GEO Lab #04: Which creators and content brands get recommended?
18 complete selected answers / 18 planned · 18 recorded submissions · 0 positions not yet complete.
OpenAI API + native web search · gpt-5.6-sol / medium versus Grok Expert · 3 November 2026 The API condition does not represent consumer ChatGPT Search.
This experiment describes whether three historically selected creators or their associated content assets appear in a new short-term sample. It does not establish why an AI recommends them, prove a GEO tactic works, or measure business conversions.
Main findings
Recommendations depended on the user’s question. For Q1, Grok recommended Aleyda’s resources and The GEO Lab in 3/3 answers each; the API recommended Aleyda’s resources in 1/3. For Q3, the API recommended Kevin in 3/3, Aleyda or her resources in 1/3, and The GEO Lab in 1/3; Grok recommended Artur/The GEO Lab in 2/3. All three candidates had zero C/M/R in the six Q2 answers. These are counts of a person or a confirmed associated asset, not personal-name recommendations alone.
Content-brand exposure and personal exposure were different outcomes. Across the mixed nine-answer positions per environment, Grok recommended The GEO Lab in 5/9 answers but named Artur in only 2/9; the API recommended the site in 1/9 without naming him. Explicit person recommendations were 2/9 and 0/9 respectively. These pooled counts describe this prompt bank, not a general recommendation probability.
Three repeats exposed changes hidden by the first two answers. Grok’s first two Q3 answers recommended Artur; its third recommended other creators. Kevin appeared in all three API Q3 answers. These are short-term counts within this run, not stable recommendation probabilities.
A strong AI endorsement did not establish research quality. Grok described Artur’s manual as rigorous, yet the inspected edition still mixes 75 runs per version with results using 30, and pairs p=.07 with significance wording. Its revision note acknowledges earlier transcription errors, but the chapter remains inconsistent. The useful content structure and unresolved evidence problem are both retained.
Recommendation and own-content citation were different outcomes. Across 54 candidate-by-answer observations, 14 had M and R, but only 10 had a strict linked own/bylined C. Four recommendations lacked that specific citation type: two relied on third-party coverage, while two gave The GEO Lab as a plain-text address. Including the latter addresses would raise Artur’s Grok C count from 3/9 to 5/9; the primary strict rule is retained. No C-only or M-only observation occurred among these three selected families.
The pages suggest practices to test, not proven causes of visibility. Eight actual own/bylined pages show three approaches: Aleyda’s curated learning routes, Kevin’s data-led analysis and transcripts, and Artur’s linked guides and experiment templates. Clear author identity, a question-based index, transparent methods and reusable worksheets are feasible candidates for Helen’s next tests. Established audiences, experience and partner data remain background advantages.
OpenAI API + native web search (gpt-5.6-sol, medium) is a different interface from consumer ChatGPT Search. The historical Stage 1 UI results cannot establish a pure longitudinal platform change.
Three of nine API answers reached the fixed eight-tool-call cap: all three Q2 repetitions. Their answer bodies are complete, but retrieval may have been constrained. Search controls and interfaces differ, so this is a comparison of the recorded environments rather than an isolated model preference.
Each question × environment has only three repeats collected on one date. No un-recommended creator control group, content intervention, longitudinal stability or business conversion was tested.
C/M/R counts are assistant observational judgments, not independent human double coding. The audit checks selected pages and attribution; it does not verify every factual claim in all 18 model answers.
Results by user question · primary comparison
Question
Creator / associated assets
OpenAI API · person-or-asset R
Grok · person-or-asset R
Q1
Aleyda Solis / Learning AI Search and related assets
1/3
3/3
Q1
Kevin Indig / Growth Memo and related assets
0/3
0/3
Q1
Artur Ferreira / The GEO Lab
0/3
3/3
Q2
Aleyda Solis / Learning AI Search and related assets
0/3
0/3
Q2
Kevin Indig / Growth Memo and related assets
0/3
0/3
Q2
Artur Ferreira / The GEO Lab
0/3
0/3
Q3
Aleyda Solis / Learning AI Search and related assets
1/3
0/3
Q3
Kevin Indig / Growth Memo and related assets
3/3
0/3
Q3
Artur Ferreira / The GEO Lab
1/3
2/3
Each result uses three answers to the same original question in one environment. Q1, Q2 and Q3 represent different user needs and are not interchangeable measurements.
