Who does AI recommend, and whose content does it cite?

A small repeated-query comparison of ChatGPT Search and Grok, covering learning resources, empirical research and practical GEO tutorials.

  • Q3 recommendations shared only 2 of 20 unique entities/resources. The 10% overlap pools two answers per platform. The shared resources were GEO research (Aggarwal et al.) and The GEO Community. The lists mix papers, communities, organizations and individual creators; this is not a comparison of 20 creators.
  • Observed own-content citations accompanied 47/48 ChatGPT Search and 39/76 Grok recommendation records. These 97.9% and 51.3% figures cover coded entity-by-answer records and captured citations only. They do not measure answer accuracy or inherent platform citation preferences.
  • Two repeats produced different recommendation sets, with important condition differences. Neither platform had the greater overlap on every question. ChatGPT Q2 spans dates; Q3 also changes reasoning effort. Account memory was observed, and Grok’s backend conditions were undisclosed. These are descriptive overlaps, not estimates of stable consistency.
3 original questions × 2 platforms × 2 repeats = 12 selected complete answers. The archive also retains 10 other submissions and 37 captures; captures are not extra prompts. There was no optimization intervention, baseline or control group, so these observations cannot establish GEO effectiveness.

Recommended entity overlap

The selected recommendation sets differed for all three questions. Q3 had the lowest pooled cross-platform overlap in this batch.

Q1 · Discover learning resources

22.2%

8 shared / 36 unique entities

ChatGPT Search: 12; Grok: 32

Q2 · Find empirical research

38.5%

5 shared / 13 unique entities

ChatGPT Search: 10; Grok: 8

Q3 · Find practical tutorials and creators

10.0%

2 shared / 20 unique entities

ChatGPT Search: 12; Grok: 10
For each platform, pool the Recommendation=true names from both selected answers, ignoring only case and whitespace. Jaccard = shared names ÷ the union. This pooled statistic differs from the repeat-pair statistic below. Original prompts and strict labels are unchanged. Longer candidate lists do not establish higher recommendation quality.

Observed Repeat Overlap

Repeat comparisons are descriptive. Some runs differed in date or reasoning effort, and identical backend conditions could not be independently verified. ChatGPT Q2 spans October 29–30; Q3 changes both date and Extra High → Medium effort. Grok’s backend model version and reasoning effort were not disclosed. Two repeats cannot estimate a stable consistency level.

Q1 · Discover learning resources

ChatGPT Search58.3%
7 shared / 12 total entities
Recorded UI conditions match; backend conditions were not independently verified
Grok53.1%
17 shared / 32 total entities
Incomplete condition information; descriptive comparison only
Not fully recorded: Model label, Reasoning effort

Q2 · Find empirical research

ChatGPT Search20.0%
2 shared / 10 total entities
Different conditions; descriptive comparison only
Test date: 2026-10-29 → 2026-10-30
Grok62.5%
5 shared / 8 total entities
Incomplete condition information; descriptive comparison only
Not fully recorded: Model label, Reasoning effort

Q3 · Find practical tutorials and creators

ChatGPT Search41.7%
5 shared / 12 total entities
Different conditions; descriptive comparison only
Test date: 2026-10-29 → 2026-10-30; Reasoning effort: Extra High → Medium
Grok40.0%
4 shared / 10 total entities
Incomplete condition information; descriptive comparison only
Not fully recorded: Model label, Reasoning effort
Each bar compares one pair of selected answers to the identical original question, using Jaccard recommendation-set overlap. Recorded UI conditions matching does not verify matching backend conditions. Memory evidence and all six condition comparisons are preserved in the appendix.

Dates, UI model labels and reasoning effort for every selected answer

Citations, mentions and recommendations are separate observations

A citation does not automatically mean a recommendation. C records a citation to content published or bylined by the entity; M records an explicit name in the answer; R records an active suggestion to use, read or follow it.

PlatformOwn/bylined citation among R recordsObserved proportionR without observed own-content C
ChatGPT Search47/4897.9%1
Grok39/7651.3%37
The denominator is coded Recommendation=true entity × answer records. The same entity in two answers counts twice. Third-party support is excluded from this own-content indicator. Missing own-content C does not mean the answer lacks supporting evidence. Source capture and assistant coding coverage limit this comparison.

Combinations among coded entities

PlatformEntity-by-answer observationsObserved combinations
ChatGPT Search57CMR: 47; C: 5; CM: 2; M: 2; MR: 1
Grok85CMR: 39; MR: 37; M: 9
These are conditional inventories of coded main recommendations and substantive discussion subjects, not every brand or source publisher. Grok source-only publishers were not exhaustively coded. Omitted letters mean false under the original coding rules; the counts cannot estimate market-wide mention rates.

Complete C/M/R tables, examples and evidence

Resource types and a bounded source review

The recommendations include several resource types. A declared post-hoc classification separates individual creators, papers/studies, communities, organizations and tools rather than treating them all as products or creators.

A purposive audit checks five Q3 recommendations per platform for existence, original experiments, reported methods/sample sizes and fit for independent builders. All ten selected pages were readable at the check.

Relevance and original evidence are different checks. The selected Seer page interprets an academic study rather than establishing a Seer-run experiment. The selected GEO Lab manual describes a Perplexity experiment but reports inconsistent sample denominators and significance wording. These are page-specific qualifications, not judgments of either entity’s entire work.

This assistant audit is supplementary, not independent human validation or a representative quality ranking. An accessible page with a stated method does not verify all claims or reproduce the study. Original names, recommendation labels and the 2/20 overlap remain unchanged.

Type counts and the source-audit table · Type mapping · Audit notes

What this cannot show

  • No causal GEO conclusion: there is no content intervention, pre-optimization baseline, control group, predefined brand panel or conversion measure.
  • No inherent platform ranking: dates, reasoning effort, account context, source capture and undisclosed backend conditions limit the comparison.
  • No finding that an absent brand has poor GEO: answers depend on the sampled questions and conditions, and the mixed recommendation inventories are not a brand census.
The assistant coded the answers; no independent human double-coding review was completed. The separate source spot-check does not remove coding subjectivity or validate every external experiment. Multiple tutorials citing the same study are not independent replications.

Methods and evidence

Read the full technical appendix for original prompts, all 12 selected answers, the 10 other submissions, capture corrections and entity-level evidence. All appendix headings remain expanded.

Download redacted Stage 1 data · Computed findings · Publication scope · Repository and learning log

Unrelated personal-context passages are omitted and marked. Private conversations, screenshots and complete local evidence are excluded. Original prompts, recommended entities, citation attribution and numerical results are preserved. Public annotations are English; verbatim answers retain only marked privacy edits.
Stage 1 remains at its original website address. The technical appendix and versioned data preserve the evidence behind this shorter reader.

Author: Helen Bird · Experiment date: 2026-10-29 to 2026-10-30 · First published: 2026-10-30 · Last revised: 2026-11-06 · Version: v1.4

Technical maintenance: consistent canonical URLs, page-specific descriptions and social metadata, truthful author/series markup, and links to individual original-answer pages. Original answers, coding, data downloads and research outcomes are unchanged. This bundled update has no measured AI visibility effect. November 6 navigation update: linked the new technical learning record; original experiment data and findings remain unchanged.

Version metadata · Revision history · Cross-report review decisions · Stage 4 review decisions · Technical maintenance