Skip to main content
FloKit surfaces recommendations when it detects a gap between your current acquisition behavior and optimal payback. Recommendations are generated daily and prioritized by projected impact on CAC payback — not by platform-reported metrics like CTR or impression share.

How recommendations are generated

FloKit continuously monitors five signal categories:
  • Budget allocation vs payback-adjusted ROAS — identifies campaigns where spend is misallocated relative to actual subscription revenue returned.
  • Creative fatigue — tracks trial conversion rate (not just CTR) on running creatives. A creative can maintain strong CTR while trial conversion declines — FloKit flags the latter.
  • Cohort LTV divergence — detects when the same channel produces materially different LTV across cohort weeks, which indicates an audience or offer shift worth investigating.
  • Offer and paywall conversion rates vs subscription quality — surfaces cases where a high-converting offer is producing low-renewal subscribers, or a lower-converting offer is producing higher-LTV cohorts.
  • Country-level payback differences — identifies markets within shared campaigns that have significantly different payback windows, which can distort blended campaign metrics.

Types of recommendations

Budget recommendations

“Shift $X from Campaign A to Campaign B.” Triggered when payback-adjusted ROAS in the target campaign is more than 15% higher than the source campaign, spend capacity exists in the target (not budget-limited), and the confidence threshold is met. FloKit calculates payback-adjusted ROAS using actual subscription revenue — not platform-reported conversion value.

Creative recommendations

“Pause creative X — trial conversion declining 35% vs 30-day average.” FloKit flags creatives where trial conversion rate is declining, independent of CTR. A creative that still generates clicks but is producing fewer trials is a spend efficiency problem that platform reporting won’t surface. Creative recommendations always include the trailing 30-day trial conversion baseline for comparison.

Audience recommendations

“Exclude lookalike segment Y — 60-day refund rate 3× account average.” Audience recommendations are based on cohort quality — renewal rate, refund rate, and 90-day LTV — not just acquisition volume. A lookalike segment that generates installs cheaply but churns fast is costing you more per retained subscriber than the CPI suggests.

Offer recommendations

“Test 7-day trial on annual plan in UK — annual plan has 2× 90-day LTV vs monthly but lower trial conversion.” FloKit suggests A/B tests for offers, not unilateral changes. Offer recommendations surface cases where LTV signals suggest an untested offer configuration could improve payback, and propose a scoped test to validate. FloKit does not modify offers or pricing directly.

Paywall recommendations

“High-performing offer X is not shown to cohorts with the strongest payback history — consider broadening exposure.” Paywall recommendations are generated from a join of paywall_viewed, offer_viewed, and subscription_started events. They surface mismatches between your best-converting paywall configurations and your highest-LTV user segments. Available during private beta.

Confidence scoring

Each recommendation includes a confidence score from 0 to 100. The score is calculated from:
  • Sample size — number of users in the affected cohorts.
  • Statistical power — signal strength relative to variance in the metric.
  • Cohort stability — whether the trend is consistent across multiple cohort weeks or driven by a single outlier period.
FloKit does not surface recommendations below 70% confidence. The confidence threshold for actions entering the approval queue is configurable in Guardrails — the default is 80%, recommended minimum is 85%.

Explaining a recommendation

Every action in the queue includes a structured explanation:
  • The signal that triggered it — the specific metric deviation FloKit detected.
  • The expected impact — projected change in CAC, ROAS, or trial conversion, with the confidence score.
  • The guardrails checked — which guardrails were evaluated and whether any were close to triggering.
  • The rollback condition — the metric threshold that would trigger a rollback if the action goes live.
If an explanation feels incomplete or the signal doesn’t match what you’re seeing in your own reporting, use the Feedback button on the action card. FloKit’s team reviews flagged recommendations and they feed directly into model improvement.