Comparison limit: 3/9 API answers reached the fixed 8-tool-call upper bound; all three were Q2 answers. Answers are complete, but retrieval may have been constrained. Interfaces and retrieval controls differ, so source-count differences cannot be assigned to model preference.
C = answer-body hyperlink to confirmed own/bylined content; M = explicit person or verified associated-asset mention; R = active recommendation. These overlapping indicators are not mutually exclusive.
Person name M means the personal name appears. Person R means the person is actively recommended. Person-or-asset R includes either target; a resource-brand recommendation alone is not a personal recommendation.
Q1 · Free learning resources
What are the best free resources to learn Generative Engine Optimization (GEO) in 2026? Include sources when available.
Creator / associated assets
OpenAI API · C/M/R
OpenAI API · person only
Grok · C/M/R
Grok · person only
Aleyda Solis / Learning AI Search and related assets
C: 1/3 M: 1/3 R: 1/3
Person name M: 1/3 Person R: 0/3
C: 2/3 M: 3/3 R: 3/3
Person name M: 3/3 Person R: 0/3
Kevin Indig / Growth Memo and related assets
C: 0/3 M: 0/3 R: 0/3
Person name M: 0/3 Person R: 0/3
C: 0/3 M: 0/3 R: 0/3
Person name M: 0/3 Person R: 0/3
Artur Ferreira / The GEO Lab
C: 0/3 M: 0/3 R: 0/3
Person name M: 0/3 Person R: 0/3
C: 1/3 M: 3/3 R: 3/3
Person name M: 0/3 Person R: 0/3
Comparison limit: 3/9 API answers reached the fixed 8-tool-call upper bound; all three were Q2 answers. Answers are complete, but retrieval may have been constrained. Interfaces and retrieval controls differ, so source-count differences cannot be assigned to model preference.
Q2 · Credible experiments
What credible experiments have measured whether GEO tactics increase AI brand mentions or recommendations? Include sources when available.
Creator / associated assets
OpenAI API · C/M/R
OpenAI API · person only
Grok · C/M/R
Grok · person only
Aleyda Solis / Learning AI Search and related assets
C: 0/3 M: 0/3 R: 0/3
Person name M: 0/3 Person R: 0/3
C: 0/3 M: 0/3 R: 0/3
Person name M: 0/3 Person R: 0/3
Kevin Indig / Growth Memo and related assets
C: 0/3 M: 0/3 R: 0/3
Person name M: 0/3 Person R: 0/3
C: 0/3 M: 0/3 R: 0/3
Person name M: 0/3 Person R: 0/3
Artur Ferreira / The GEO Lab
C: 0/3 M: 0/3 R: 0/3
Person name M: 0/3 Person R: 0/3
C: 0/3 M: 0/3 R: 0/3
Person name M: 0/3 Person R: 0/3
All three API Q2 answers reached the eight-call cap. This is a retrieval-budget and interface confound, not evidence that an environment intrinsically prefers or rejects these authors.
Q2 requests credible experiments on effects of GEO tactics; Q1 and Q3 ask more broadly for resources or practical tutorials. No C/M/R for these three candidates in Q2 does not establish that they lack research ability.
Q3 · Practical creators
Who publishes practical, data-backed GEO tutorials for independent builders? Explain why and include sources when available.
