{
  "id": "C-Q2-R1-A4",
  "platform": "ChatGPT Search",
  "question_id": "Q2",
  "repeat": 1,
  "date_local": "2026-10-29",
  "started_at": "2026-10-29T15:55:04.406000+00:00",
  "finished_at": "2026-10-30T05:06:39.767426+00:00",
  "prompt": "What credible experiments have measured whether GEO tactics increase AI brand mentions or recommendations? Include sources when available.",
  "submitted_prompt": "\u7f51\u9875\u641c\u7d22 What credible experiments have measured whether GEO tactics increase AI brand mentions or recommendations? Include sources when available.",
  "observed_submitted_prompt": "\u7f51\u9875\u641c\u7d22 What credible experiments have measured whether GEO tactics increase AI brand mentions or recommendations? Include sources when available.",
  "normalized_submitted_prompt": "What credible experiments have measured whether GEO tactics increase AI brand mentions or recommendations? Include sources when available.",
  "submitted_prompt_provenance": "raw_record",
  "mode_label": "Web Search \u00b7 Medium",
  "model_label": "GPT-6 (UI selection label)",
  "model_evidence": "The original UI record showed GPT-6. The saved share HTML for the same conversation has gpt-6-thinking in both message model_slug and default_model_slug. These are platform identifiers, not independent proof of an exact backend build.",
  "search_status": "observed",
  "search_evidence": "The UI reported 6 websites searched. The saved conversation contains web.run tool messages, and a Web Search chip appears beside the original question.",
  "status": "completed",
  "answer_text": "# What credible experiments show about GEO effectiveness\n\nAs of October 2026, the evidence suggests that **GEO tactics can improve visibility in AI-generated answers, but the strength of evidence varies substantially by tactic and study design.** The clearest experimental evidence comes from academic research on content optimization. Large-scale industry studies support the importance of third-party brand mentions, while commercial case studies report substantial improvements after coordinated GEO campaigns.\n\nThe central distinction is whether a study demonstrates that a tactic *causes* more AI recommendations or merely shows that brands using that tactic tend to receive more mentions.\n\n## 1\\. The strongest evidence, ranked by study design\n\n\ue200image_group\ue202{\"query\":[\"Princeton University computer science research campus\"],\"aspect_ratio\":\"3:4\",\"layout\":\"carousel\",\"num_per_query\":1}\ue201\n\n(Strongest experimental evidence)\n\n**1\\. Princeton researchers \u2014 GEO: Generative Engine Optimization (KDD 2024)**\n\nPeer-reviewed academic paper \u00b7 Published August 2024\n\nThe researchers introduced GEO-bench and systematically tested content-optimization strategies across diverse queries and domains. They reported that optimization could improve content visibility in generative-engine responses by **up to 40%**\\. Tested approaches included adding relevant citations, statistics and quotations, and improving the presentation of information.\n\n**What it establishes\\:** Content-level changes can influence how generative engines represent and surface information under experimental conditions.\n\n**Limitations\\:** The headline is a maximum result, not an average guaranteed lift. The outcome is content visibility, which does not necessarily mean more brand recommendations, purchases or conversions.\n\nSource: \ue200url\ue202KDD 2024 paper and DOI\ue202https://doi.org/10.1145/3637528.3671900\ue201 \u00b7 \ue200url\ue202Princeton research record\ue202https://collaborate.princeton.edu/en/publications/geo-generative-engine-optimization/\ue201\\.\n\n---\n\n\ue200image_group\ue202{\"query\":[\"Semrush company logo on white background\"],\"aspect_ratio\":\"3:4\",\"layout\":\"carousel\",\"num_per_query\":1}\ue201\n\n(Real-world intervention)\n\n**2\\. Semrush \u2014 Its own AI visibility improvement program**\n\nCompany-reported case study \u00b7 May 2026 retrospective\n\nSemrush reported that its AI share of voice for target prompts increased from **13% to 32% in one month** during summer 2025, nearly tripling its share of answers mentioning the company relative to competitors. It subsequently refined its measurement and optimization process.