Generative Engine Optimization: How to Dominate AI Search

Kaiwen Chen; Mahe Chen; Nick Koudas; Xiaoxuan Wang

REVIEW 4 major objections 5 minor 2 cited by

Generative Engine Optimization: How to Dominate AI Search

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper tries to establish that generative AI search systems—ChatGPT, Perplexity, Gemini, Claude—share a systematic sourcing bias: they draw their answers overwhelmingly from earned media (third-party editorial and review sites) rather t

desk verdict Useful empirical snapshot of AI search sourcing, but the universal "earned media bias" claim rests on ranking-style prompts and no statistics. read the letter →

arxiv 2509.08919 v1 pith:5ZSFY6RB submitted 2025-09-10 cs.IR cs.AIcs.CLcs.LGcs.SI

Mahe Chen , Xiaoxuan Wang , Kaiwen Chen , Nick Koudas This is my paper · ORCID

classification cs.IRcs.AIcs.CLcs.LGcs.SI MSC 68P2068T50

keywords generativeengineoptimizationAIsearchSEOearnedmediabrandvisibilityinformationretrievallargelanguagemodelscitationanalysis

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper attempts to prove that AI search engines source information with a systematic and overwhelming bias toward earned media (third-party editorial and review sites) over brand-owned and social content, in stark contrast to Google's more balanced mix. It also claims that AI search services differ significantly among themselves in domain diversity, freshness, cross-language stability, and sensitivity to phrasing. If correct, the practical consequence is that winning visibility in AI search requires cultivating third-party citations rather than relying on owned content or social presence, and the strategy must be tailored to each engine and language.

What carries the argument

The empirical engine is a controlled comparative pipeline: ranking-style prompts issued to each engine, citation URLs normalized to registrable domains, domains classified by GPT-4o-search-preview into Brand / Earned / Social, and outputs compared with overlap metrics (Coverage@k, Jaccard index). This machinery lets the authors separate stable structural tendencies (earned-media dominance) from engine-specific and language-specific variation.

What would settle it

Run the same citation-classification pipeline on conversational and task-oriented queries from the paper's own taxonomy (coding help, self-improvement, creative writing) and check whether the earned-media share falls below the measured 60–90%; or repeat the August 2025 measurements in a later window and check whether the brand/social shares have converged toward Google's mix.

Watch

Extended reading notes

Core claim

The central discovery is that under standardized consumer ranking queries, web-enabled AI engines allocate most of their citations to earned media domains (typically 60–90% depending on engine and vertical), nearly exclude social sources (often 0–5%), and relegate brand-owned sites to a secondary role, while Google returns a more balanced mix of brand, earned, and social domains. The paper also documents that engines differ in how they handle language: Claude reuses English authority domains across languages, GPT swaps to local-language ecosystems, and Perplexity/Gemini sit in between; paraphrase shifts move citations less than language shifts; and local-service queries produce very low over

Load-bearing premise

The results rest on the assumption that standardized ranking-style prompts (e.g., 'Top 10... brands') in a handful of consumer verticals capture how people actually use AI search; if real queries are predominantly conversational, long-tail, or non-ranking, the measured dominance of earned media may be an artifact of the prompt format.

Editorial extensions

If this is right

Content strategists should treat earned media placements as the primary lever for AI search visibility; on-page SEO remains necessary but not sufficient.
Websites should be structured for machine scannability (schema markup, comparison tables, explicit value propositions) so AI agents can extract justification.
Multilingual visibility requires a dual path: local-language earned media for GPT/Perplexity, English-language authority for Claude.
Niche brands face an inherent big-brand bias; they need niche authority and Perplexity-style social/YouTube presence to break in.
Because engine behaviors are a moving target, conclusions need periodic re-measurement.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The results imply a shift in marketing budgets: if AI search favors earned media, brand-owned content's ROI may decline relative to PR and third-party coverage; brands that currently invest heavily in owned content may need to rebalance.
The near-zero social share in GPT answers suggests that community platforms (Reddit, forums) are losing a channel in AI-mediated discovery, which could alter traffic patterns for those platforms.
The fact that engines overlap little with each other means the same query can yield materially different brands/answers across assistants; this raises a testable question about verifiability: a user who relies on only one assistant may get systematically different recommendations.
The methodology (ranking-style prompts) suggests an untested boundary: the ground truth for 'earned bias' under conversational or long-tail queries may differ, so the next natural experiment is to run the pipeline on the taxonomy's non-ranking categories.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The paper delivers a genuinely useful comparative measurement of how four AI search engines and Google cite sources across languages and paraphrases. That's new relative to the prior GEO optimization work (Aggarwal et al.) and manipulation studies (Kumar/Lakkaraju). If you work on IR or GEO, this is worth reading.