Creator / associated assets
OpenAI API · C/M/R
OpenAI API · person only
Grok · C/M/R
Grok · person only
Aleyda Solis / Learning AI Search and related assets
C: 1/3 M: 1/3 R: 1/3
Person name M: 1/3 Person R: 1/3
C: 0/3 M: 0/3 R: 0/3
Person name M: 0/3 Person R: 0/3
Kevin Indig / Growth Memo and related assets
C: 2/3 M: 3/3 R: 3/3
Person name M: 3/3 Person R: 3/3
C: 0/3 M: 0/3 R: 0/3
Person name M: 0/3 Person R: 0/3
Artur Ferreira / The GEO Lab
C: 1/3 M: 1/3 R: 1/3
Person name M: 0/3 Person R: 0/3
C: 2/3 M: 2/3 R: 2/3
Person name M: 2/3 Person R: 2/3
Comparison limit: 3/9 API answers reached the fixed 8-tool-call upper bound; all three were Q2 answers. Answers are complete, but retrieval may have been constrained. Interfaces and retrieval controls differ, so source-count differences cannot be assigned to model preference.
These small denominators are short-term observations, not long-term stability or a pure platform effect. All three candidates are displayed, including zeros and Aleyda’s API Q3 1/3 result.
Personal name and content brand are different outcomes
A content brand can be recommended while its author remains unnamed. The GEO Lab received person-or-asset R in 5/9 Grok answers, but Artur’s name appeared in 2/9; the API recommended the site in 1/9 without naming him.
Creator / associated assets
Environment
Person-or-asset R
Person name M
Person R
Aleyda Solis / Learning AI Search and related assets
OpenAI API + web search
2/9
2/9
1/9
Aleyda Solis / Learning AI Search and related assets
Grok
3/9
3/9
0/9
Kevin Indig / Growth Memo and related assets
OpenAI API + web search
3/9
3/9
3/9
Kevin Indig / Growth Memo and related assets
Grok
0/9
0/9
0/9
Artur Ferreira / The GEO Lab
OpenAI API + web search
1/9
0/9
0/9
Artur Ferreira / The GEO Lab
Grok
5/9
2/9
2/9
The denominator of nine mixes three different user questions. Use the question-level results above for intent-specific comparisons.
Comparison limit: 3/9 API answers reached the fixed 8-tool-call upper bound; all three were Q2 answers. Answers are complete, but retrieval may have been constrained. Interfaces and retrieval controls differ, so source-count differences cannot be assigned to model preference.
For Helen Bird, future measurements should distinguish a helen-bird.github.io page citation, a GEO Lab content-brand recommendation, and an explicit Helen Bird name mention or recommendation. Clear and consistent author attribution is a candidate change to test; its effect is not established here.
All-question C/M/R · secondary summary
Secondary pooled person-or-asset recommendations. The question-level results above are primary. Bars use nine planned answers per environment; they do not count explicit personal recommendations alone.
Aleyda Solis / Learning AI Search and related assets
OpenAI API + web search2/9
Grok3/9
Kevin Indig / Growth Memo and related assets
OpenAI API + web search3/9
Grok0/9
Artur Ferreira / The GEO Lab
OpenAI API + web search1/9
Grok5/9
Comparison limit: 3/9 API answers reached the fixed 8-tool-call upper bound; all three were Q2 answers. Answers are complete, but retrieval may have been constrained. Interfaces and retrieval controls differ, so source-count differences cannot be assigned to model preference.
Creator / associated assets
OpenAI API · C/M/R
Grok · C/M/R
Aleyda Solis / Learning AI Search and related assets
C: 2/9 M: 2/9 R: 2/9
C: 2/9 M: 3/9 R: 3/9
Kevin Indig / Growth Memo and related assets
C: 2/9 M: 3/9 R: 3/9
C: 0/9 M: 0/9 R: 0/9
Artur Ferreira / The GEO Lab
C: 1/9 M: 1/9 R: 1/9
C: 3/9 M: 5/9 R: 5/9
Secondary pooled view: nine planned answers per environment across three different questions. It describes this prompt bank, not a general recommendation probability. Unknown authorship remains unknown C.
Comparison limit: 3/9 API answers reached the fixed 8-tool-call upper bound; all three were Q2 answers. Answers are complete, but retrieval may have been constrained. Interfaces and retrieval controls differ, so source-count differences cannot be assigned to model preference.
Separate reference pages from experiment logs; publish version notes.
Create one public index with task labels, suggested reading order and authorship.
Publish a blank measurement sheet and distinguish templates from completed results.
