Methodology · the loop, named
You don't run ASO by hand — ASOLOOP runs the loop; your job is to check it. Below: the five-step loop it runs across both stores, then the governance spine that makes letting it act safe. Every step has a receipt; every action has an undo path.
How it runs
Propose → generate → launch → apply the winner → learn — across both stores, in your storefront’s language: App Store winners and Google Play SLE winners apply live; CSL winners stage for your publish. You grant how much it does on its own; you can check, edit, or revert any move.
ASOLOOP's AI reads your per-app signal graph — store data, your keyword tools, the MMP, every past experiment — and proposes what to test next across both the App Store and Google Play: ranked hypotheses, each with a signal-trace rationale. This is the step other tools leave to you.
It drafts the creative — copy and visual briefs — in your storefront's listing language, for both the App Store and Google Play listings, and runs every output through the claim-safety validator before it reaches you.
It launches the test as a native store experiment — PPO or CPP on the App Store, a store listing experiment (SLE, the Android default) on Google Play — or stages a custom store listing for your send-for-review. Both stores, scoped to the storefront locale it detected — PPO, CPP, and SLE end to end; CSL send-for-review staged for you.
When a variant wins, ASOLOOP applies it to your live listing — end to end on the App Store and on Google Play store listing experiments (SLE) — and stages the winning CSL treatment ready for your send-for-review. Every other tool stops at a recommendation and hands the doing back to you; ASOLOOP closes the loop. Reversible within 14 days.
Every result, from both the App Store and Google Play, writes to the per-app evidence library, so the next cycle starts smarter than the last. The loop compounds.
Reversible · gated · explained · logged
You never have to believe, only check. Every action the loop takes on its own carries an undo path, a gate it passed, the evidence behind it, and a log entry — the Linear-Method credibility beat that makes letting it act safe.
Reversible
Revoke a signal within 7 days, suspend a running experiment, roll back an applied winner to the previous live listing within 14 days, or turn Autopilot off effective next cycle. Project Rules edit history is versioned too — rollback is one click, everywhere.
14-day winner revert windowGated
Un-sourced and point-estimate revenue claims are rejected outright; regulated-category claims need an evidence link or a Rights Administrator attestation. The gate runs per storefront language, with a deterministic-template fallback when a draft is rejected.
5 gate rules, every generationExplained
Every LLM-rendered surface keeps its deterministic evidence beside it — source-traced to the experiment that produced it, queryable per app, and revocable within 7 days. You check the claim; you don't just trust it.
18 reasoning surfacesLogged
Timestamp, actor, and the signals behind it — recorded for every revocation, suspension, applied winner, and Autopilot change. Nothing is silently dropped.
Exportable audit logHow confidence shapes commitment
Many candidates when evidence is thin; fewer when it’s reliable; one when it’s rich enough to act on. Each app sits in one mode at a time.
Cold start
Posterior shape is unstable; ASOLOOP surfaces multiple candidate directions on every cycle so the operator picks based on judgment, not a borrowed score. Inconclusive PPO/CPP results still write contribution; precision-tier disclosure carries forward.
Signals reliable
Auto-progresses from Learning once the app clears all three evidence gates — three active signals, signal health at or above 1.5 sigma, and three qualified hypotheses. Experiment outcomes feed those gates as signal rows; there is no experiment count to hit. Each candidate carries a signal-trace rationale; the operator can inspect every contributing experiment.
Signal-rich; commitment
Surfaces a single recommendation when accumulated evidence is dense enough to back one candidate over the alternatives. Required precondition for Autopilot. Never auto-promoted from Classic — Agent is always an explicit operator opt-in, with the implication understood.
Auto-progression rules: Learning → Classic fires when all three evidence gates clear (the thresholds are product-set, not per-app dials — what you control is a per-app mode override); Classic → Agentnever auto-fires — Agent is always an explicit operator opt-in. Reversion in either direction is a single click, takes effect on the next cycle.
The gate before anything renders
A mandatory gate that runs before every generative output reaches you. Nothing renders without passing.
The proof stays visible
Every LLM-rendered surface in ASOLOOPkeeps its deterministic evidence in view — the model writes the sentence, but the data it reasoned from stays visible, source-traced, and revocable. You check the claim; you don’t just trust it.
evidence-in-viewBanners, rationales, hypothesis-card recommendations
The model renders the sentence; the signals, contributions, and numbers it reasoned from stay visible beside it — never replaced by the prose. The rationale is editable; the recommendation can be revoked.
source-tracedEvery signal, every claim
Each claim traces to the source experiment that produced it. The signals are queryable per app; click into any surface and verify the evidence the model claims to summarize is the evidence in the system.
revocableEvery signal you disagree with
Suppress or revoke any signal within a 7-day window and it disappears from future rankings. The audit trail keeps the record; nothing is silently dropped.
From posterior to revenue range
CVR posteriors don’t survive a CMO meeting. Revenue ranges do. Today ASOLOOPrenders one workspace-level revenue band, computed from your connected AppsFlyer revenue, on the Teams-tier stakeholder dashboard. Translating each hypothesis’s Bayesian CVR posterior into an expected revenue range from the inputs below, with assumptions disclosed on every projection, is in progress — this is the model it will run.
Bayesian CVR posterior
Native PPO / CPP / SLE / CSL result
Posterior probability distribution over the conversion-rate delta between treatment and control variants. Inconclusive results render as overlapping distributions; ASOLOOP does not collapse to a point estimate.
ARPU
MMP-derived (AppsFlyer-derived today; Adjust / Branch / Firebase Analytics connectors live)
Average revenue per user. Operator-MMP-sourced exclusively — no operator-attested fallback per the claim-safety contract.
Install-to-paid
MMP-derived
Ratio of installs that convert to paying customers within the operator-defined window.
Lifespan
MMP-derived
Expected duration a paying user stays subscribed or retained, used to project lifetime-value range from a single CVR posterior.
Output will render as ~$X–$Y/month additional revenue at N% probability with the four inputs disclosed inline once the per-hypothesis translation ships. Stakeholder-readable; defensible upward.
The compounding substrate
The substrate that compounds. Every signal ASOLOOPwrites is scoped per app, per audience. No global ASO model. Signals from one Operator’s app do not feed another Operator’s recommendations. Cross-app pattern aggregation is a planned Teams capability under a separate consent surface.
The library is the difference between “we ran four tests” and “we’ve learned four things from those four tests.”
Where to next
The 7-day trial drops you into a workspace where the architecture above is live. Connect an app, log an experiment, watch the evidence library accumulate.