What it does well: the experiments are extensive. 1,000 ranking prompts across verticals, 100 base queries translated into five languages, seven paraphrase templates, multiple engines. The consistent pattern—AI engines cite earned/editorial sources much more than social, while Google includes more social and brand—appears across many settings. The engine differences (Claude and GPT earned-heavy, Perplexity more socially inclusive, Gemini more brand-prone) are interesting and plausible. They also acknowledge temporal and classification limitations, which is honest.

Soft spots: the biggest is that nearly all experiments force queries into a "Top 10..." ranking template. That naturally selects for listicles and review roundups, so the near-zero social share and 70–90% earned mix may be partly an artifact. The paper's own Reddit-derived taxonomy shows real AI queries are dominated by coding, creative writing, self-improvement—none of which are tested. Section 4.3 gestures at non-ranking intents but provides no query inventory or sample sizes, so it can't rescue the general claim. The stress-test concern holds up. Second, all percentages are point estimates with no confidence intervals or statistical tests. Given the measurement noise, "overwhelming" is too strong without error bars. Third, using GPT-4o to classify domains and extract dates introduces same-model bias, though not fatally—the object is the engine's citations, not the classifier's behavior. Fourth, the intro says "no published studies" then cites Aggarwal et al. as prior GEO work—a self-contradiction. Minor but sloppy. No code or data released, which limits reproducibility.

Who this is for: practitioners in SEO/GEO, and researchers studying AI search source selection. The central qualitative finding likely holds up—AI search favors earned media and suppresses social relative to Google—but the exact percentages and the universal phrasing should be treated as provisional. The paper deserves peer review because it's a large, structured empirical study with falsifiable claims; a good referee can push for non-ranking queries, uncertainty quantification, and data release. Send it to review, expecting major revision.

Referee Report

4 major / 5 minor

Summary. The paper compares web-enabled AI search systems (ChatGPT/GPT-4o-search-preview, Perplexity, Gemini, Claude) with traditional Google search across consumer verticals, languages, and paraphrases. It measures domain overlap, media-type mix (Brand/Earned/Social), freshness, and cross-engine diversity, concluding that AI search exhibits a systematic and overwhelming bias toward earned media relative to Google, and that engines differ in diversity, freshness, cross-language stability, and paraphrase sensitivity. Based on these results, the paper proposes a strategic Generative Engine Optimization (GEO) agenda for brands and practitioners.

Significance. If the empirical claims hold, the paper provides timely, multi-engine evidence about a major shift in information access, with direct implications for SEO/GEO practice and for the study of AI-mediated discovery. The strengths are the breadth of the comparison (four AI engines plus Google, multiple verticals and languages), the use of direct API outputs rather than simulated relevance judgments, and the explicit acknowledgement of temporal and black-box limitations in Section 7. However, the central empirical claim is currently supported by a restricted query protocol and by measurement instruments that share the same model family as the objects of study; these issues need to be addressed before the broader conclusions can be accepted.

major comments (4)

