Reading the numbers

Every number on the platform is a forecast, not a measurement. AI answers shift day to day, so a single-day exact number would be misleading. Instead we show a middle number with a range around it. Read the trend over weeks and the gap versus competitors. Ignore tiny day-to-day wiggles.

Every score, rank, volume estimate and revenue figure on Appkitekt AI is a forecast grounded in agent logic — not a measurement of a stable truth. AI agents are stochastic at the output layer: the same prompt can yield different answers across users, sessions, and model updates. But the logic the agents are running — the documented APIs, the codebases, the manifest formats, the retrieval rules, the signals each agent reads when picking a brand — is stable enough to model directly.

So we publish a central estimate (the geometric mean of the modelled band, the unbiased centre on a log-uniform distribution) with the full range visible next to it. Read the trend, the gap, and the relative rank across agents. Treat the single-day exact figure as a forecast, not a measurement. See the full methodology for the deterministic-vs-probabilistic split.

A single composite score per app, refreshed each audit cycle. It is the weighted mean of the 12 pillar scores (see Scoring & weighting). We catalog and score the signals AI agents actually use to discover, select, and trust brands — signal selection is observed from agent behavior, so the Index is best read as a relative, indicative read rather than an absolute grade. Two apps with the same Index can have very different underlying problems.

Bands: 0–39 Invisible, 40–59 Mentioned, 60–74 Considered, 75–89 Recommended, 90–100 Default-of-record.

Each pillar is scored independently on its own rubric. A 70 in AEO means something different from a 70 in AAIO — the rubrics are not interchangeable. Always read the pillar score alongside the per-agent breakdown.

Every pillar shows six bars, one per model in the agent panel. A low average masking one strong model is common — and it usually points to a structural signal (schema, MCP, freshness) that one model rewards and the others ignore. Optimizer recipes are agent-specific for exactly this reason.

The share of category prompts in which your app appears in the agent's answer, weighted by rank position. A first-place mention counts more than a buried one. Mindshare is the agent-era equivalent of share-of-voice, not market share.

Note: we use the term Mindshare for this metric. Other tools may market a similar idea under their own product names — those are not interchangeable, and our scores will not match theirs because the prompt sets, models, and rank weighting differ.

A delta is the change since the last completed audit, not since you last opened the dashboard. Trend lines plot the last 12 audit cycles. A single-cycle swing of ±3 points is within normal sampling noise; treat ±7 or sustained movement across two cycles as signal.

When you see "85th percentile," the cohort is other audited apps in the same category band over the last 30 days. It is a positioning hint, not a ranking — the cohort can shift week to week as new apps audit in.

Small dots next to a score: solid = high confidence (full prompt set, all agents responded), half = partial sample (a model rate-limited or returned an empty answer), hollow = insufficient sample, treat as directional only.

All the numbers above apply to both web and app audits. The metrics below replace their web-only counterparts when the target is an App Store or Play listing.

• Listing DST score (0–100) — the App-mode counterpart of the Appkitekt AI Index, scored on 200+ factors. Same bands, same formula.
• Agent recommendation share — % of category prompts where the agent names your app. App-mode equivalent of Mindshare.
• Install-intent share — % of install-intent prompts ("which app should I install for X") where you appear.
• Handoff CVR — observed agent → store → install resolution rate, derived from instrumented Universal Links / asset links.
• Rival app overrides — count of queries where a rival app was recommended instead of yours.
• Listing freshness age — days since last what's-new / version bump.

Web-only metrics like Citation share and Retrieval inclusion disappear in App mode, and vice-versa. The Index, pillar scores, deltas, percentiles and confidence dots stay common to both surfaces.

When the audit target is an MCP server or public API rather than a site or app, the headline metrics swap again. The Index, pillar scores, deltas, percentiles and confidence dots remain identical.

• Tool DST score (0–100) — the MCP-mode counterpart of the Appkitekt AI Index, scored on 210+ factors. Same bands, same formula.
• Tool-call share — % of agent turns in your category where the agent invoked one of your tools instead of a rival's. MCP-mode equivalent of Mindshare.
• Registry coverage — % of public MCP registries and tool catalogs that list your server with a complete manifest.
• Tool-selection precision — share of agent attempts that pick the right tool from your tools/list on the first try; low values flag description / naming collisions.
• Schema fidelity — share of tool calls that succeed without parameter coercion or retry; tracks JSON Schema completeness and required-field clarity.
• Latency p50 / p95 per tool — agents downweight slow tools fast; both percentiles are tracked per tool, per agent.
• Token footprint — average tokens per response. Verbose tools get dropped in long agent loops.
• Hallucinated-call rate — share of agent answers that describe calling your tools without actually calling them, often citing wrong parameters or fictional methods.

Scores answer "how ready are you"; the dashboards also turn that into agentic visits, human impressions and a monthly USD band. The math, the defaults (conversion rate, AOV, baseline visits) and the coarse → partial → refined confidence steps are all in Traffic & revenue.