GEO Lab #04 · S4-GROK-Q2-R1 · Original response

This page presents one selected response from GEO Lab #04. It adds a smaller reading and citation entry point; it does not replace the full archive or collect a new answer.

The OpenAI arm uses the Responses API with web search and does not represent consumer ChatGPT Search. Grok records identify their actual consumer-interface mode and available model label.

Recorded conditions

Record ID
S4-GROK-Q2-R1
Platform
Grok
Question / condition
Q2
Repeat
1
Model label
Exact backend version undisclosed
Visible mode
Expert
Reasoning
Not independently selectable
Submitted / first recorded time
2026-11-03T09:03:14.607Z
Completeness
Recorded body complete
Search
Actual search activity observed; exact controls and evidence remain in the raw record.

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

The most credible controlled experiments are academic benchmarks (especially the foundational Princeton/KDD 2024 GEO study), with supporting work from later papers, large-scale observational analyses, and some industry case studies. Evidence is stronger for increasing the share of text or prominence of already-retrieved content inside AI answers than for boosting organic retrieval, brand discovery, or cross-platform causal effects. A 2026 critical survey of ~45 studies concluded that GEO tactics reliably change how an already-retrieved page is cited/used, but no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability.⁠Geoaura

1. Princeton GEO Study (Foundational Peer-Reviewed Work)

Paper: “GEO: Generative Engine Optimization” by Pranjal Aggarwal et al. (IIT Delhi, Princeton, etc.), presented at KDD 2024 (arXiv:2311.09735).⁠arXiv

Design: GEO-bench with ~10,000 real-world-style queries across domains + relevant web sources. They tested nine content rewrite tactics on sources already in a fixed retrieval context (top results fed to a generative engine setup, also validated on Perplexity). Visibility was measured primarily as position-adjusted word count (share of the answer text attributable to the source, weighted by position) and a subjective impression metric.
Key results (relative lifts on position-adjusted word count vs. baseline):
Quotation/expert quotation addition: ~+41% (highest; e.g., 19.3 → 27.2 in reported figures).
Statistics addition: ~+33%.
Fluency optimization: ~+29%.
Citing sources: ~+28%.
Other style/language tweaks: smaller positive effects.
Keyword stuffing: −8% (negative).
Gains were largest for lower-ranked sources (e.g., rank-5 sources saw ~100%+ lifts with citation/quotation/statistics methods; top-ranked sources sometimes lost share). Effects varied by domain. On Perplexity they saw up to ~37% gains on some metrics.
Caveat: This measures absorption/visibility inside the answer once the source is already retrieved and in context. It does not test improving initial retrieval or discoverability. Later replications and critiques note the numbers are conditional on the fixed testbed and metric.⁠Geoaura

This remains the most-cited rigorous experiment establishing that certain content edits (evidence-rich additions) can increase a source’s presence in generated answers.

2. Product Recommendation Manipulation Experiment

Paper: “Manipulating Large Language Models to Increase Product Visibility” by Aounon Kumar and Himabindu Lakkaraju (arXiv:2404.07981, 2024).⁠arXiv

Design: Catalog of fictitious coffee machines. They added carefully crafted “strategic text sequences” (STS) to product pages and measured changes in LLM top recommendations for two target products (one rarely appearing, one typically ranking second).
Results: STS significantly raised the chance of both targets appearing as the top recommendation. This demonstrated that content on a product page can manipulatively increase LLM recommendation rates in a controlled setting.
Relevance: Direct measurement of increased product recommendations via GEO-style content changes.
3. C-SEO Bench (NeurIPS 2025 Datasets & Benchmarks)

Paper: “C-SEO Bench: Does Conversational SEO Work?” by Haritz Puerto et al. (arXiv:2506.11097).⁠arXiv

Design: Tested content rewrite methods (including several from the Princeton paper) plus traditional ranking approaches on question-answering and product-recommendation tasks across domains. Metric focused on citation rank in the output, with varying numbers of competing actors.
Results: Most content-level C-SEO/GEO rewrites were largely ineffective or negative on citation ranking. Traditional SEO approaches that improve a source’s ranking/position in the LLM’s context were substantially more effective. Gains diminished as more actors adopted the same tactics (zero-sum/congested).

This challenges some of the earlier positive findings when the metric shifts from word-share to ranking under competition.