\n\n**What it establishes\\:** A company can measure and materially improve its own visibility over a defined campaign period.\n\n**Limitations\\:** This is a first-party case study, not an independently replicated randomized experiment. The reported change does not isolate the causal contribution of individual tactics or rule out changes in competitors, prompts or AI systems.\n\nSource: \ue200url\ue202Semrush: How we're driving AI visibility\ue202https://www.semrush.com/blog/how-we-are-using-semrush-to-drive-llm-visibility/\ue201\\.\n\n---\n\n\ue200image_group\ue202{\"query\":[\"Ahrefs SEO software company logo\"],\"aspect_ratio\":\"3:4\",\"layout\":\"carousel\",\"num_per_query\":1}\ue201\n\n(Large-scale observational evidence)\n\n**3\\. Ahrefs \u2014 Brand visibility factors across 75,000 brands**\n\nIndustry data analysis \u00b7 2025\n\nAhrefs analyzed millions of AI-generated responses and found that brand mentions across the web correlated strongly with mentions in Google's AI Overviews. Its May 2025 study reported a Spearman correlation of **0.664** for web mentions, compared with **0.218** for backlinks and **0.392** for branded search volume.\n\nA later study extended the analysis to ChatGPT, AI Mode and AI Overviews, again finding strong associations between third-party brand mentions and AI visibility.\n\n**What it establishes\\:** Off-site brand presence is a promising GEO signal worth testing.\n\n**Limitations\\:** Correlation cannot establish whether additional mentions cause AI visibility to rise. Larger, more established brands may naturally have both more mentions and greater AI visibility.\n\nSources: \ue200url\ue2022025 study\ue202https://ahrefs.com/blog/ai-overview-brand-correlation/\ue201 \u00b7 \ue200url\ue202Cross-platform follow-up study\ue202https://ahrefs.com/blog/ai-brand-visibility-correlations/\ue201\\.\n\n## 2\\. What does the evidence say about individual GEO tactics?\n\nThe studies above support different conclusions. The table below separates tactics with experimental support from those with mainly observational or commercial evidence.\n\n| GEO tactic | Evidence | Current conclusion |\n|--------|--------|--------|\n| Add statistics, citations and expert quotations | Princeton controlled evaluation | Promising content-level intervention; test in your own niche |\n| Earn third-party brand mentions | Large-scale Ahrefs correlation analysis | Strong association; causal effect remains uncertain |\n| Get included in \u201cbest of\u201d lists and comparison articles | Commercial case studies and AI citation analyses | Promising for product recommendations; limited controlled evidence |\n| Publish more pages | Ahrefs observational analysis | Weak evidence that page volume alone increases AI visibility |\n| Build backlinks | Ahrefs observational analysis | Weaker association than brand mentions in its 2025 AI Overview study |\n| Publish social content | Comparative AI search research | Effects vary by platform; social content may be less influential than earned media |\n\nTwo additional sources help explain these findings.\n\n**4\\. Chen et al. \u2014 *Generative Engine Optimization: How to Dominate AI Search* \\(2025)**\n\nThis arXiv preprint uses large-scale controlled experiments to compare AI search with traditional search across verticals, languages and query variations. It finds that AI search systems favor earned media\u2014third-party authoritative sources\u2014over brand-owned and social content, with substantial differences between engines. \ue200url\ue202Read the paper\ue202https://arxiv.org/abs/2509.08919\ue201\\.\n\nThe distinction matters: this research provides evidence about *which sources AI systems tend to use*, but it does not establish that purchasing or securing a particular media placement will cause a brand to receive more recommendations.\n\n**5\\. Smoking Gun PR \u2014 Four-month GEO campaign**\n\nA home-appliance brand's campaign combined editorial coverage, expert commentary, product comparison articles and community placements. The agency reported\\:\n\nAI share of voice\n\n# 3.2% \u2192 9.1%\n\nBrand mentions\n\n# 274 \u2192 1,424\n\nReported results over four months; the campaign combined several tactics.\n\nSource: \ue200url\ue202Smoking Gun PR case study\ue202https://smokinggun.agency/our-work/geo-casestudy/\ue201\\.