AI field guide · Verified 2026-10-12
ASO with AI, without invented certainty
Use a model for synthesis, critique, and drafts while keeping facts, product truth, and final judgment human-verifiable.
In brief
AI can organize evidence, propose alternatives, validate mechanics, and expose inconsistencies. It cannot know private store algorithms, ground-truth keyword volume, your unshared product behavior, or whether generated copy is true.
Give the model evidence, not vibes
Provide the audience, job, shipped features, exclusions, business model, current metadata, screenshots, reviews, support themes, and primary platform documentation. Mark private material explicitly.
Use AI for bounded jobs
Good jobs include clustering language, finding duplicated terms, generating alternatives, checking consistency, drafting a screenshot brief, or turning an experiment note into a structured record.
Keep uncertainty visible
A model can suggest search phrases but cannot supply authoritative search volume or ranking causality. It can draft localization but cannot replace an informed native reviewer.
Review as a diff
Preserve the original, inspect each proposed change, trace claims to product evidence, validate limits, and reject generic wording that erases the brand voice.
Run six checks before anything ships
Verify product truth, private-data separation, current platform policy, field and asset mechanics, brand voice, and market context. A fluent draft is not evidence that a claim is true.
- Trace every capability and outcome to the shipped product.
- Keep credentials, reviewer notes, private analytics, and customer data out of public output.
- Validate characters, UTF-8 bytes, formats, locales, and the live console preview.
- Require a human to approve the diff and record unresolved unknowns.
Primary sources
Rules and interfaces move. Recheck these sources before submission-sensitive work.