[§4.1.1, §4.2.1, §4.3] The abstract's universal claim that 'AI Search exhibit a systematic and overwhelming bias towards Earned media' is established almost exclusively through ranking-style templates. §4.1.1 states that prompts are standardized to 'Top 10 ...' and §4.2.1 uses 1,000 such prompts; §4.2.3–4.2.4 and §5.2.x similarly use ranking or identification prompts. §4.3 is the only section with non-ranking intents, but it gives no query inventory, sample size, or engine coverage, and Figures 18–20 appear without methodology. Ranking prompts naturally elicit listicles and review roundups, so the measured earned-media dominance and near-zero social share may be a template artifact. The paper's own §3 taxonomy includes coding, creative writing, and self-improvement queries that are not 'Top 10' lists. Please either restrict the central claim to ranking-style prompts or add a comparable non-ranking experiment u
[§4.1.4, §5.1] The Brand/Earned/Social classification is performed by GPT-4o-search-preview, and §5.1 further describes 'GPT-assisted classification prompts' for the brand experiments. The classifier is from the same model family as several systems under study (GPT itself, and to a lesser extent the other LLM-based engines), so the measured category shares could reflect the classifier's own sourcing biases rather than the engines' behavior. No validation against human labels, inter-annotator agreement, or an independent classifier is reported. Please provide a human-labeled validation set, report classification accuracy, and show that the main distributions are robust to using a rule-based or external classifier.
[§4.2.1, §5.2.2] All reported percentages and overlap values are point estimates. For example, Section 4.2.1 gives figures such as 69.1% Earned and 0% Social for AI search in Canada, and Section 5.2.2 reports Claude and GPT at approximately 93.7% and 93.6% Earned, but no confidence intervals, hypothesis tests, or query-level variance are provided. The claims of 'systematic' and 'overwhelming' differences require statistical support, especially because several comparisons are based on 10 verticals and 100 queries, where sampling noise is non-negligible. Report per-query distributions, error bars, and formal tests (e.g., paired comparisons of the Earned share between AI and Google).
[§5.2.6] The 'Big Brand Bias' experiment depends on curated sets of 20 major and 20 niche brands (Table 1). The selection of these specific brands is not justified, and the conclusion that Coca-Cola and Pepsi dominate may be a property of the chosen lists rather than a general property of the systems. Please describe how the brand sets were constructed, or show that the result is stable under alternative brand lists.

minor comments (5)

[§4.2.4, §5.2.5] Cross-reference errors: §4.2.4's interpretation refers to 'language choice (§4.2.2)', but the language experiment is §4.2.3; §5.2.5's objective refers to 'prior experiments (§5.3)', which should be §5.2.4 or §5.2.2.
[References] Reference [10] is labeled 'WIRED' but points to a Search Engine Land URL; the inline citation to 'reddit []' has an empty bracket. Please correct these.
[Acknowledgments] The acknowledgment to ktau.ai, a commercial GEO company, should be accompanied by a conflict-of-interest statement, given the paper's prescriptive GEO agenda.
[§4.3] The Informational/Consideration/Transactional taxonomy in §4.3 is not defined or linked to the Reddit-derived taxonomy in §3. It would be helpful to state how the two taxonomies relate and to provide example queries for each intent bucket.
[Reproducibility] The paper states reproducibility 'in mind' (e.g., §5.2.4) but does not release query lists, raw outputs, or code. Given the API-based methodology, making these available as an artifact would substantially strengthen the paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical measurement study; its central claims do not reduce to fitted parameters, self-citations, or definitional equivalences.