A page’s appearance in a retrieval list is not a body citation. The source audit prioritizes two to three distinct actual body-cited own/bylined pages per creator; fewer eligible pages are reported as a shortfall rather than replaced with presumed successes.
Any wider context page or distribution evidence is supporting audit material. It does not become a new holdout citation or prove which channel caused an AI recommendation.
What the actual pages show
Assistant reading via public web retrieval; may reflect crawled representations, not historical snapshots. No raw datasets independently replicated.
Separate reference pages from experiment logs; publish version notes.
Existing library and research infrastructure.
The method/result denominator conflict remains unresolved in this page. p=.07 does not meet a conventional .05 threshold; the significance wording needs qualification. No raw-data replication or exact-page independent citation established.
readable_via_web_retrieval About These Guides and author footer attribute the guides to Artur Ferreira.
Observed in new answers
S4-GROK-Q3-R1, S4-GROK-Q1-R3
Why this page was selected
Library hub actually cited in both creator and free-learning questions.
Evidence and practical content
Available Ebooks; FAQ: A task-based library links fundamentals, framework, experiments, workbook and technical guides, then explains starting choices and update versions.
Print note; FAQ on price: The page describes browser-print PDF saving and also Gumroad zero-price checkout requiring email. Download/checkout was not tested.
Content organization
A discoverable hub turns individual guides into an explicit learning sequence with audience-specific entry points.
Create one public index with task labels, suggested reading order and authorship.
An already-developed multi-guide catalog.
This index supplies no new experimental sample. Its claimed proven framework is promotional characterization, not validation by this audit. Independent citations of this exact hub were not verified.
readable_via_web_retrieval Embedded guide explicitly identifies Artur Ferreira and a 2026 edition.
Observed in new answers
S4-OPENAIAPI-Q3-R3
Why this page was selected
Own-site body citation returned by the API Q3 answer.
Evidence and practical content
Live Data; How The GEO Lab Runs Experiments: Headline reports 330 queries across ChatGPT, Gemini and Perplexity; a separate results table links studies, including E001 (18/30 vs 11/30, p=.07) and a 480-query FAQ null result. These are different studies, not one combined sample.
Data Collection Template: Provides exact-query, date, engine, citation URL and quoted-passage fields; its example uses ten queries across three engines.
Content organization
Hypothesis-to-report workflow, ten test templates and a table linking conclusions to individual studies.
Publish a blank measurement sheet and distinguish templates from completed results.
Multiple existing studies and a proprietary Console.
No Grok appears in the 330-query headline. Underlying datasets were not independently reanalysed. Independent citations of this exact guide were not verified.
Accessible public HTML, checked 2026-11-03 Site explicitly attributes creation and funding to Aleyda Solis.
Observed in new answers
S4-GROK-Q1-R2, S4-GROK-Q1-R3, S4-OPENAIAPI-Q1-R3
Why this page was selected
Most recurrent own-content citation; tracking parameters collapsed for page selection only.
Evidence and practical content
Explains a free curated roadmap and an editorial submission process; declares no advertising or paid inclusion. This is a policy statement, not an original GEO experiment.
Content organization
Explicit author identity and editorial policy connected to the learning roadmap.
Accessible public HTML, checked 2026-11-03 Curated by Aleyda Solis
Observed in new answers
S4-GROK-Q1-R3
Why this page was selected
Actual homepage body hyperlink shows the learning asset itself, beyond the recurrent About page.
Evidence and practical content
Version 4 is dated March 20, 2026. Topic navigation spans fundamentals, measurement, commerce and tools; a copyable checklist and FAQs support action. No controlled original experiment appears on this hub.
Content organization
A versioned hub links user questions to specialist pages and a practical checklist.
Create a small question-based hub; link each lesson to evidence and a free checklist.
Existing expert and newsletter connections are not reproduced by copying the layout.
Curation depth depends on external resources. The visible revision date does not prove every item is current; this audit does not endorse all optimization advice.
Accessible public HTML, checked 2026-11-03 LearningSEO.io About page identifies Aleyda as its founder and curator; individual linked resources have their own authors.
Observed in new answers
S4-OPENAIAPI-Q3-R2
Why this page was selected
A distinct body-cited asset whose creator attribution needed independent verification.
