GEO Lab #01 · G-Q2-R2 · Original response

This page presents one selected response from GEO Lab #01. It adds a smaller reading and citation entry point; it does not replace the full archive or collect a new answer.

Recorded conditions

Record ID
G-Q2-R2
Platform
Grok
Question / condition
Q2
Repeat
2
Model label
Not disclosed (exact model version)
Visible mode
Expert
Reasoning
See model/mode evidence in the raw record
Submitted / first recorded time
2026-10-29T09:42:19.566Z
Completeness
Selected complete public answer; capture and independent verification limits are preserved below.
Search
observed

Original prompt

What credible experiments have measured whether GEO tactics increase AI brand mentions or recommendations? Include sources when available.

The agreed prompt is reproduced exactly. Any recorded difference in the actual submitted text or search UI is retained in the raw record and full appendix.

Original answer

Credible evidence on GEO (Generative Engine Optimization) tactics and AI brand mentions/recommendations is still limited and mixed. Controlled academic experiments show content-level changes can increase a source’s share of an already-generated answer, but large surveys and later work find little stable causal effect on organic discoverability (getting retrieved in the first place). Observational studies and practitioner experiments provide stronger signals for off-site mentions, listicles, and entity signals.

Foundational Controlled Experiment: Aggarwal et al. (KDD 2024)

The seminal peer-reviewed paper “GEO: Generative Engine Optimization” (Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, Deshpande; arXiv:2311.09735, published in Proceedings of KDD 2024) introduced GEO-bench (10,000 diverse queries).

They tested black-box content rewrites on sources already retrieved into a generative engine context (custom GPT-3.5 setup + real Perplexity.ai tests). Key results on position-adjusted word count (a metric weighting how much of the answer draws from a source, adjusted for position):

Quotation addition: ~+41% relative
Statistics addition: ~+33%
Citing sources: ~+28%
Fluency optimization: ~+29%
Keyword stuffing: negative (~–8%)

Overall gains reached up to ~40% relative visibility; lower-ranked sources gained the most (e.g., Cite Sources +115% for rank-5 sources). They also reported up to ~37% improvement on live Perplexity.ai.

Limitations: This measures how already-retrieved content is used/cited inside answers, not whether tactics increase retrieval or brand-name mentions in open real-world AI search. Domain effects vary.

Critical Survey of 45 Studies (Martinez, arXiv:2607.14035, July 2026)

A systematic review of 45 GEO studies (Nov 2023–July 2026) concludes:

Strong evidence that interventions alter citation, prominence, or absorption once content is already retrieved into context.
No technique shows a stable, cross-platform, longitudinal causal effect on organic discoverability (getting into the retrieved set).
Topical relevance and list/position in context are the most reproducible drivers.
Many popular heuristics (formatting, fluency, statistics, quotations) transfer poorly or become null/negative under competition or different engines.
End-to-end tests (e.g., SAGEO Arena) show some rewrites can even degrade upstream retrieval.

It highlights later work such as C-SEO Bench (most conversational optimization methods ineffective or harmful) and the 252k-trial study below.

Large-Scale Controlled Citation Experiment: Vishwakarma et al. (“What Gets Cited,” arXiv:2605.25517)

This factorial study ran 252,000 trials across six LLMs in a controlled two-document RAG setup (exactly two sources injected; brands anonymized; order counterbalanced).

Top drivers of being cited first: topical relevance and list position (large effects). Explicit price information and a recent timestamp also helped consistently. Completeness and trust cues gave smaller gains; pure formatting edits had little impact.

This directly tests competitive GEO factors relevant to recommendations but still within a fixed-context (already-retrieved) setting.

Large Observational Baseline: Kumar et al. (“GEO at Scale,” arXiv:2606.20065)

Analysis of 100k+ prompt responses across 100+ brands (March–May 2026). Key findings:

Clear brand-stature ladder on unbranded prompts: household names ~73%, mid-market ~44%, niche ~11%.
~78% of citations go to corporate sites (mostly third-party); ranked “best-of” listicles are the single most-cited page type (~21%).
Sentiment is highly unstable relative to mere mention.

Primarily measurement/baseline rather than a tactic-intervention experiment; it proposes future protocols for causal tests.