full rationale

The paper's derivation chain is a direct API-measurement pipeline: it issues queries to Google and AI engines, extracts cited domains, classifies them into Brand/Earned/Social, and aggregates shares. No parameter is fitted and then renamed a prediction; no uniqueness theorem is imported from the authors' prior work; and the only prior-work citation (Aggarwal et al. for the GEO term) is not load-bearing and is not a self-citation. The central 'earned media bias' claim is an empirical distribution, not an identity: using ranking-style prompts may affect external validity, but the observed dominance of earned domains is not true by construction. The use of GPT-4o-search-preview as the domain classifier is a measurement-validity concern, since the instrument overlaps with one studied engine; the paper itself discloses the subjectivity of its classification in §7 (Assumptions and Limitations). This is a methodological limitation, not a circular reduction. Similarly, the Reddit-derived taxonomy in §3 covers a broader query space than the standardized prompts, which is a generalizability limitation rather than circularity. The cola 'big brand bias' experiment also uses popularity-framed queries, but this is an experimental-design confound, not a derivation circle. Overall, no circular step meeting the evidence bar is present.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central measurements rest on several unvalidated modeling choices: the top-k truncation, the ad hoc freshness scorer, the hand-built major/niche brand lists, and the three-way source taxonomy. No fitting is performed, so the circularity burden from parameters is low, but the precision of every reported percentage depends on these choices.

free parameters (3)

top-k cutoff = 10 (and 5 in some analyses)
The results (domain overlap, source mix) are computed for the top-10 citations/results; changing k changes the distributions. Chosen by the authors, not data-driven.
freshness scoring formula = freshness = mean(1/(1+age)); coverage-adjusted = freshness * coverage
This weighting is ad hoc and affects comparative freshness scores between engines. No justification, and coverage differences (e.g., Claude automotive 61%) directly change the adjusted score.
curated major/niche brand sets = 20 major and 20 niche cola brands listed in Table 1
The "big brand bias" results depend on which brands the authors placed in each bucket. The split is hand-chosen and could be shifted to change the measured bias.

assumptions (5)

domain assumption Citations returned by web-enabled AI engines are an accurate proxy for the sources used to generate the answer and for visibility in AI search.
The entire GEO strategy is premised on citation presence being the optimization target. The paper does not test whether uncited sources influenced the answer.
domain assumption GPT-4o-based classification of domains into Brand/Earned/Social is sufficiently accurate to support the reported percentages.
All distributions rely on labels from GPT-4o-search-preview with a predefined social-domain list; errors would propagate to every claim about source mix. See Sec. 4.1.4.
domain assumption Ranking-style prompts are representative of real user interactions with AI search.
Prompts are standardized to "Top 10..." to allow scoring, but no evidence is provided that this is a common query type; the Reddit taxonomy in Sec. 3 suggests broader usage.
domain assumption The outputs of these AI services are stable enough within the data-collection window that pooling queries and citing domains is meaningful.
Responses are stochastic and APIs may serve different results per call; no repeated calls or variance estimation are reported.
ad hoc to paper The three-category Brand/Earned/Social taxonomy is a valid projection of the information ecosystem.
The authors acknowledge the categories are constructed and that other taxonomies would change numbers (Sec. 7). The central "earned bias" claim depends on this taxonomy.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Generative Engine Optimization: How to Dominate AI Search." pith.science (2026). https://pith.science/paper/5ZSFY6RB

@misc{pith2026250908919,
  author       = {Pith},
  title        = {Pith review of: Generative Engine Optimization: How to Dominate AI Search},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5ZSFY6RB}},
  note         = {Machine review of arXiv:2509.08919}
}

read the original abstract

The rapid adoption of generative AI-powered search engines like ChatGPT, Perplexity, and Gemini is fundamentally reshaping information retrieval, moving from traditional ranked lists to synthesized, citation-backed answers. This shift challenges established Search Engine Optimization (SEO) practices and necessitates a new paradigm, which we term Generative Engine Optimization (GEO). This paper presents a comprehensive comparative analysis of AI Search and traditional web search (Google). Through a series of large-scale, controlled experiments across multiple verticals, languages, and query paraphrases, we quantify critical differences in how these systems source information. Our key findings reveal that AI Search exhibit a systematic and overwhelming bias towards Earned media (third-party, authoritative sources) over Brand-owned and Social content, a stark contrast to Google's more balanced mix. We further demonstrate that AI Search services differ significantly from each other in their domain diversity, freshness, cross-language stability, and sensitivity to phrasing. Based on these empirical results, we formulate a strategic GEO agenda. We provide actionable guidance for practitioners, emphasizing the critical need to: (1) engineer content for machine scannability and justification, (2) dominate earned media to build AI-perceived authority, (3) adopt engine-specific and language-aware strategies, and (4) overcome the inherent "big brand bias" for niche players. Our work provides the foundational empirical analysis and a strategic framework for achieving visibility in the new generative search landscape.

