; the script flags it. 4a. A blended number says almost nothing. Any eCPM, revenue or rate pooled across segments (country, format, platform, app, ad unit/rung, network, user cohort, new vs returning) moves when the MIX moves, even if no segment’s price changed, and it can move opposite to every segment (Simpson’s paradox). Floors are priced per country x format, so read them there. Before any claim, split the change into mix vs within-segment price: mix = sum((share_new - share_old) * eCPM_old), price = sum(share_new * (eCPM_new - eCPM_old)). Example (illustrative): Game S’s blended eCPM fell $10.00 -> $8.00 in one week while UA lifted the US share 60% -> 80%; split by country x format, about 40% of the drop was mix and 60% was real price decline inside countries. Neither part is visible in the blend. Name the segment every number is for; never report a blended figure alone.
  5. Week 1 is a poor baseline. It can include a one-off surge (one hypercasual app did in early September); also compare latest vs the previous week.
  6. Show rate cannot be computed before ~14 days ago. MAX keeps requests/attempts/responses only that long and requests cannot be combined with network. Do not invent a show rate; the MAX dashboard export is the only older source.
  7. The group gap is not causal. Floor apps start from a lower eCPM and are not randomly chosen.

analysis/floor_uplift.py and analysis/rung_uplift_monitor.py are project-local to ApplovinPersonalizedBiddingFloor; run them only in a checkout that has them.

ToolWhat it gives
analysis/floor_uplift.pyDifference-in-differences vs hypercasual controls, with a noise band; verdict inconclusive unless outside it. Needs the1platform on 127.0.0.1. Output analysis/floor_uplift.json, dashboard /uplift.
Per-country on/off via the kill-switchThe only clean measurement. Costs revenue while it runs: needs the operator’s explicit before->after confirmation, never start it unasked.
  • Valid report columns here: day, package_name, platform, ad_format, max_ad_unit, impressions, estimated_revenue, requests (ad_unit, ad_unit_name, ad_unit_id are rejected with HTTP 400).
  • estimated_revenue for the newest days can still shift; do not read one day.
  • Units are classified by trailing _N (0 = auto). Apps with other naming (e.g. Game D, Game F) would misclassify; check unit names before trusting a per-rung split for them.
  • analysis/rung_uplift_monitor.py’s docstring still describes “parallel load, highest wins”; its dead-rung and utilization numbers are fine, its uplift reading is not (see 1-2).
  • A “floor” app can still have dead upper rungs (e.g. Game A: rung 1 only). Say so; it is a wasted-slot finding, not a pass.

The script is mechanical: run it inline in the main session or hand it to a read-only analyst agent if the project defines one. The interpretation (rules above) stays with the main session; do not delegate it.