The Citation Index panel: a standing record of who gets cited

What this check measures

Separate from any audit you run, we operate our own standing measurement panel: a fixed set of representative questions per vertical that we put to the AI engines on a regular, slow schedule. Each time the panel runs, it records who the engines actually cited for those questions, carrying the same receipts every other measurement carries, the verbatim answer, the engine, and the timestamp. Over time this accumulates into a longitudinal record of how citation behavior in a vertical moves: which brands and sources are gaining presence, which are fading, and how the engines differ from one another. The panel is our own research instrument. It is owned and run by us, not driven by any single customer, and it is kept strictly separate from the questions individual customers choose to monitor, so the two are never conflated.

Why it matters

A single audit is a snapshot; the interesting questions are about movement and context. Is your AI visibility actually improving, or is the whole vertical shifting under everyone at once? Who are the brands the engines keep citing in your category, and are you closing the gap or losing ground? Those questions can only be answered against a consistent, repeated measurement that predates your interest in any single result, which is exactly what a standing panel provides and what a one-off cannot. Building this record is slow and deliberate by design; it compounds with every cycle and cannot be manufactured after the fact. That is why we run it continuously in the background rather than waiting until the data is needed.

How we score it

The panel reuses the same measurement engine and the same receipts that every other check uses; nothing about how an individual citation is recorded is special here. What is specific to the panel, the questions it asks, why those questions were chosen, the cadence it runs on, and how results are organized into the longitudinal record, is our research design and is deliberately not published. That design is the work that makes the eventual Index valuable, and disclosing it would let it be copied without being earned. When the Index is published in a later release, each published figure will carry its own receipts: the date of the reading, the sample behind it, and openable evidence, so the output is verifiable even though the construction behind it is not disclosed.

Confidence-flag rules

The panel runs on a slow, deliberately bounded schedule rather than continuously, both because longitudinal trend is the goal and because cost is strictly controlled. Every single measurement the panel makes passes through the exact same budget breaker that governs every other measurement in the product: before any engine is called, the spend is checked against a hard per-provider ceiling, and once that ceiling is reached the call is paused and returns at no cost rather than overrunning. The panel cannot bypass that ceiling, so a misconfiguration can at worst pause measurement, never run up unbounded spend. To keep cost predictable, the panel leans on the cheaper engines by default, with the more expensive engines included only deliberately. Each cycle is recorded with its own date and is never pooled across time, so a trend is always read as a sequence of dated readings rather than a single blended figure that would hide when something actually changed.

Common mistakes

• Reading a single audit as a trend. One snapshot cannot tell you whether you are improving or whether the whole vertical is moving; only a repeated, consistent measurement can.
• Assuming a longitudinal dataset can be assembled retroactively. A standing record only exists if you were measuring consistently before you needed the answer.
• Confusing the panel with your own monitored questions. The panel is our own research instrument and is kept strictly separate from the questions any customer chooses to track.
• Expecting unbounded measurement. The panel is intentionally slow and cost-bounded; every cycle runs through the same hard budget ceiling as the rest of the product.

How to fix it

There is nothing for you to fix here; the panel is our own measurement instrument, not a check on your site. What it gives you, once the Index is published, is context: the ability to read your own results against how the vertical as a whole is behaving, so you can tell genuine progress from a category-wide shift. Until then, the actionable checks remain your own audit and your own monitored questions; the panel runs quietly in the background so that the longitudinal record exists when it is time to publish. When the Index does ship, every published reading will be dated and backed by openable receipts, so you will be able to verify the output even though the design of the panel behind it stays our own.

Primary sources

• OpenAI, reproducibility and the seed / temperature parameters (why repeated measurement is needed)

  OpenAI

• Anthropic, Claude API temperature and sampling

  Anthropic

• Perplexity, Sonar API and online models (cross-engine citation behavior)

  Perplexity

• Google Search Central, AI features and your website (AI surfaces evolve over time)

  Google Search Central

Changelog

• 2026-06-23 · Initial publication. Documents the standing measurement panel: a service-owned, slow-cadence cross-engine poll that accumulates a longitudinal record of citation behavior per vertical. The published Index is a later release; cost is hard-bounded by the same budget breaker every other measurement runs through.