Generative Search Engines Overview
- Generative search engines are information systems that combine large language models with traditional search techniques to produce synthesized, citation-bearing responses.
- They enable multi-turn conversations and generate explanations, summaries, comparisons, and code, shifting users from simple lookup to complex knowledge work.
- Optimization paradigms like GEO and SAGEO focus on enhancing retrieval, citation integrity, and evidence synthesis, impacting authority and user experience.
Searching arXiv for papers on generative search engines, GEO, and evaluation. Generative Search Engines (GEs) are information systems that combine the capabilities of a LLM with a traditional search engine, thereby shifting online information access from ranked retrieval toward synthesized, citation-bearing, natural-language responses. In this operationalization, a GE differs from a conventional search engine along three main dimensions: capabilities, because it can generate explanations, summaries, comparisons, code, and other synthesized artifacts; interaction pattern, because it supports multi-turn conversational interaction; and output format, because it returns direct natural-language responses rather than primarily a list of links (Suri et al., 2024). Subsequent work treats this shift as more than an interface change: it alters what users do with search, what content becomes visible, how source authority is established, how optimization should be evaluated, and how exposure, trust, and monetization are redistributed across the web ecosystem (Kirsten et al., 13 Oct 2025).
1. Historical emergence and definitional scope
Early empirical work defined a generative search engine through a concrete operational case: Bing Copilot, described as a search system that “marries the new capabilities of LLMs with a traditional search engine which can retrieve information from the web” (Suri et al., 2024). In this framing, GEs do not merely retrieve and rank existing documents; they generate new artifacts such as explanations, summaries, analyses, comparisons, and code, while supporting multi-turn interaction and direct answer presentation. This suggests that the relevant unit of analysis is no longer only document retrieval, but also the transformation of retrieved material into synthesized output.
Later work generalized this formulation in several directions. One line defines generative search as a form of web search in which an LLM retrieves web pages related to a query and then generates a single coherent natural-language response, in contrast to traditional search, where the output is a ranked list of largely independent web pages (Kirsten et al., 13 Oct 2025). Another line narrows the concept to Search-Augmented Generative Engines (SAGE), formalizing systems that answer a query through a pipeline in which a retriever , reranker , and generator interact over a corpus : Under this view, generation-stage visibility is downstream of retrieval and reranking, so a document must first survive earlier stages before it can ever be cited (Kim et al., 12 Feb 2026).
The literature also distinguishes GEs from adjacent systems such as general chatbots. Audit work on ChatGPT, Bing Chat, and Perplexity argues that these systems occupy the social role of search engines because they accept user queries, return information about the web, and in some cases integrate and cite current web sources while synthesizing direct answers rather than simply ranking links (Li et al., 2024). A commentary on “search engines post-ChatGPT” likewise emphasizes that search engines are beginning to generate, index, and distribute AI-generated content, thereby changing the epistemic role of search from helping users locate information to actively composing and circulating knowledge claims (Memon et al., 2024).
2. Behavioral shift from lookup to knowledge work
The first large-scale behavioral portrait of a GE in the wild found that usage differs sharply from conventional web search. Using eight consecutive weeks of production logs from May 28, 2023 through July 22, 2023, the study sampled 80,000 de-identified Bing Copilot conversations and matched them to a comparable Bing Search sample. After exclusions, the final fully classified datasets contained 77,303 Bing Copilot conversations and 78,701 Bing Search sessions (Suri et al., 2024). The paper’s central question was how people use generative search engines and how that usage compares to traditional search.
Topically, Bing Copilot usage concentrated in domains such as Business and economics, Translation and language learning, Web development, Creative writing and editing, Biology and medicine, Programming and scripting, Academic writing and editing, Education and learning, Engineering and design, Law and politics, and Data analysis and visualization (Suri et al., 2024). When Bing Search sessions were reclassified using the Bing Copilot taxonomy, the top 16 Bing Copilot categories accounted for 86.4% of the Copilot sample. Within those categories, nine were substantially more prevalent in Bing Copilot, at roughly 1.5x to 2x the Bing Search proportion: Translation and language learning; Creative Writing and editing; Biology and medicine; Programming and scripting; Academic Writing and editing; Small talk and chatbot; Engineering and design; Law and politics; Data analysis and visualization (Suri et al., 2024). This supports the interpretation that GEs are disproportionately used in knowledge-intensive domains.