Figures

Figures reproduced from arXiv: 2509.08919 by the authors.

**Figure 2.** Automotive: Distribution of source types (Brand, [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 5.** Local Search Coverage by Business Category. Cov [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

Figures from the paper (39 more)

**Figure 7.** Figure 7: Language Sensitivity: Domain overlap heatmap, [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

**Figure 8.** Figure 8: Language Sensitivity: Domain overlap heatmap, [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]

**Figure 6.** Figure 6: Language Sensitivity: Domain overlap heatmap, [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

**Figure 10.** Figure 10: Language Sensitivity: Domain overlap heatmap, [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]

**Figure 12.** Figure 12: Language Sensitivity: Overall website-language [PITH_FULL_IMAGE:figures/full_fig_p010_12.png]

**Figure 11.** Figure 11: Language Sensitivity: Overall domain-type distri [PITH_FULL_IMAGE:figures/full_fig_p010_11.png]

**Figure 14.** Figure 14: Paraphrase Sensitivity: Domain overlap heatmap, [PITH_FULL_IMAGE:figures/full_fig_p011_14.png]

**Figure 15.** Figure 15: Paraphrase Sensitivity: Domain overlap heatmap, [PITH_FULL_IMAGE:figures/full_fig_p011_15.png]

**Figure 16.** Figure 16: Paraphrase Sensitivity: Domain overlap heatmap, [PITH_FULL_IMAGE:figures/full_fig_p011_16.png]

**Figure 18.** Figure 18: Media-type distribution (Brand, Earned, Social) [PITH_FULL_IMAGE:figures/full_fig_p012_18.png]

**Figure 17.** Figure 17: Paraphrase Sensitivity: Overall domain-type dis [PITH_FULL_IMAGE:figures/full_fig_p012_17.png]

**Figure 19.** Figure 19: Media-type distribution (Brand, Earned, Social) [PITH_FULL_IMAGE:figures/full_fig_p012_19.png]

**Figure 20.** Figure 20: Media-type distribution (Brand, Earned, Social) [PITH_FULL_IMAGE:figures/full_fig_p013_20.png]

**Figure 21.** Figure 21: Comparison of Claude vs. Gemini for well-known [PITH_FULL_IMAGE:figures/full_fig_p014_21.png]

**Figure 22.** Figure 22: Comparison of GPT vs. Gemini for well-known [PITH_FULL_IMAGE:figures/full_fig_p015_22.png]

**Figure 23.** Figure 23: Comparison of Perplexity vs. Gemini for well [PITH_FULL_IMAGE:figures/full_fig_p015_23.png]

**Figure 24.** Figure 24: Distribution of article age (days) across verticals [PITH_FULL_IMAGE:figures/full_fig_p016_24.png]

**Figure 25.** Figure 25: Domain classification shares (Brand, Earned, So [PITH_FULL_IMAGE:figures/full_fig_p016_25.png]

**Figure 26.** Figure 26: Domain composition (Brand, Earned, Social) across [PITH_FULL_IMAGE:figures/full_fig_p017_26.png]

**Figure 28.** Figure 28: Domain diversity for automotive across Claude, [PITH_FULL_IMAGE:figures/full_fig_p018_28.png]

**Figure 27.** Figure 27: Domain diversity for consumer electronics across [PITH_FULL_IMAGE:figures/full_fig_p018_27.png]

**Figure 31.** Figure 31: Domain diversity for Moving Company queries [PITH_FULL_IMAGE:figures/full_fig_p019_31.png]

**Figure 29.** Figure 29: Domain diversity for Auto Repair queries across [PITH_FULL_IMAGE:figures/full_fig_p019_29.png]

**Figure 30.** Figure 30: Domain diversity for Dentist queries across Claude, [PITH_FULL_IMAGE:figures/full_fig_p019_30.png]