\n\nThis is useful real-world evidence, but it is agency-reported, lacks a published control group and does not isolate which intervention drove the results.\n\n## 3\\. How to interpret the evidence correctly\n\nThree measurement distinctions are essential when evaluating GEO claims.\n\n**Brand mention rate**\n\nHow frequently an AI answer names the brand. This is the most direct measure of whether a brand appears at all.\n\n**Recommendation and position**\n\nWhether the AI actually recommends the brand, places it among the top options, or merely mentions it in passing. A citation alone does not establish a positive recommendation.\n\n**Business impact**\n\nWhether the intervention increases qualified visits, leads or purchases. Greater AI visibility does not automatically translate into commercial results.\n\nAI answers can also change across repeated runs of the same prompt. Semrush explicitly describes this non-determinism in its measurement discussion. Consequently, before-and-after comparisons based on only a handful of prompts can produce misleading conclusions. \ue200url\ue202Source and methodology discussion\ue202https://www.semrush.com/blog/how-we-are-using-semrush-to-drive-llm-visibility/\ue201\\.\n\n## 4\\. What would constitute convincing evidence?\n\nFor someone conducting independent GEO experiments, I would prioritize the following design.\n\n1. **Establish a baseline.** Select 30\u2013100 realistic prompts in one topic area and record brand mentions, recommendation rate, ranking within recommendations and citations.\n2. **Change one variable at a time.** For example, test a structured, evidence-rich article against a comparable article without the optimization. Alternatively, compare prompts associated with earned-media placements against a matched control group.\n3. **Repeat across multiple runs.** Use the same prompts, models, settings and observation schedule. Record changes in the AI system and its retrieval sources.\n4. **Measure the treatment effect.** Compare changes in the treated group with changes in the control group, ideally using a difference-in-differences design. For content experiments, account for differences in topic, baseline authority and publication timing.\n5. **Publish the results, including failures.** Share the prompt set, intervention, sample size, observation period, outcome definitions and uncertainty. This makes the findings independently testable.\n\n## Bottom line\n\nThe strongest conclusion currently supported by the evidence is that **GEO is measurable and content optimization can improve AI visibility, but reliable causal evidence for many popular tactics remains limited**\\.\n\nFor practical experimentation, I would start with two hypotheses\\:\n\n1. **Content-level optimization\\:** Does adding concrete statistics, attributable citations and useful comparisons increase the probability that an AI answer surfaces your content or mentions your brand?\n2. **Earned-media optimization\\:** Does getting your brand included in relevant third-party comparisons increase its recommendation rate relative to a matched control?\n\nThese hypotheses connect the academic evidence to real-world brand outcomes, and both can be tested without relying on vendor claims or assuming that correlation proves causation.\n\n---\n\nIf you want, I can:\n\n- Suggest a detailed plan for running a GEO experiment to test brand mention increase\n- Provide example prompts and control set for a GEO brand mention experiment\n- Explain how to distinguish AI mentions vs positive brand recommendations in GEO testing",
  "notes": "Historical note (Cursor had not yet coded this record; coding is now complete): 2026-10-29 web supplement using GPT-6 + Web Search. This was attempt 4 for C-Q2-R1; text was extracted from the share-page fallbackMarkdown. Coding was not yet complete. Replacement for C-Q2-R1 / C-Q2-R1-A2/A3.\ntext preserves the complete fallbackMarkdown verbatim, including URL and citation tokens. links contains only explicit URL tokens in that answer, not the broader retrieved-source list.\nThe existing HTML snapshot was created by Cursor on 2026-10-29; capturedAt is this offline recovery time.\nCitations and individual observations were reviewed on October 30; Cursor's original files and all recovery captures remain preserved locally.",