The same study operationalized knowledge work at both the domain level and the conversation level. Domain-level coding aligned domains to four U.S. Bureau of Labor Statistics sectors: Professional, Scientific, and Technical Services; Information; Finance and Insurance; Educational Services. The result was a large difference: 72.9% of Bing Copilot conversations fell in domains coded as knowledge work, compared with 37% of Bing Search sessions (Suri et al., 2024). At the conversation level, highly knowledge-work-heavy Copilot domains included Data analysis and visualization (89% classified as knowledge work), Academic writing and editing (89%), and Web development (81%). By contrast, in Bing Search only Science (64%) had a majority of sessions classified as knowledge work, while Social Media (1%) and Shopping (4%) were much lower (Suri et al., 2024).
Task complexity was analyzed using Anderson and Krathwohl’s taxonomy: Remember, Understand, Apply, Analyze, Evaluate, Create. In the final complexity analysis, 37.0% of Bing Copilot conversations fell into the higher-complexity categories Apply, Analyze, Evaluate, or Create, whereas only 13.4% of Bing Search sessions did so (Suri et al., 2024). The study explicitly interprets this as a “large qualitative shift”: users are not simply looking information up but doing tasks that involve processing information, for example by analyzing or evaluating it. This suggests that generative search reorganizes search behavior around task assistance rather than navigation or lookup.
A later comparative study broadens the behavioral picture. Across Google organic search and four GEs from Google and OpenAI, the work shows that generative search engines differ from traditional search not only in output modality but also in retrieval breadth, internal-versus-external knowledge balance, and concept exposure (Kirsten et al., 13 Oct 2025). Generative systems often consult source pools that only partially overlap with Google’s top-10 or top-100 organic results, and they vary sharply in how much they rely on parametric knowledge versus external retrieval. A plausible implication is that the behavioral shift observed in Bing Copilot is accompanied by a structural shift in what evidence enters the answer in the first place.
3. Output structure, source mediation, and authority construction
A core property of GEs is that they compress multiple sources into a single narrative. This changes how authority is displayed and how users encounter evidence. An audit of ChatGPT, Bing Chat, and Perplexity across 1,008 responses to 48 authentic queries on climate change, vaccination, alternative energy, and trust in the media found that these systems are not merely summarizing neutrally; they construct responses in ways shaped by query framing, topic, system design, and source ecology (Li et al., 2024). The paper defines “Generative AI search engines” as hybrid systems that combine LLMs with search results to generate responses and links to current web content.
The audit found a moderate positive correlation between query polarity and response polarity, reported overall as “(1,006) = .46, p < .001,” with query polarity explaining 21% of variation in response polarity. By system, the correlations were ChatGPT , Bing Chat , and Perplexity , all (Li et al., 2024). Combined with the coding result that about 81% of Bing and Perplexity responses supported the query, this indicates that GEs often reinforce user framing rather than counterargue it. The paper interprets this as a risk for confirmation bias and reduced perspective diversity.
The same audit analyzed rhetoric and source ecology in 672 Bing and Perplexity responses. Bing used first person singular in 81.0% of responses, whereas Perplexity used third person/passive voice in 86.0% of responses (Li et al., 2024). Both systems frequently cited generic authorities, named authorities, and quantification, while also relying heavily on hedging. This suggests that GEs construct authority through a mixture of fluent narrative, references to expertise, selective sourcing, and statistical cues, while softening responsibility through hedging and balance devices.
