Research hub

AirPulse AI-Search Measurement Research

This is the canonical index for AirPulse measurement research. Use it to find the current stability study, the permanent calculation methodology, the prompt-selection methodology, public aggregate downloads and the update record. Individual pages link back here so readers do not have to reconstruct the evidence across unrelated posts.

Maintained by the AirPulse research and data team

Why AI-search measurement needs repeated observations

An AI answer is generated fresh each time, not retrieved from a fixed ranked list. The same prompt on the same engine can name a brand today and omit it tomorrow. A single screenshot is therefore evidence of one answer, not an estimate of visibility. AirPulse research measures each prompt-engine pair as a repeated series over time and reports the resulting rates with their uncertainty.

Current research finding

AirPulse's frozen 30-day study covered 30,504 production responses. Mention status changed between consecutive observations 9.7% of the time. 35.9% of sufficiently repeated prompt-engine cells changed state at least once. Consecutive cited-domain sets had 0.396 mean Jaccard overlap.

A separate rolling validation extracted on July 20, 2026 covered 30,146 responses and found a 10.0% consecutive mention flip rate, 33.9% mixed-state cells and 0.390 mean source overlap. The validation is reported separately so the original study does not become a moving target: the frozen cohort keeps its published numbers, and the rolling window checks whether the pattern still holds.

Diagram showing one prompt and one engine measured repeatedly across time.
The measurement unit behind every AirPulse research number: one prompt × one engine, observed repeatedly over time.

Research library

Three pages, three jobs. Start with the study if you are questioning a one-run result, the methodology if you are defining how your team measures, and the prompt-selection page if you are deciding what to measure.

Public downloads

These files contain aggregate statistics and templates only. They exclude customer names, tenant identifiers, raw prompts and raw answers.

Core definitions

TermDefinition
Prompt-engine cellOne exact prompt paired with one fixed answer engine
Valid runOne timestamped execution that returned a usable answer under the declared failure rules
Mention rateValid runs naming the brand divided by valid runs in the same scope
Own-domain citation rateValid runs citing the brand's registrable domain divided by valid runs
Source coverageDistinct valid cited URLs or domains observed across repeated runs
Jaccard similarityIntersection divided by union for two comparable source sets

The full definitions, calculations, uncertainty rules and known limitations are maintained in the measurement methodology. Every study on this hub is an observational production cohort, not a random sample of all brands or prompts, and none of it estimates the causal effect of a single content change. The study page lists what its findings do and do not support.

How the research connects to AirPulse

  • Prompt Visibility applies repeated measurement to prompt-engine cells.
  • Citation Visibility separates own-domain, third-party and source-history outcomes.
  • Prompt Analysis helps teams group and maintain the prompt roster.
  • AI Visibility provides the broader product context.
  • AI Traffic measures downstream referral sessions separately from answer visibility.
  • The free GEO Audit is the diagnostic entry point; the methodology explains why crawler access alone is not measured visibility.

The research pages explain the measurement rules. Product pages explain the workflow. Neither should substitute for the other.

Update record

DateChange
July 13, 2026Frozen production study and corrected evidence strategy established
July 20, 2026Rolling validation added; first research cluster prepared for publication

Future updates will list material changes to the cohort, parser, prompt roster, source normalisation or conclusions. Silent rewrites should not alter a published headline number. New studies will be added to the research library above as they are frozen and reviewed.

How this page was made

AirPulse analysed aggregate records from its production read replica in a read-only transaction. The public study excludes customer names, tenant identifiers, raw prompts and raw answers. The analysis code groups observations by brand, prompt and engine, orders them by time, and compares consecutive valid runs. Harsh Songra drafted the page; the AirPulse research and data team reviewed the claims and the downloadable aggregates before publication. AI assisted with structure and editing, but the published numbers must reproduce from the frozen evidence file.

Primary references

  1. Don't Measure Once: Measuring Visibility in AI Search (arXiv)
  2. Quantifying Uncertainty in AI Visibility (arXiv)
  3. How many AI-search runs are enough? AirPulse's frozen study
  4. Which AI search prompts should your brand track?

See the measurement running on your brand.

The free GEO audit runs your buyers' real questions against four engines. The research above explains how to read what comes back.