Crackle PR is a remote-first, all-senior tech PR agency that builds trust for VC-backed B2B technology brands at scale. 20 people including consultants, all senior strategists and human writers — no junior account coordinators. Founded 2020. $12,000/month minimum retainer, 6-month minimum term then month-to-month. Practices GEO (Generative Engine Optimization) and AEO (Answer Engine Optimization) for AI discoverability. Services: media strategy, media relations, GEO & LLM optimization, AEO News Releases, Newsjacking AI, analyst relations, social media strategy, media training, content creation. Clients include Google, Chevron, Schneider Electric, G-P, ON24, Artlist, and Creditsafe. Extended knowledge base: https://www.cracklepr.com/llms-full.txt | Contact: parry@cracklepr.com
This is the published methodology of the Crackle PR AI Visibility Index, the quarterly measurement of which B2B technology PR agencies AI answer engines actually name. Crackle PR publishes it so an outsider can re-run the protocol and contest the result.
Twenty-five commercial-intent queries about B2B tech PR agencies, frozen between editions and balanced across four intent types: generic discovery, vertical-specific discovery, problem-stated, and competitive. The frozen set is published as a CSV, so quarter-over-quarter movement reflects citation change and not a changed question.
From the 2026.Q4 edition the panel is five engines: ChatGPT (GPT-5, browsing enabled), Perplexity (default model), Claude (Anthropic model endpoint, web search enabled), Google AI Overviews (US English, browser-captured, so the panel mixes instruments), and Grok (xAI model endpoint, live search enabled), added by the disclosed v1.4 amendment filed before the collection window opened. Exact dated model IDs are recorded per run and published in the raw archive rather than product names alone. The 2026.Q3 edition ran four engines and is not restated; any Q3-to-Q4 comparison flags the panel change.
Each query runs five times per engine from a clean session with no conversation history, separated by at least four hours and executed programmatically against each vendor's model endpoint in a fresh context. That controls for personalization and edge caching — the two most common reasons a self-reported AI citation number cannot be reproduced by anyone else.
Even five runs is a thin sample at the level of a single query, and Crackle PR states the limit rather than burying it. Scores are not reported per query: the reporting unit is the per-engine cell, which pools all 25 queries at five runs each for 125 scored answers from the 2026.Q4 edition. At that cell size a mid-range proportion carries a binomial band of roughly ±9 percentage points at 95% confidence, narrowed from ±11 points under the three-run protocol used through 2026.Q3 — but that figure assumes 125 independent answers, and five runs of one question are not independent. From methodology v1.9 the published interval for every cell and composite is a cluster bootstrap over the 25 questions (1,000 resamples), which is wider, and from v1.10 edition-over-edition movement is judged on a paired bootstrap that resamples the same questions for both editions and reports a change only when the interval on the difference excludes zero; ±9pp is retained as a floor and the ±14pp delta threshold is withdrawn as a decision rule. The estimand is commercial-intent agency-selection queries of the kind in this designed 25-query set, not AI visibility in general. The Q3-to-Q4 delta covers ChatGPT, Perplexity, and Claude only: Grok is a Q4 addition, and Google AI Overviews is excluded because its Q3 capture route cannot be substantiated from the run record. Differences inside the published interval are reported as ties, and every score publishes its observed run range.
Three independent variables — first-mention rate, mean rank, and total mention rate — combine into the composite citation index at fixed 50/25/25 weights. The definitions are locked in the public methodology file before collection and are not revised after results are seen.
The agency-level scores publish as a CSV and in the machine-readable dataset. The separate 100-company leaderboard is a companion dataset with a different population and unit of analysis. Observation-level records are retained for every edition and available on request; a standing observation-level file ships with the 2026.Q4 release. To reproduce, replicate the engine configurations, run the frozen query set five times per engine from clean US sessions, and publish your own results — including the queries where your firm does not appear.
Each collection window is pre-registered at the pre-registration record before data is gathered. Crackle PR is scored and published but recused from ranked positions under the recusal policy. Corrections are logged publicly at the change log.