The source analysis found 3,448 citations across 355 unique domains. The most common domain categories were News and Media (21.8%), Business (17.9%), Government (16.5%), Non-Profit (13.3%), and Digital Media Site (11.7%), while Educational Institutions (2.6%), Academic Publishers/Journals (3.0%), Policy Institutes (1.7%), and Associations (3.1%) were much rarer (Li et al., 2024). Geographic skew was also substantial: 95% of domains originated in the U.S., UK, Canada, or international organizations, and 65% of references were to U.S. sources despite all queries being issued from Canada. This suggests that GE authority is partly a product of source selection biases rather than only answer-generation style.
A commentary on post-ChatGPT search sharpens the epistemic critique. It argues that GEs can reduce reliability through contextual misgeneration, hallucination, reluctance to say “I don’t know,” weakened provenance, and an “efficiency-reliability trade-off” in which convenience comes at the cost of depth, diversity, accuracy, and verifiability (Memon et al., 2024). The commentary cites prior work reporting that across four major generative search engines only 51.5% of generated sentences are fully supported by citations and only 74.5% of citations support their associated sentence (Memon et al., 2024). Since this statistic is reported through cited prior work rather than direct experimentation in that commentary, it is best interpreted as a secondary indicator rather than a primary result of the paper itself.
4. Source selection, diversity, and robustness
Because GEs must decide which retrieved material to use, source selection becomes a central object of analysis. Comparative work on generative and traditional web search shows that GEs often draw from broader and different source pools than ordinary search results (Kirsten et al., 13 Oct 2025). For Google AI Overviews, 53% of consulted domains were not contained in top-10 Google results, and 27% were not contained in top-100 Google results; overlap could be even lower for other generative systems (Kirsten et al., 13 Oct 2025). This indicates that GEs are not simply paraphrasing the first-page SERP.
The same study shows strong heterogeneity in the balance between external retrieval and internal model knowledge. On average, Google AI Overviews consulted only 0.4 web pages per query, compared with 8.6 for Gemini and 8.5 for GPT-4o Search (Kirsten et al., 13 Oct 2025). Yet aggregate topic coverage remained surprisingly similar across systems: average topic coverage was 0 for all Google organic results combined, 1 for Gemini, 2 for GPT-4o Search, 3 for AI Overviews, and 4 for GPT-4o with Search Tool (Kirsten et al., 13 Oct 2025). The important difference was not total topic count but which concepts were surfaced, especially on ambiguous or time-sensitive queries. A plausible implication is that GEs can preserve average conceptual breadth while still changing the interpretive center of an answer.
Robustness under adversarial conditions is a separate concern. An evaluation of Bing Chat, PerplexityAI, YouChat, and non-retrieval LLM baselines on adversarial factual questions found that average pre-attack accuracy was 95.8%, but average attack success rate (ASR) was 25.1% (Hu et al., 2024). Generative search engines were more vulnerable than non-retrieval LLMs on average: average ASR of search engines was 31.6%, compared with 24.4% for LLMs without retrieval (Hu et al., 2024). Attack effectiveness varied by perturbation type, with Numerical Manipulation averaging 52.1% ASR and Multihop Extension 43.2%, while Facts Reversal averaged 14.4% (Hu et al., 2024). The paper argues that retrieval augmentation does not guarantee robustness and may fail when the system does not verify the manipulated factual component against evidence.
Security-oriented work on citation vulnerabilities extends the robustness question from answer correctness to source manipulability. In political question answering across Japan and the United States, a study of GPT-5, Claude Sonnet 4, and Gemini Flash 2.0 defined the content-injection barrier as the practical difficulty for an actor to publish content on a domain so that it can be retrieved and cited for a poisoning attack (Mochizuki et al., 8 Oct 2025). It found that primary-source citations accounted for about 60%–65% of political citations in Japan, but only 25%–45% in the U.S. (Mochizuki et al., 8 Oct 2025). The same paper reports that low-barrier sources that can be published with only registration account for approximately 30% of citations in answers, yet are poorly reflected in answer content (Mochizuki et al., 8 Oct 2025). This suggests that citation itself can be an attack surface, and that GE evaluation must consider source robustness in addition to sentence-level faithfulness.