**Figure 33.** Figure 33: Big Brand Bias: Brand counts (ChatGPT) [PITH_FULL_IMAGE:figures/full_fig_p020_33.png]

**Figure 34.** Figure 34: Big Brand Bias: Brand counts (Perplexity) [PITH_FULL_IMAGE:figures/full_fig_p020_34.png]

**Figure 32.** Figure 32: Big Brand Bias: Individual brand counts (com [PITH_FULL_IMAGE:figures/full_fig_p020_32.png]

**Figure 35.** Figure 35: Big Brand Bias: Category summary (combined) [PITH_FULL_IMAGE:figures/full_fig_p021_35.png]

**Figure 38.** Figure 38: Big Brand Bias: Top consulted domains (ChatGPT) [PITH_FULL_IMAGE:figures/full_fig_p021_38.png]

**Figure 39.** Figure 39: Big Brand Bias: Top consulted domains (Perplexity) [PITH_FULL_IMAGE:figures/full_fig_p021_39.png]

**Figure 40.** Figure 40: Bank Queries Across Personas: Overall domain [PITH_FULL_IMAGE:figures/full_fig_p022_40.png]

**Figure 42.** Figure 42: Bank Queries Across Personas: Overall top 10 do [PITH_FULL_IMAGE:figures/full_fig_p022_42.png]

**Figure 43.** Figure 43: Language Sensitivity: Brand overlap heatmap, Per [PITH_FULL_IMAGE:figures/full_fig_p023_43.png]

**Figure 44.** Figure 44: Language Sensitivity: Brand overlap heatmap, GPT [PITH_FULL_IMAGE:figures/full_fig_p023_44.png]

**Figure 45.** Figure 45: Language Sensitivity: Brand overlap heatmap, [PITH_FULL_IMAGE:figures/full_fig_p023_45.png]

**Figure 46.** Figure 46: Language Sensitivity: Brand overlap heatmap, [PITH_FULL_IMAGE:figures/full_fig_p024_46.png]

**Figure 47.** Figure 47: Paraphrase Sensitivity: Brand overlap heatmap, [PITH_FULL_IMAGE:figures/full_fig_p024_47.png]

**Figure 48.** Figure 48: Paraphrase Sensitivity: Brand overlap heatmap, [PITH_FULL_IMAGE:figures/full_fig_p025_48.png]

**Figure 49.** Figure 49: Paraphrase Sensitivity: Brand overlap heatmap, [PITH_FULL_IMAGE:figures/full_fig_p025_49.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026)
cs.IR 2026-07 conditional novelty 5.0 of 10

A critical review of GEO research concludes that already-retrieved content can improve citation and use, but no tested technique reliably raises organic discoverability or downstream traffic across engines.
AI Answer Engine Citation Behavior An Empirical Analysis of the GEO16 Framework
cs.AI 2025-09 conditional novelty 5.0 of 10

A new audit framework, GEO-16, links 16 on-page quality signals to AI answer engine citations, identifying thresholds associated with higher citation odds.

Reference graph

Works this paper leans on

11 extracted references · 3 linked inside Pith · cited by 2 Pith papers

[1]

Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan, and Ameet Deshpande. 2024. GEO: Generative Engine Optimization. arXiv:2311.09735 [cs.LG] https://arxiv.org/abs/2311.09735

arXiv 2024
[2]

2025.CEO says Perplexity hit 780M queries in May

Perplexity AI. 2025.CEO says Perplexity hit 780M queries in May

2025
[3]

2025.About a third of U.S

Pew Research Center. 2025.About a third of U.S. adults have used ChatGPT; usage has increased over the past year. https://www.pewresearch.org/short-reads/2025/ 06/25/34-of-us-adults-have-used-chatgpt-about-double-the-share-in-2023/ [Online]. Available

2025
[4]

2025.Google users are less likely to click on links when an AI summary appears in the results