4. Large-Scale Observational and Other Academic Work
Kumar et al. (arXiv:2606.20065, 2026): Analysis of 100,000+ prompt responses across 100+ brands. Established a clear brand-stature ladder (household names cited in ~73% of relevant answers, mid-market ~44%, niche ~11%). Ranked “best-of” listicles were a top citation format (~21%). Provides a baseline more than a controlled tactic experiment.⁠Geoaura
University of Toronto work (Chen, Koudas et al.): Multiple papers showing strong earned-media (third-party) bias over brand-owned content in AI search citations (often 6.5× advantage for third-party platforms). A 2026 preprint (“What Drives AI Citations?”) on 158k+ answers found that page-level tactics (FAQ blocks, schema, summary boxes, etc.) showed little or no within-website association with citations once retrieval is controlled; retrieval itself explained far more variance. Brand names also influence rankings even with identical facts.⁠Aivisibilityjournal
Seer Interactive GEO Olympics Study (2026): 231,347+ LLM responses across 7 platforms over 52 days of Olympics data. Tested hypotheses on signals (entity authority, third-party validation, community discussion). Strong signal stacks produced ~7.8× more AI mentions. Different sources drive listing vs. recommendation; platform-specific patterns (e.g., Reddit, editorial, official sites) were clear. Large-scale observational rather than pure controlled experiment.⁠Seerinteractive
Other: E-GEO e-commerce testbed; VoteGEO-Bench on electoral visibility; various GEO-Bench ranking-manipulation studies; Stacker/Scrunch earned-media distribution tests reporting large citation lifts from third-party placement.
Industry Experiments and Case Studies (Less Controlled, but Directional)
SIXT (via Peec AI, 2026): Controlled experiment editing partner listings (not their own site) on high-traffic travel domains. Full-treatment domains saw ~320% increase in citation frequency (adjusted ~308%); control rose only ~12%. Also gains in visibility and share of voice.⁠Peec
Search Engine Land experiments (2026): Multi-month and 30-day tests tracking commercial keywords across platforms. Reported upward trends in citations via PR, listicles, guest posts, and updates; one cold-start produced hundreds of appearances. One original conclusion did not fully hold on retest.⁠Searchengineland
OtterlyAI cold-start: New brand with zero authority reached competitive share of voice / rank ~7 via external mentions/directories in 14 days.
Ahrefs and others: Large correlation studies (e.g., 75k brands) linking unlinked brand mentions strongly to AI visibility (correlations ~0.65–0.71), stronger than raw backlinks.
Some schema/JSON-LD and certain page-feature tests have returned null or mixed results in controlled designs.

Overall takeaway from the evidence: Adding statistics, quotations, citations, and fluency improvements can increase a source’s share and prominence inside AI answers when the content is already retrieved (Princeton results). Being retrieved in the first place (via relevance, third-party mentions, listicles, earned media, and model prior knowledge) appears more decisive for brand mentions and recommendations. Effects are platform- and domain-dependent, often stronger for lower-ranked or lesser-known brands, and can be zero-sum under competition. Keyword stuffing is consistently unhelpful. Most commercial claims of large, general “GEO lifts” exceed what the controlled experiments rigorously support.

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 answer-body links (10)

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

Read full rendered answer, extract native visible DOM text and inline hrefs; copy button returned empty browser clipboard and was not used as evidence of body completeness.

Search evidence

Native activity shows search groups of 3 and 2 plus SIXT query; only four exact query labels captured so far, raw activity snapshot preserved.

Record notes

[
"Initial real-time connection warning cleared without clicking Retry or resubmitting.",
"No candidate names/assets appear in full answer body."
]

Body review

{
"reviewer": "assistant",
"complete_body_read": true,
"coherent_ending": true,
"inline_url_coverage_checked": true,
"external_claims_verified": false,
"independent_human_double_coding": false
}

Existing public-text privacy policy

{
"body_changed": false,
"policy": "Only actual URL tracking, signed/session URL parameters and necessary identifying information are removed; original wording, including analytics parameter examples outside URLs, remains.",
"annotation_positions": "Original position indexes omitted because URL/privacy edits can change character locations."
}

Prompt SHA-256: 84f18194f3ea8dd08151f46577cb5614af08d775b76bd516de9d149a7fc01118
Answer SHA-256: fdca601cc9ecf4479a4fb7da6f9ed9dc72e2bd6dbda8cb2db78c6d87953bd725