Deterministic about logic, probabilistic about results

The mistake every other tracker is making

The current generation of AI-visibility tools all measure the same thing in three different costumes. Some poll ChatGPT, Claude, Gemini, and Perplexity through their public APIs every day. Some replay a giant historical corpus of prompts against cached answers. Some pay panels of real users to share their AI-surface traffic and sell the rollup back as "prompt demand." The wrappers look different. The substrate is identical: they record what the model produced on a given Tuesday, average it across runs, and put it on a line chart.

That approach has one fatal property. The thing it measures — model output — is the most volatile, least durable, most easily perturbed signal in the entire stack. Sampling temperature, system-prompt edits, personalization, conversation history, A/B rollouts, silent model upgrades, even the time of day a vendor re-tunes a retrieval index — any one of them moves the number. None of them are decisions you, the brand, can act on. By the time the line chart wiggles, the cause is already three releases old.

Worse, the output-tracking posture forces every one of those tools into a small lie: they present a single number as if it were a measurement. It isn't. It's one draw from a distribution. Run the same prompt an hour later and you'd get a different draw. The screenshot in their dashboard is true and meaningless at the same time.

The asymmetry we built on

There is an asymmetry hiding in plain sight in how modern AI agents work, and it is the foundation of our methodology.

Agent output is stochastic. Large language models generate text by assigning probabilities to candidate next tokens and sampling among them. The same prompt, asked twice, can produce different answers. That is a property of the architecture, not a bug. Nobody — not us, not the vendor, not the model itself — can promise you a specific sentence tomorrow.

Agent logic is not. The codebase that wraps the model, the retrieval index it consults, the manifest formats it accepts, the tool schemas it can call, the crawler user-agents it respects, the ranking heuristics it uses to pick a citation, the trust signals it weighs when it decides whether to recommend you — these change on a release cadence, not a per-prompt one. They are documented, observable, and reverse-engineerable. They are the levers that, when pulled, move the distribution of outputs in a direction you can predict.

Once you internalize that asymmetry, the right product to build becomes obvious. You don't chase the volatile output. You model the stable logic, and you forecast the output from there. That is what AIAO is.

What we actually model

Every agent in our scored panel — ChatGPT, Claude, Gemini, Perplexity, Copilot, Grok, Meta AI, plus Apple Intelligence as a personal-context surface — is treated as a stack of decisions, not a black box. For each agent we maintain a living catalog of:

• The crawlers it operates and the robots directives it actually obeys.
• The retrieval indexes it queries (its own, a partner's, or a live web fetch) and the freshness windows on each.
• The manifest formats it parses — llms.txt, /.well-known/mcp.json, App Intents, App Actions, ChatGPT Apps SDK, Copilot connectors, Perplexity actions, the X App Card, AASA, assetlinks.
• The tool-selection rubric it applies when more than one tool could answer.
• The trust signals it visibly weights — Wikipedia, Wikidata, review-site density, knowledge-panel eligibility, structured-data validity, citation provenance.
• The handoff surfaces it can hand a user off into — web click, deep link, App Intent, MCP tool call, in-app messenger.
• The refusal and safety filters that can quietly remove you from an answer for reasons unrelated to merit.

None of that is guessed. It comes from agent code where it is public, vendor documentation, controlled simulations at constant temperature, parallel sample answers across logged-in and incognito variants, and the shape of the tool calls the agents actually make. When the logic changes — a new crawler ships, a manifest format gets new required fields, a vendor flips on a retrieval partner — our catalog moves and your scores move with it. Not because we re-sampled today's chat output. Because the rules of the game changed.

Why this produces better forecasts than output tracking

Imagine two ways to predict whether a chess engine will beat a human next Tuesday. The first: have it play thousands of games this week and extrapolate the win rate. The second: read its opening book, evaluate its search depth, inspect its evaluation function, and reason from there. The second method is cheaper, more stable, and survives the engine being patched on Monday night. The first method has to start over.

AI agents are the chess engine. Our methodology is the second method. We do still run prompts — see the five-layer triangulation stack below — but we use them to validate that our model of the logic still matches the agent's behavior, not as the primary source of truth. The model is the source; the prompts are the audit.

The five-layer triangulation stack

A single signal source can be wrong. Five independent ones converging cannot all be wrong in the same direction. Every finding that lands in your report has to clear a quorum of:

1. Real devices. Headed sessions on physical iOS, Android, macOS, Windows, and major browsers — what an actual user actually sees, including consent walls, app-clip prompts, and store rendering quirks.
2. Controlled simulations. Prompt panels held at constant temperature, re-run on a rolling schedule across logged-in, incognito, regional, and tier variants. Designed to isolate which variable moved the answer when the answer moves.
3. API investigations. Direct calls to provider, store, registry, and structured-data APIs to ground-truth what the agent's retrieval layer can see about you, independent of what it chooses to say.
4. Agent interrogations. Multi-turn probes of frontier agents in their default consumer configuration — brand, category, comparison, and intent prompts — to measure selection, rank, attribution, and refusal behavior in the wild.
5. Hallucination and anomaly detection. Cross-layer reconciliation that flags fabricated facts, mis-attributed quotes, stale citations, and out-of-distribution swings before they land in a report and mislead a fix.

The confidence score next to every finding on the platform is literally derived from how many of these five layers agreed. Two-of-five is a hypothesis. Four-of-five is a fix. Five-of-five with stable trend is a moat.

What this means for the numbers on your dashboard

Scores, ranks, mindshare

Treat them as forecasts conditioned on the logic the agents are running today, not as measurements of a stable truth. The honest unit is the trend, the gap between agents, and the relative position against rivals — not a single day's exact figure. When we publish a central estimate we use the geometric mean of the modelled band so the headline number is not biased by a single anomalous draw, and we show the full range so you can see how wide the uncertainty is.

Revenue, traffic, query volume

Same posture. The central estimate is a forecast; the band is the uncertainty; the trend is the trustworthy unit. We will never quote you a precise daily count and pretend we measured it. We will quote you a forecast and tell you what would have to move in the agent logic to make that forecast wrong.

Recommendations and Push-to-Live

Because the logic is the durable layer, recommendations target the logic. "Add an Applebot-Extended allow." "Donate App Shortcuts for these three intents." "Register your MCP server in the Copilot Studio catalog." Each recipe has a modelled expected lift, an effort estimate, a confidence drawn from the five layers, and a verification window — and every recipe pushed live enqueues a Lift Harness probe set against a control group so the causal claim is later proven or disproven on the actual agents.

What we deliberately do not do

• We do not present a single sampled output as a measurement.
• We do not average across the agent panel by default — each model weights signals differently and averaging hides the per-agent gap you most need to see.
• We do not scrape or republish model weights, system prompts, or training data.
• We do not synthesize prompts designed to manipulate an agent against its own guidelines.
• We do not use any competitor's product name or proprietary metric label as our own.
• We do not promise "100% accurate" or "real" numbers. The honest answer is a calibrated forecast, and that is what we ship.

The one sentence to remember

We model how agents decide — not what they'll say tomorrow. Output is stochastic. Logic is not. Once you measure the logic, the results become predictable enough to act on, and the recommendations you ship can be proven causally rather than guessed at. That is the methodology, and to our knowledge no one else in the AEO/GEO/AIAO category is built on it.