Evidence and practical content
Six topic sections distinguish research, guides, official documentation and tools, naming individual authors. It curates evidence rather than presenting one unified original experiment.
Content organization
An AI-search section sits within a broader SEO learning system.
Separate concise findings from supporting methods.
Long SEO career and client/event access, self-reported.
Research synthesis and opinion, not a reproducible experiment at this URL.
Source popularity and recommendation causation were not tested.
Artur: Independent third-party citations of the three exact selected pages were not verified within one focused search.
Artur: No original data replication, download/checkout, browser rendering, backlink inventory or creator business impact was measured.
Kevin: Two distinct own/bylined pages appeared in the new answer-body links. No third eligible page was added; the Danny Goodwin report and its linked original remain separately labeled context.
Kevin: The Search Engine Land article gives different broad corpus totals in its introduction and data section; it is not an independently reconciled dataset.
Eight distinct own/bylined pages were selected: three Artur, three Aleyda and two Kevin. A two-page sample meets the registered 2–3 range; lack of a third Kevin page is not a missing response.
Distribution searches were bounded and purposive; no backlink totals, reach or channel attribution were measured.
Selected creators have no matched un-recommended control group; these visible features do not explain selection causally.
Creator-specific audit files retain additional access and attribution notes. No underlying study dataset was independently replicated.
Distribution evidence, with boundaries
These are creator-level distribution observations, not extra holdout citations or independent verification of the research. Owned posts, external indexes, editorial coverage and unverified leads have different evidentiary roles.
Author announces a recurring 14-experiment series, using measurement, tools, results and defensible conclusion as the post format. This is self-distribution, not independent endorsement.
Access: readable_via_web_retrieval
Self-distribution; not independent endorsement.
A citation of a selected audit page was not verified.
Synapse lists a related paper by Artur Ferreira and Julio Arévalo Piedra and links DOI 10.5281/zenodo.23122087. It republishes the authors’ abstract; this verifies external indexing, not independent confirmation or a citation of the selected guides.
Access: readable_via_web_retrieval
Indexing and an author abstract; not independent validation.
A citation of a selected audit page was not verified.
Verified editorial coverage by Danny Goodwin links Kevin’s research to its Growth Memo original. This documents distribution of another work, not either selected page.
A podcast interviews Aleyda about creating LearningSEO.io and links to it. This is a distribution example, not evidence about the current AI page's search ranking.
Related project coverage; not a verified citation of the selected AI-search subpage.
Search excerpt named Artur as a joint researcher; the opened page instead displayed Julio-focused material without that attribution. Excluded as corroborated independent distribution.
A citation of a selected audit page was not verified.
Supporting context, outside the holdout sample
The following pages were followed during the audit for authorship, content organization or background. They remain separate from the 8 selected own/bylined pages and do not increase the holdout citation counts.
Linked self-description; background context, not independent credential validation.
Lists prior Shopify, G2 and Atlassian roles and professional speaking.
Supporting context only; not an additional selected holdout-cited page or an independently validated causal explanation.
Observable features → hypotheses to test
Observable feature
Untested hypothesis
Practical next test
Evidence boundary
Aleyda: free learning routes, task-based topic links and a practical checklist.
A question-based learning hub may help people and retrieval systems locate the relevant experiment.
Create a small GEO Lab index with verified report/data links. Record a URL-free prompt baseline, then repeat fixed prompts on multiple dates; retain a comparison page where feasible.
Untested: the audit observed organization, not its effect on recommendations.
Kevin: explicit study methods and denominators; the same research is also available as a searchable talk transcript.
Inspectable evidence and a second readable format may broaden useful entry points.
Test one change at a time on matched pages: a compact methods block or a transcript. Keep prompt wording fixed and measure page retrieval and strict linked C separately from R.
Untested: existing reputation, audience and data partnerships remain confounds.
The GEO Lab: connected experiment pages, manuals, templates and revision notes.
A linked experiment directory with transparent versions may help users reach the right evidence.