A later audit focused specifically on synthetic sources. Across 712 real-world queries on politics, health, and environment, ChatGPT, Copilot, Gemini, and Perplexity produced 2,848 responses with 26,266 unique cited URLs from 7,675 domains. Of the 19,154 scraped cited URLs, Pangram classified 15.2% as Highly Likely AI and 0.7% as Likely AI; excluding the ambiguous “Possibly AI” label, this yielded an estimate of about 16% AI-generated cited sources (Allaham et al., 22 May 2026). Provider-level prevalence varied sharply: Copilot 27.8%, Gemini 14.7%, Perplexity 9.4%, ChatGPT 7.3% (Allaham et al., 22 May 2026). The paper also found a strong head-plus-long-tail structure: the top 25 domains accounted for 23.8% of all citations, while 75.6% of cited domains were cited at most twice (Allaham et al., 22 May 2026). This indicates that GEs combine a repeated head of familiar domains with a broad periphery in which synthetic-source exposure is widely distributed.
5. Optimization paradigms: GEO, SAGEO, and content influence
As GEs became established as citation-mediated systems, optimization research shifted from SEO-style rank manipulation toward answer-level inclusion. A travel-domain study framed Generative Engine Optimization (GEO) as rewriting raw website content 5 into an optimized version 6 more likely to be surfaced in generative responses (Lüttgenau et al., 3 Jul 2025). Using a BART-base model fine-tuned on 1,905 cleaned travel-domain pairs, the paper reported intrinsic improvements over baseline BART—ROUGE-L 0.249 vs. 0.226, BLEU 0.200 vs. 0.173, PPL 1.50 vs. 1.71—and, more importantly, improved visibility in Llama-3.3-70B responses by 15.63% in absolute word count and 30.96% in position-adjusted word count (Lüttgenau et al., 3 Jul 2025). The optimization signal came from synthetic rewrites that integrated “credible-sounding citations,” improved fluency, inserted “compelling statistics,” and reorganized content into separable sections with descriptive headings and structured paragraphs.
Later work argued that generation-only GEO is incomplete because retrieval and reranking gate whether a document can ever be seen by the generator. SAGEO Arena addresses this by evaluating optimization end to end over a corpus of 171,003 unique web documents built from 2,700 test queries across nine domains (Kim et al., 12 Feb 2026). Its pipeline uses BM25 retrieval, Qwen3-Reranker-4B reranking, and gpt-5-mini generation over top-10 candidates (Kim et al., 12 Feb 2026). The benchmark shows that body-text-only optimization often hurts performance in realistic conditions: on average, retrieval H@20 drops from 0.58 to 0.53, reranking H@10 drops from 1.00 to 0.84, and citation rate drops from 0.50 to 0.47 (Kim et al., 12 Feb 2026). By contrast, structural-only optimization raises retrieval H@20 from 0.58 to 0.71 and citation rate from 0.50 to 0.52, while jointly optimizing both structure and body produces mixed outcomes (Kim et al., 12 Feb 2026). This leads to a stage-aware view of optimization: retrieval favors concise, keyword-aligned, entity-rich structural fields; reranking favors direct relevance and answer prominence; generation favors self-contained, evidence-rich body passages (Kim et al., 12 Feb 2026).
Several frameworks push GEO further toward adaptive, engine-specific strategy learning. MAGEO treats GEO as a strategy learning problem and introduces a multi-agent framework with a Preference Agent, Planner Agent, Editor Agent, and Evaluator Agent, together with a Skill Bank that stores reusable, engine-specific editing skills (Wu et al., 21 Apr 2026). Its evaluation metric, DSV-CF, combines normalized surface semantic visibility, normalized intrinsic semantic impact, and an attribution-accuracy penalty: 7 with 8 and 9 in experiments (Wu et al., 21 Apr 2026). On MSME-GEO-Bench and GEO-Bench, MAGEO consistently outperformed heuristic baselines across GPT-5.2, Gemini-3 Pro, and Qwen-3 Max, while ablations showed strong dependence on engine-specific preference modeling and the Skill Bank (Wu et al., 21 Apr 2026).