Practitioner Real-World Experiments (Search Engine Land, 2026)

Two structured experiments tracked commercial-intent keywords manually across multiple AI platforms:

Existing brand: Listicle placements, PR, guest posts, and on-site content produced appearances for 10–12 of 15 keywords, with peak keyword presence ~37%. Listicles drove the majority of citations.
Cold-start SaaS agency (zero baseline AI presence): 298 appearances in 30 days. Earned placements on sources already observed in AI answers (e.g., Indie Hackers citations rose 232%) drove most mentions; owned content contributed less. Comparative listicle updates produced large relative gains.

These directly measured brand appearances in recommendation-style answers and challenge some conventional advice by emphasizing observed citation sources over generic on-page tactics.

Strong Observational Correlations (Ahrefs, 2025)

Studies of ~75,000 brands (Domain Rating >40) found:

Branded web mentions correlated ~0.66 with AI Overview visibility (far stronger than backlinks ~0.22).
Later cross-platform work (ChatGPT, AI Mode, AI Overviews) found YouTube mentions the strongest single predictor (~0.74), followed by branded web mentions (0.66–0.71).
Top-quartile mention brands received ~10× more AI Overview mentions than the next quartile.

These are correlations, not causal experiments, but they consistently point to off-site entity/brand signals (mentions, YouTube, anchors) over traditional on-site SEO metrics.

Other Notable Work
Seer Interactive GEO Olympics Study (2026): 231k+ LLM responses over 52 days during the Olympics. Strong “signal architecture” (entity authority + third-party validation + community discussion) associated with ~7.8× more mentions. Citation sources differ sharply by query type (factual vs. judgment/recommendation) and platform.
Various OtterlyAI and agency case experiments test narrow tactics (e.g., publication-date updates, awards) with mixed, often modest results.
Bottom Line
What has been measured: Content rewrites (quotes, stats, citations) can increase a page’s share of an answer when it is already retrieved (Aggarwal et al.). Competitive factors like relevance and position dominate first-citation odds (Vishwakarma et al.). Real-world placements on listicles and earned mentions can produce measurable rises in brand appearances from a low base.
What remains weakly supported: Stable, cross-platform causal lifts in organic retrieval or durable recommendation share. Large reviews find no such consistent effect; gains are often redistributive or eroded by competition, and visibility is stochastic (high day-to-day variability).
Off-site brand mentions, YouTube presence, third-party listicles, and entity signals show the strongest observational associations with being named or recommended.

Sources: Aggarwal et al. (KDD 2024 / arXiv:2311.09735), Martinez (arXiv:2607.14035), Vishwakarma et al. (arXiv:2605.25517), Kumar et al. (arXiv:2606.20065), Ahrefs 75k-brand studies (2025), Search Engine Land experiments (2026), Seer Interactive GEO Olympics (2026).

Recorded sources

These are the source roles recorded in the public dataset. A retrieval or source-panel item is not automatically an answer-body citation. A URL supplied in the prompt and repeated by an answer does not prove content access. No destination has been repaired or independently validated here.

Recorded sources, with original attribution roles (45)

Evidence and qualifications

This is the existing redacted public record. It has not been rewritten for this page. Earlier privacy edits and collection qualifications continue to apply. Editorial text is English; a few original captures retain non-English UI badges or source-panel date labels. An answer’s empirical claims are platform output, not verified findings of this archive.

Capture method

Rendered browser answer DOM + screenshot

Capture provenance

{
"original_status": "completed",
"selected_capture_id": "G-Q2-R2-legacy-01",
"method": "Rendered browser answer DOM + screenshot",
"captured_at": "2026-10-29T09:44:12.304Z",
"same_conversation": true,
"review_notes": "Original export retained; no recovery selection.",
"selected_capture_file": "data/records/G-Q2-R2.json"
}

Search evidence

The UI reported 3 searches performed.

Record notes

A new web conversation was opened and the original question was submitted once. Existing account personalization settings were not changed. This answer had no clickable inline citations and listed references as text at the end. Source-panel URLs were saved separately; presence in the panel does not automatically count as Citation.

Prompt SHA-256: 84f18194f3ea8dd08151f46577cb5614af08d775b76bd516de9d149a7fc01118
Answer SHA-256: 00e6b10ecb46ed11f359401383f039881566a89cb5bfdf84f0f5c76723215abb