  "capture_method": "Offline extraction of exact fallbackMarkdown from saved share-page HTML; no network, new prompt, or new public share",
  "citation_mapping_complete": true,
  "attempt_number": 4,
  "replacement_for": "C-Q2-R1",
  "slot_id": "C-Q2-R1",
  "attempt_id": "C-Q2-R1-A4",
  "capture_id": "C-Q2-R1-A4-restored-20261030-01",
  "selected_eligible": true,
  "selection_reason": "earliest_complete_attempt",
  "coding_reviewed_for_capture": true,
  "sources": [
    {
      "url": "https://doi.org/10.1145/3637528.3671900",
      "title": "KDD 2024 paper and DOI",
      "domain": "doi.org",
      "kind": "answer_link"
    },
    {
      "url": "https://collaborate.princeton.edu/en/publications/geo-generative-engine-optimization/",
      "title": "Princeton research record",
      "domain": "collaborate.princeton.edu",
      "kind": "answer_link"
    },
    {
      "url": "https://www.semrush.com/blog/how-we-are-using-semrush-to-drive-llm-visibility/",
      "title": "Semrush: How we're driving AI visibility",
      "domain": "www.semrush.com",
      "kind": "answer_link"
    },
    {
      "url": "https://ahrefs.com/blog/ai-overview-brand-correlation/",
      "title": "2025 study",
      "domain": "ahrefs.com",
      "kind": "answer_link"
    },
    {
      "url": "https://ahrefs.com/blog/ai-brand-visibility-correlations/",
      "title": "Cross-platform follow-up study",
      "domain": "ahrefs.com",
      "kind": "answer_link"
    },
    {
      "url": "https://arxiv.org/abs/2509.08919",
      "title": "Read the paper",
      "domain": "arxiv.org",
      "kind": "answer_link"
    },
    {
      "url": "https://smokinggun.agency/our-work/geo-casestudy/",
      "title": "Smoking Gun PR case study",
      "domain": "smokinggun.agency",
      "kind": "answer_link"
    }
  ],
  "retrieval_sources": [],
  "observations": [
    {
      "entity": "GEO research (Aggarwal et al.)",
      "entity_type": "research",
      "citation": true,
      "mention": true,
      "recommendation": true,
      "evidence": "Princeton researchers \u2014 GEO: Generative Engine Optimization (KDD 2024)",
      "notes": "Selected as the strongest experimental evidence. The original paper's full title appears explicitly, and the same DOI links it to the existing GEO research entity. Aggarwal is not named in this answer; M=true is based on the study title. Both the DOI and Princeton record are explicit URL tokens in the body and are retained as observed.",
      "source_urls": [
        "https://doi.org/10.1145/3637528.3671900",
        "https://collaborate.princeton.edu/en/publications/geo-generative-engine-optimization/"
      ]
    },
    {
      "entity": "Semrush",
      "entity_type": "brand",
      "citation": true,
      "mention": true,
      "recommendation": true,
      "evidence": "Semrush \u2014 Its own AI visibility improvement program",
      "notes": "Selected as a separately numbered real-world intervention case. R=true means the study material is recommended for evaluation, not that its commercial claims have been verified. The answer directly cites Semrush's own retrospective article. The same URL reappears in the methodology discussion but is retained only once.",
      "source_urls": [
        "https://www.semrush.com/blog/how-we-are-using-semrush-to-drive-llm-visibility/"
      ]
    },
    {
      "entity": "Ahrefs",
      "entity_type": "brand",
      "citation": true,
      "mention": true,
      "recommendation": true,
      "evidence": "Ahrefs \u2014 Brand visibility factors across 75,000 brands",
      "notes": "Selected as a separately numbered large-scale observational study. Both Ahrefs research URLs are explicitly present in the body. The answer preserves the limitation that correlation does not establish causation; R=true denotes a reading/evidence recommendation.",
      "source_urls": [
        "https://ahrefs.com/blog/ai-overview-brand-correlation/",
        "https://ahrefs.com/blog/ai-brand-visibility-correlations/"
      ]
    },
    {
      "entity": "Chen et al. (How to Dominate AI Search)",
      "entity_type": "research",
      "citation": true,
      "mention": true,
      "recommendation": true,
      "evidence": "Generative Engine Optimization: How to Dominate AI Search",
