Measure when ChatGPT, Gemini, Perplexity and assistant-style experiences mention, recommend, cite or link to your app—and strengthen the store and source evidence those outcomes depend on.
Built for CMOs who need a defensible answer to “Are AI systems sending people to us?” without confusing one screenshot with a market trend or one mention with qualified recommendation visibility.
App Store store + analytics
Google Play store + acquisition
Google Ads app campaigns
Meta app promotion
OpenAI observed answers
Gemini observed answersThe correct unit is prompt × engine × mode × locale × account state × date × repetition. Change one dimension and the answer may change. That variability is part of the method, not noise to hide.
| Dimension | Example | Why it matters | Recorded as |
|---|---|---|---|
| Prompt | “best offline meditation app” | Intent and wording shape the shortlist | Exact text + prompt cluster |
| Engine + mode | ChatGPT web search | Retrieval and citation behavior differ | Product + mode + observed version context |
| Locale | US · English | Store availability and source landscape differ | Market + language |
| Account state | Logged out | Personalisation can alter the answer | Declared state |
| Date + repetition | 24 Jul · 5 repeats | Volatility requires repeated sampling | Timestamp + run number |
A brand can be named but not proposed. It can be recommended without a visible citation. It can be linked with an inaccurate capability description. Each state implies a different intervention.
The store, page or source can be accessed and understood.
It enters the result or answer-generation context.
The app name appears in the answer or shortlist.
The app is explicitly proposed for the user’s job.
A source supporting the app is visibly referenced.
The path is reconciled to install, activation and value.
The panel spans category choice, jobs, alternatives, comparisons, constraints, capabilities, trust, locality and device context. Branded prompts are separated from category discovery.
Which app should I use for this job?
Which product fits the user’s constraint?
Can the product explicitly do the required thing?
Is the product credible enough for the category?
Does the answer change by market or language?
Is the app suitable for the actual use environment?
We map which facts are explicit on the store listing, product site, help centre, documentation, review ecosystem and independent sources. Store and web language must agree without manufacturing consensus.
Availability, category, metadata, screenshots, ratings, release notes and in-app events provide a native, high-control source. The app must be legible here before external evidence can compensate.
Product, help, comparison and category pages can state capabilities, constraints and trust facts with far more depth than a store field. Claims require a clear owner and support source.
Editorial, expert, publisher and review sources can strengthen authority when the relationship is genuine. Paid placement and fabricated review activity are never represented as independent evidence.
Every observed answer is coded for capability accuracy, availability, price, platform and material caveats. Incorrect prominence can be commercially worse than absence.
The method remains consistent while engine, mode and visible source treatment are recorded separately. We report what was observed—not assumptions about private training data.
Observe model mode, web-search state, visible sources, direct links and repeated answer framing.
Record answer mode, Google ecosystem surfaces, source links, locality and store-route continuity.
Inspect visible citations, source mix, comparison framing and whether app-store/product links are exposed.
Track explicit recommendations, factual descriptions and source behavior in available modes.
Separate conventional organic/store results from AI answer modules and recommendation surfaces.
Treat Siri and Apple discovery behavior as observed native experiences, not a single stable ranking index.
The denominator is the declared set of repeated observations. We keep prompt coverage, mention frequency, recommendation frequency, citation frequency, link exposure and accuracy separate before comparing brands.
| Metric | Definition | Decision it supports | Must include |
|---|---|---|---|
| Prompt coverage | Relevant prompt themes represented in the panel | Sampling completeness | Panel design + exclusions |
| Mention frequency | Observations in which the app is named | Basic visibility | Engine · mode · repetitions |
| Recommendation frequency | Observations explicitly proposing the app | Qualified choice visibility | User job + answer context |
| Citation frequency | Observations showing a supporting source | Source authority | Visible URL or source label |
| Accuracy rate | Correct coded factual statements / statements checked | Reputational risk | Claim set + reviewer |
| Attributed action | Measurable visits, store opens or installs from exposed routes | Commercial contribution | Instrumentation limitations |
The programme improves answer readiness without pretending to control proprietary systems. Every intervention is tied to a source gap and a future observation date.
Connect sources, verify baselines, map discovery doors, define the commercial event and expose evidence limitations.
Publish the highest-confidence metadata, conversion, review or source-evidence intervention.
Run store-native tests, custom routes and repeated recommendation observations with explicit decision rules.
Read store movement, value cohorts and source outcomes; scale, iterate or stop each workstream.
Finance used job language and trust evidence. Wellness made offline capability explicit. Education separated student and parent discovery language.
The product was trustworthy. The language was not how people described the money job they needed done.
A beautiful brand was invisible for practical, high-intent needs: offline meditation, sleep support and anxiety routines.
Students and parents searched the same outcome in different language.
Recommendations can be commercially important and highly variable. The operating standard should make both truths visible.
No provider controls inclusion. We improve store and source eligibility, product clarity and evidence quality, then observe outcomes under a documented sampling protocol.
There is no universal count. The panel must cover the app’s meaningful jobs, comparisons, constraints, markets and audience language. Repetition and sampling design matter more than a large vanity count.
No. The programme records observable answers and visible sources in declared user-facing modes. It does not claim access to training data or proprietary ranking factors.
Sometimes, partially. Direct links, referral paths, branded-search lift and assisted journeys can be instrumented, but dark or cross-device behavior creates unavoidable limits that must be stated.
A mention names the app. A recommendation explicitly proposes it for the user’s job. Citation and link exposure are coded separately.
Enough to detect durable change without overreacting to day-to-day volatility. Ongoing programmes typically use a stable core panel with scheduled repeat windows and event-driven checks.
No. Source quality, explicit product facts, useful category education and genuine authority matter more than mechanically producing pages around every prompt.
The AI Discovery Sprint establishes the baseline, competitor contrast, source gaps and a prioritised 90-day remediation plan.
The one-time sprint returns a declared prompt panel, coded outcomes, source map, competitor contrast and 90-day evidence plan.