The AI Visibility Index in one page

What this Index does not measure

• It does not measure client satisfaction, campaign results, media outcomes, or revenue. Nothing here tells you whether an agency does good work.
• It does not measure what a buyer sees. It measures cold, fresh-context recommendation behavior on programmatic endpoints, not a signed-in consumer app with memory, personalization, and an enterprise purchase history behind it.
• It does not measure quality per dollar. Agency size drives raw visibility upward and drives CEI downward; a small firm will tend to lead CEI for arithmetic reasons, which is why the raw table leads and a size-bracket view publishes beside it.
• It does not measure causation. An agency's movement between editions can come from its earned media, from a vendor changing models underneath the instrument, or from chance across 25 questions. The Index reports which of those it can and cannot separate.

1. The intervals cover the questions, not everything else. They resample the 25 benchmark questions. They do not cover prompt-selection bias, the gap between this instrument and a consumer app, or errors in a reported senior headcount — CEI intervals treat the denominator as correct.
2. Four scoring rules were not in the v1.10 text. Rank normalization, the treatment of an agency that does not appear, the line between a recommendation and an incidental mention, and alias handling were settled in writing on 2026-09-22 as protocol amendment v1.11, before the readiness gate and before collection opened, and publish with a worked example that recomputes one cell by hand.
3. It is an API instrument, not a consumer one. Answers come from vendor model endpoints, not from a person's browser. Consumer apps carry memory, location, and personalization that this instrument deliberately does not reproduce, so these are not the answers any one buyer would see.
4. The prompt set is fixed and designed, not sampled. The questions were written to cover generic, vertical, problem-stated, and competitive intent. They are not a random draw from all buyer questions, so generalization rests on how well that design represents real purchase behavior.
5. Engines are weighted equally. Each engine counts the same in the composite, which is a choice rather than a fact about usage. Every edition publishes a usage-weighted sensitivity line so a reader can see how much the ranking depends on that choice.
6. Dozens of agencies means some movement is chance — and the test can also miss real change. Every agency is tested at 95% confidence, so roughly one or two apparent movements per edition are expected by luck alone. Crackle PR publishes the unadjusted deltas as preregistered, plus a clearly labelled exploratory multiplicity-adjusted view that no rank rests on. The paired test runs on 25 question clusters and is conservative: a flat delta means the design could not see a change, not that nothing changed, so each edition publishes a minimum-detectable-effect estimate. Q4 also runs newer model IDs than Q3 on the delta engines, so most movement mixes agency change with model change and is labelled that way.
7. The CEI denominator is self-reported in substance. Senior headcount is read from public LinkedIn profiles and agency team pages, which agencies control and can title-inflate or title-deflate, and "Director" is not equivalent seniority across agencies. It counts every Director-and-above practitioner at the whole agency, not only its technology practice. Each agency's count, source URL, and capture date publish in the dataset and are sent to that agency for correction before release.
8. Crackle PR's own press footprint can move Crackle PR's own score. Each edition creates more text linking Crackle PR to B2B tech PR and AI visibility, which can feed future model training. Crackle PR's raw composite is watched quarter over quarter, and a rise faster than client work explains is logged as a suspected artifact of this Index rather than reported as performance.