      "notes": "The answer explicitly names Chen et al. and the full study title, with a Read the paper link to the original arXiv work. It is a separate study entry used to explain the evidence; M/C/R are all true.",
      "source_urls": [
        "https://arxiv.org/abs/2509.08919"
      ]
    },
    {
      "entity": "Smoking Gun PR",
      "entity_type": "organization",
      "citation": true,
      "mention": true,
      "recommendation": true,
      "evidence": "Smoking Gun PR \u2014 Four-month GEO campaign",
      "notes": "A separately numbered case described as useful real-world evidence, supporting R=true. The answer also explicitly says it is agency-reported, has no public control group, and cannot isolate an individual intervention. The agency's own case page is cited; the unnamed appliance brand is not inferred.",
      "source_urls": [
        "https://smokinggun.agency/our-work/geo-casestudy/"
      ]
    }
  ],
  "raw_file": "data/records/C-Q2-R1-A4.json",
  "capture_history": [
    {
      "capture_id": "C-Q2-R1-A4-legacy-01",
      "attempt_id": "C-Q2-R1-A4",
      "slot_id": "C-Q2-R1",
      "method": "Browser UI + share page HTML extract",
      "captured_at": "2026-10-29T14:05:30.000Z",
      "status": "completed",
      "body_completeness": "legacy_unverified",
      "answer_chars": 8999,
      "source_count": 7,
      "notes": "2026-10-29 web supplement using GPT-6 + Web Search. This was attempt 4 for C-Q2-R1; text was extracted from the share-page fallbackMarkdown. Coding was not yet complete. Replacement for C-Q2-R1 / C-Q2-R1-A2/A3.",
      "capture_kind": "initial_export",
      "selected": false,
      "file": "data/captures/C-Q2-R1-A4-legacy-01.json",
      "public_evidence_scope": "Public capture summary; original screenshots, full HTML and private conversation information remain local."
    },
    {
      "capture_id": "C-Q2-R1-A4-restored-20261030-01",
      "attempt_id": "C-Q2-R1-A4",
      "slot_id": "C-Q2-R1",
      "method": "Offline extraction of exact fallbackMarkdown from saved share-page HTML; no network, new prompt, or new public share",
      "captured_at": "2026-10-30T05:06:39.767426+00:00",
      "status": "completed",
      "body_completeness": "completed",
      "answer_chars": 10147,
      "source_count": 7,
      "notes": "Historical note (Cursor had not yet coded this record; coding is now complete): 2026-10-29 web supplement using GPT-6 + Web Search. This was attempt 4 for C-Q2-R1; text was extracted from the share-page fallbackMarkdown. Coding was not yet complete. Replacement for C-Q2-R1 / C-Q2-R1-A2/A3.\ntext preserves the complete fallbackMarkdown verbatim, including URL and citation tokens. links contains only explicit URL tokens in that answer, not the broader retrieved-source list.\nThe existing HTML snapshot was created by Cursor on 2026-10-29; capturedAt is this offline recovery time.\nCitations and individual observations were reviewed on October 30; Cursor's original files and all recovery captures remain preserved locally.",
      "capture_kind": "share_preview",
      "selected": true,
      "file": "data/captures/C-Q2-R1-A4-restored-20261030-01.json",
      "public_evidence_scope": "Public capture summary; original screenshots, full HTML and private conversation information remain local."
    },
    {
      "capture_id": "C-Q2-R1-A4-share-preview-metadata",
      "attempt_id": "C-Q2-R1-A4",
      "slot_id": "C-Q2-R1",
      "method": "Saved handoff summary; not a body capture",
      "captured_at": null,
      "status": "metadata_only",
      "body_completeness": "not_applicable",
      "answer_chars": null,
      "source_count": null,
      "notes": "This file contains summary metadata only. Missing body fields do not establish an empty or failed answer. The share dialog indicated that the answer was complete.\nThe body was extracted from fallbackMarkdown in the share-page HTML.\nUI: GPT-6 \u00b7 Medium \u00b7 Web Search.\nC/M/R coding had not yet been performed.",
      "capture_kind": "metadata_summary",
      "selected": false,
      "file": "data/captures/C-Q2-R1-A4-share-preview-metadata.json",
      "public_evidence_scope": "Public capture summary; original screenshots, full HTML and private conversation information remain local."
