What GEO actually influences
Assistants may draw from model training memory, live web retrieval, connected files, store pages, publisher pages, third-party reviews, and current conversation. The mixture varies by engine, mode, query, account, locale and date. GEO improves the eligibility and quality of what assistants can find, it does not dictate the answer. Any vendor promising 'guaranteed AI ranking' is selling superstition.
The assistant-retrieval flow
User prompt → intent interpretation → retrieval decision (yes/no) → source selection → generation → observable outcome (mention, recommendation, position, citation, store link). Every stage is conditional. The operator can improve evidence and measurement; the operator cannot force retrieval. Log each stage separately and mark hidden stages as unknown, that discipline alone separates honest GEO from marketing theatre.
Why a store page is often insufficient
Store pages are authoritative for publisher facts, availability and assets. They rarely answer comparative or constraint-heavy questions, 'works offline', 'best for two people', 'privacy-first', 'supports Hindi', 'good for irregular income', with the depth assistants need to ground an answer. Assistants seek corroboration: current pricing, reviews, help content, capability pages. Your job is to make that corroboration accurate and reachable, not to fake it into existence.
The evidence map
For each high-value non-branded intent, build a capability-evidence row: the intent, the natural-language question that expresses it, the owned proof (help article, capability page, schema, in-app evidence), the third-party proof (legitimate reviews, comparisons, category coverage), and the gap. Ten rows to start. The gap column is your work backlog for the quarter.
The prompt panel
Select a stable panel of prompts covering intent clusters: capability ('best app for X'), competitive ('X vs Y'), price ('free app for X'), compliance ('privacy-safe X'), geography ('X in India'), persona ('X for couples'). Fixed panel is essential, month-over-month comparability is the entire methodology. Never brand-lead the prompt; that measures ego, not discovery.
Coding what you observe
For every prompt × engine × mode: was your app mentioned (yes/no), recommended (yes/no), positioned (rank in list if any), cited (source link present), linked to store, linked to web, sentiment (positive/neutral/negative), accuracy (fact-check the assistant's claim about your app). Eight independent fields. Never collapse them into a single 'AI score', the loss of resolution is the whole information.
Entity consistency across surfaces
Assistants perform entity recognition: is 'Example Co' on this comparison site the same app as 'Example Co' on the App Store? Publisher name, app name, developer URL, package/bundle IDs, structured data (MobileApplication schema), Wikipedia entry if warranted, and Crunchbase-style directory entries must all agree. One inconsistency and the assistant hedges.
Third-party authority (legitimate only)
Map which sources each assistant favours in your category, Wirecutter, The Verge, category-specialist publications, aggregator sites, community forums. Earn coverage in those sources through legitimate editorial outreach with real stories. Never buy placements, never manufacture Reddit posts, never pay for fake reviews. Assistants will eventually detect the pattern and the damage compounds against you.
What is prohibited
Fake comparison sites. Paid fake reviews. Manufactured Reddit or Quora threads. Fabricated citations. Cloaked content that serves different HTML to crawlers than users. Undisclosed paid placements presented as editorial. Doorway pages repeating 'best X app' with no genuine content. Every one of these is short-term visible and long-term catastrophic, and the honest client base you want will never touch a vendor caught doing them.
How we defend the number
Share of Recommendation is defensible because the panel is fixed, sampling is public, engines and modes are named, and raw outputs are preserved. Boards accept the metric when methodology is transparent. Every monthly report includes the raw panel outputs on request, no black box. If a competitor's number is higher than ours and their methodology can't be shown, that is the finding.