GEO guide · Measurement guide · Download aggregate JSON

ARCADE: CRM discovery findings and methods

Selected first-party aggregate evidence for Gwenth’s research-led articles. Collection: 11–13 September 2026. Numerical method: 22 September. Extract prepared: 25 September.

This is not the complete raw dataset or an independent reproduction. The source corpus remains in the project repository. Human semantic validation of the labels is incomplete. The figures are descriptive, not product recommendations or a test of GEO interventions.

Study design and selection

Six buyer needs × three paraphrases × three runs × four consumer AI platforms = 216 selected answers, with 54 Google snapshots. The original wave had 204 eligible AI answers. Twelve later Claude answers complete the view after 16 supplement attempts; four unsuccessful supplement attempts remain preserved. There were 303 total collection attempts, not 303 independent answers. The current view is conditional on retry selection and date.

Natural questions did not force browsing or citations. The July forced-citation study is separate. Intended meaning-equivalence does not make this a representative sample of how buyers ask questions.

Definitions

Jaccard

size(intersection) / size(union); one empty set gives 0; two empty sets give undefined, not 0 or 1.

Mention

An entity named in answer content; not necessarily recommended.

Citation

A visible source link; not proof of complete retrieval or claim-level support.

Recommendation

A product positively presented as an option. Primary definition includes clear and hedged positive recommendations, not integration-only or source-only mentions.

Source Presence

A domain counted once per eligible answer, irrespective of repeated links within that answer.

RQ1: visible source presence

Automindz recruitment presence is 2/9 for ChatGPT, 0/9 for Claude, 3/9 for Gemini and 9/9 for Perplexity. Citation-bearing answer counts are 54/54, 25/54, 52/54 and 54/54 respectively across the whole study. Publisher ownership is distinct from content format; unknown ownership stays unknown.

RQ2a: exact-repeat stability

Average 3 exact-repeat pairs per exact question, then equally weight available questions within each platform.

Every platform has 54 eligible recommendation pairs across 18 questions. Claude has fewer eligible citation pairs because both-empty citation comparisons are undefined. Recommendation membership is clear-plus-hedged positive options; no semantic rank is inferred from display position.

RQ2b: wording beyond the repeat baseline

Mean across-paraphrase Jaccard distance minus mean exact-repeat distance, within buyer-need/platform blocks; equal weighting of prompt and phrasing-pair cells. Mean delta: 0.0543074852. Of 24 blocks, 15 are positive, eight negative and one zero. Positive does not imply statistically significant or commercially large.

RQ2c: cross-platform variation

All 3 x 3 cross-platform run pairs per exact question, then equal question weighting. Reused answers are not independent observations.

RQ3: individual AI versus direct-organic Google

Individual AI recommendation set versus direct-organic vendor brands in the same exact-question/repetition Google snapshot. 3 matched runs per question/platform; 3 questions per buyer need/platform; equal weight across 4 platforms and 6 buyer needs.

Overall mean: 0.1877376610. Direct-vendor visibility excludes brands appearing only within linked review-page bodies, ads or AI Overviews. The difference between 0.161 combined-union overlap and 0.188 individual-answer overlap is a change of estimand, not a temporal improvement. Linked-page, ad and AI Overview layers are separate; later linked-page overlap still uses the combined-AI secondary unit.

Targeted corrections and interpretation

The 21 September release corrected 21 brand records across 15 selected responses. A NetSuite existing-stack mention was not a recommendation; WorkHorse source ownership was corrected to CRM vendor. These are targeted adjudications, not a full annotation error-rate estimate. Old evidence and earlier releases remain preserved.

Limitations that apply to every table

• Fixed purposive question set, not a representative sample of buyers or all software discovery.
• Twelve later Claude answers are selected replacements; timing and retry selection are confounded.
• Repeated pair comparisons reuse answers and are dependent.
• Human semantic validation is incomplete; labels and vendor identity can still contain errors.
• Visible citations do not expose complete internal retrieval, and complete source-to-brand or sentence-to-source support was not established.
• Later linked-page retrieval cannot reconstruct original search-day page content.
• These are visibility observations, not product accuracy, traffic, conversion, causal GEO effects or Gwenth customer outcomes.
• The July forced-citation study, archived 19 September paper and older methods must not be pooled with this extract.

Provenance

Repository: cyberjj999/arcade. Snapshot: 76546e70ca5826573d4a7802032849272dbd4600. Current release: outputs/crm-rq3-individual-ai-google-20260922/. Original calculations and correction records were reviewed for the article work; the full collection and analysis pipeline was not rerun during this editorial extraction. The underlying repository is not presented as a publicly accessible citation.

ARCADE means AI Retrieval and Citation Analysis for Discovery Evaluation. Its purpose is to examine discovery behaviour, not decide which CRM is best. Gwenth is the commercial publisher of these educational articles; the project is not an independent product benchmark.

Read the annotation explainer · Read the SEO–GEO interpretation