    }
  ],
  "capture_provenance": {
    "original_status": "completed",
    "selected_capture_id": "C-Q2-R1-A4-restored-20261030-01",
    "method": "Offline extraction of exact fallbackMarkdown from saved share-page HTML; no network, new prompt, or new public share",
    "captured_at": "2026-10-30T05:06:39.767426+00:00",
    "same_conversation": true,
    "review_notes": "Historical note (Cursor had not yet coded this record; coding is now complete): 2026-10-29 web supplement using GPT-6 + Web Search. This was attempt 4 for C-Q2-R1; text was extracted from the share-page fallbackMarkdown. Coding was not yet complete. Replacement for C-Q2-R1 / C-Q2-R1-A2/A3.\ntext preserves the complete fallbackMarkdown verbatim, including URL and citation tokens. links contains only explicit URL tokens in that answer, not the broader retrieved-source list.\nThe existing HTML snapshot was created by Cursor on 2026-10-29; capturedAt is this offline recovery time.\nCitations and individual observations were reviewed on October 30; Cursor's original files and all recovery captures remain preserved locally.",
    "selected_capture_file": "data/records/C-Q2-R1-A4.json"
  },
  "public_copy": true,
  "metadata_corrections": [
    {
      "field": "model_evidence",
      "original_value": "\u6a21\u578b\u83dc\u5355\u6cbf\u7528 GPT-6\uff1b\u601d\u8003\u5f3a\u5ea6 Medium\uff08\u964d\u9891\uff09",
      "previous_effective_value": "\u6a21\u578b\u83dc\u5355\u6cbf\u7528 GPT-6\uff1b\u601d\u8003\u5f3a\u5ea6 Medium\uff08\u964d\u9891\uff09",
      "value": "The original UI record showed GPT-6. The saved share HTML for the same conversation has gpt-6-thinking in both message model_slug and default_model_slug. These are platform identifiers, not independent proof of an exact backend build.",
      "evidence": "Offline decoded conversation/message metadata from saved HTML; unrelated feature-flag model names are excluded."
    },
    {
      "field": "search_evidence",
      "original_value": "UI \u542f\u7528\u7f51\u9875\u641c\u7d22 chip\uff1b\u5206\u4eab\u5bf9\u8bdd\u6846\u663e\u793a\u201c\u56de\u7b54\u5df2\u5b8c\u6210\u201d\uff1b\u6b63\u6587/\u5206\u4eab\u9875\u542b\u591a\u4e2a utm_source=chatgpt.com \u94fe\u63a5",
      "previous_effective_value": "UI \u542f\u7528\u7f51\u9875\u641c\u7d22 chip\uff1b\u5206\u4eab\u5bf9\u8bdd\u6846\u663e\u793a\u201c\u56de\u7b54\u5df2\u5b8c\u6210\u201d\uff1b\u6b63\u6587/\u5206\u4eab\u9875\u542b\u591a\u4e2a utm_source=chatgpt.com \u94fe\u63a5",
      "value": "The UI reported 6 websites searched. The saved conversation contains web.run tool messages, and a Web Search chip appears beside the original question.",
      "evidence": "Decoded actual conversation search activity and user prompt, not inferred from tracking query parameters."
    },
    {
      "field": "started_at",
      "original_value": "2026-10-29T14:02:00.000Z",
      "previous_effective_value": "2026-10-29T14:02:00.000Z",
      "value": "2026-10-29T15:55:04.406000+00:00",
      "evidence": "Saved user-message create_time conflicts with raw recorded 14:02Z; shared HTML backing_conversation_id matches the raw conversation."
    },
    {
      "field": "finished_at",
      "original_value": "2026-10-29T14:05:30.000Z",
      "previous_effective_value": "2026-10-30T05:06:39.767426+00:00",
      "value": "2026-10-30T05:06:39.767426+00:00",
      "evidence": "Actual current offline recovery/extraction time. Original raw 14:05:30Z predates saved user-message creation; historical first extraction time cannot be verified. This value is not model generation completion time."
    }
  ]
}