Pew Research Center. 2025.Google users are less likely to click on links when an AI summary appears in the results. https://www.pewresearch.org/short- reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai- summary-appears-in-the-results/ [Online]. Available

2025
[5]

Aounon Kumar and Himabindu Lakkaraju. 2024. Manipulating Large Language Models to Increase Product Visibility. arXiv:2404.07981 [cs.IR] https://arxiv.org/ abs/2404.07981

arXiv 2024
[6]

2024.ChatGPT Surges to 3.1B Visits in September 2024

Similarweb. 2024.ChatGPT Surges to 3.1B Visits in September 2024. https://www.similarweb.com/blog/insights/ai-news/chatgpt-topped-3- billion-visits-in-september/ [Online]. Available

2024
[7]

2025.AI Chatbot Market Share Worldwide

StatCounter Global Stats. 2025.AI Chatbot Market Share Worldwide. https: //gs.statcounter.com/ai-chatbot-market-share [Online]. Available

2025
[8]

2025.Search Engine Market Share Worldwide

StatCounter Global Stats. 2025.Search Engine Market Share Worldwide. https: //gs.statcounter.com/search-engine-market-share [Online]. Available

2025

Show all 11 references

[9]

Alexander Wan, Eric Wallace, and Dan Klein. 2024. What Evidence Do Language Models Find Convincing? arXiv:2402.11782 [cs.CL] https://arxiv.org/abs/2402. 11782

2024 arXiv
[10]

2024.Google Cut Back AI Overviews in Search Even Before Its ’Pizza Glue’ Fiasco

WIRED. 2024.Google Cut Back AI Overviews in Search Even Before Its ’Pizza Glue’ Fiasco. https://searchengineland.com/google-ai-overviews-visibility-drops-15- percent-queries-442850?utm_source=chatgpt.com [Online]. Available

2024
[2025]

Available

https://www.perplexity.ai/page/ceo-says-perplexity-hit-780m-q- dENgiYOuTfaMEpxLQc2bIQ [Online]. Available

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan, and Ameet Deshpande. 2024. GEO: Generative Engine Optimization. arXiv:2311.09735 [cs.LG] https://arxiv.org/abs/2311.09735

arXiv 2024

[2] [2]

2025.CEO says Perplexity hit 780M queries in May

Perplexity AI. 2025.CEO says Perplexity hit 780M queries in May

2025

[3] [3]

2025.About a third of U.S

Pew Research Center. 2025.About a third of U.S. adults have used ChatGPT; usage has increased over the past year. https://www.pewresearch.org/short-reads/2025/ 06/25/34-of-us-adults-have-used-chatgpt-about-double-the-share-in-2023/ [Online]. Available

2025

[4] [4]

2025.Google users are less likely to click on links when an AI summary appears in the results

Pew Research Center. 2025.Google users are less likely to click on links when an AI summary appears in the results. https://www.pewresearch.org/short- reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai- summary-appears-in-the-results/ [Online]. Available

2025

[5] [5]

Aounon Kumar and Himabindu Lakkaraju. 2024. Manipulating Large Language Models to Increase Product Visibility. arXiv:2404.07981 [cs.IR] https://arxiv.org/ abs/2404.07981

arXiv 2024

[6] [6]

2024.ChatGPT Surges to 3.1B Visits in September 2024

Similarweb. 2024.ChatGPT Surges to 3.1B Visits in September 2024. https://www.similarweb.com/blog/insights/ai-news/chatgpt-topped-3- billion-visits-in-september/ [Online]. Available

2024

[7] [7]

2025.AI Chatbot Market Share Worldwide

StatCounter Global Stats. 2025.AI Chatbot Market Share Worldwide. https: //gs.statcounter.com/ai-chatbot-market-share [Online]. Available

2025

[8] [8]

2025.Search Engine Market Share Worldwide

StatCounter Global Stats. 2025.Search Engine Market Share Worldwide. https: //gs.statcounter.com/search-engine-market-share [Online]. Available

2025