Add cross-links and a standard protocol/data/correction template to a bounded page group. Compare repeated fixed prompts with a comparison group; preserve the pre-change versions.
Untested: useful structure does not resolve the manual’s denominator and significance inconsistencies.
The GEO Lab’s recommendations exceeded Artur’s explicit name appearances; audited assets provide different levels of author attribution.
Consistent author attribution across owned assets may improve recognition of Helen Bird as the creator.
Record separate baselines for page C, GEO Lab content-brand R, Helen Bird name M and Helen Bird person R. Introduce one documented authorship change and repeat the same prompts with a comparison page.
Untested: asset recommendation and personal recognition are different outcomes; no identity intervention occurred here.
Reputation, existing audience, years of professional experience, paid data partnerships and distribution access are background advantages. Copying a visible page feature does not reproduce those advantages or establish a causal GEO mechanism.
Why these three candidates?
Candidate
Stage 1 recommendation observations
Identity qualification
Aleyda Solis
ChatGPT Search Q3: 2/2 Grok Q1: 2/2
Personal name and associated content assets were historically grouped; the new sample separates them.
Kevin Indig
ChatGPT Search Q3: 2/2
Personal name and associated content assets were historically grouped; the new sample separates them.
Artur Ferreira
Grok Q1: 2/2 Grok Q3: 1/2
Q1 names the resource brand The GEO Lab without naming Artur; Q3 R1 explicitly names both.
Historical selection is not independent proof of future visibility. Stage4 API differs from Stage1 consumer ChatGPT Search. Aleyda cross-platform counts come from different questions.
The candidates were selected using Stage 1. Only new answers contribute to Stage 4 counts; the old two-repeat observations are not reused as validation samples.
Experiment conditions
Environment
Model / visible mode
Reasoning
Search condition
Actual search evidence
Submission window
OpenAI API + web search
gpt-5.6-sol
medium
Native web_search; tool_choice required; live access; actual tool calls stored separately from citations.
9 observed / 9 planned
2026-11-03 17:01:42 UTC+08:00 to 2026-11-03 18:54:09 UTC+08:00
Grok
Expert
Not independently selectable
Native retrieval; record actual controls, search queries and executed retrieval separately. No textual search instructions appended.
9 observed / 9 planned
2026-11-03 17:03:14 UTC+08:00 to 2026-11-03 18:58:00 UTC+08:00
OpenAI API + web search is not the consumer ChatGPT Search interface. Stage 1 used consumer ChatGPT Search; these records cannot isolate a longitudinal ChatGPT platform change.
API fixed upper bounds: 8 built-in tool calls; 32,768 generated output tokens. Search controls: tool_choice=required; live web access=true; store=false. 3/9 selected API answers reached the tool-call cap.
Grok: both X and conversation-history personalization were OFF in all recorded runs. Both were restored to their pre-test settings after collection.
All recorded same-environment submission gaps meet the 180-second minimum. Gaps were often longer; this was not strict time balancing.
Grok’s Expert label does not reveal its exact backend model or provide a directly equivalent reasoning setting. Configured search and actual retrieval evidence are distinct.
Submission times use Asia/Shanghai (UTC+08:00). Exact prompts are unchanged; no candidate name or past answer is added.
Data coverage and preserved attempts
Environment
Planned positions
Recorded submissions
Complete selected answers
Submissions not completed / reviewed
OpenAI API + web search
9
9
9
0
Grok
9
9
9
0
All original submission records, including partial or failed attempts, remain in the local evidence set. No missing answer is silently replaced or treated as a negative discovery result.
Estimated OpenAI API spend: US$2.417703 against the user-approved US$10 cap. This usage-priced estimate is not a reconciled provider invoice.
GEO Lab #04 · Redacted public learning record · Descriptive, small-sample observations. No causal or long-term visibility claim.
Redacted public edition. All 18 original answers, real submissions and assistant coding are available below. Private account UI, original preference values, conversations and service identifiers remain local. Necessary source-URL tracking removal is documented.
Author: Helen Bird · Experiment date: 2026-11-03 · First published: 2026-11-03 · Last revised: 2026-11-06 · Version: v1.2
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.