AutoGEO approaches the same problem by automatically learning GE preference rules from contrasting document pairs and using those rules either as context engineering for a prompt-based system or as rule-based rewards for a compact RL-trained model (Wu et al., 13 Oct 2025). On GEO-Bench and two newly constructed benchmarks, AutoGEO0 and AutoGEO1 improved GEO metrics while preserving utility, with the paper reporting a 35.99% average improvement and an inference cost for AutoGEO2 of about 3 that of AutoGEO4 (Wu et al., 13 Oct 2025). The extracted rules emphasize citation/trustworthiness, factual accuracy, comprehensiveness, structure, self-containment, balanced tone, and domain-specific usefulness.
AgenticGEO formulates GEO as a content-conditioned control problem under black-box, non-stationary engine behavior (Yuan et al., 2 Mar 2026). It combines a MAP-Elites archive of diverse strategies, a Co-Evolving Critic that predicts impression gains, and multi-turn inference-time planning. In-domain and cross-domain experiments across GEO-Bench, MS MARCO, and E-commerce show state-of-the-art performance over 14 baselines, with strong transferability and reduced GE interaction cost due to the critic (Yuan et al., 2 Mar 2026).
A parallel line reframes GSEO as semantic influence rather than citation count. CC-GSEO-Bench is constructed around 1,030 unique source articles and 5,353 query-article pairs, with six influence dimensions: Citation Prominence, Attribution Accuracy, Faithfulness, Key Information Point Coverage, Semantic Contribution, and Answer Dominance (Chen et al., 6 Sep 2025). Its multi-agent optimizer MACO outperforms all 18 baseline metrics and shows that iterative, content-centric, article-level optimization is stronger and more stable than static heuristic rewriting (Chen et al., 6 Sep 2025). This suggests that the natural object of optimization in GEs may be an article’s influence over a family of related questions rather than its behavior on a single query.
Not all optimization work is text-only. A production-scale framework for Pinterest describes a visual GEO pipeline in which a VLM predicts what users would actually search for, rather than merely captioning images; agents mine external trend signals; semantically coherent collection pages are created via ANN infrastructure; and a hybrid ranking system propagates authority signals across billions of images (Zhang et al., 3 Feb 2026). The paper reports 20% organic traffic growth contributing to multi-million monthly active user growth and frames this as evidence that multimodal platforms must transform weak individual assets into intent-aligned, semantically rich, authority-bearing retrieval surfaces (Zhang et al., 3 Feb 2026).
6. Economic, governance, and societal implications
The rise of GEs changes not only search behavior and publisher optimization, but also monetization and exposure dynamics. An economic model of GEs treats the platform as choosing between with-ad and ad-free generated responses in a dynamic Stackelberg game, where immediate ad payoff must be balanced against long-run retention and subscription conversion (Zhang et al., 30 Mar 2026). The model yields a cutoff rule: ads are shown if and only if the with-ad action value exceeds the ad-free value. Competition from rival GEs shifts the optimal policy toward ad-free responses, because stronger outside options reduce the sustainability of ad-heavy monetization (Zhang et al., 30 Mar 2026). This suggests that answer synthesis changes not only publisher traffic flows but also the platform’s own dynamic optimization problem.
Exposure allocation inside GEs is itself subject to external attention structures. In an audit of 44 Web3 enterprises and their creator ecosystems on X, citation panels from Grok, Perplexity, and Grok(X-only) showed systematic exposure bias toward already prominent creators (Alipour et al., 5 Jan 2026). Top-10 creators received significantly more citation exposure than bottom-half creators: +4.32 percentage points for Grok, +4.79 for Perplexity, and +7.16 for Grok(X-only), each with bootstrap confidence intervals excluding zero (Alipour et al., 5 Jan 2026). The paper interprets this as evidence that external attention markets can become exposure allocations inside GEs, entrenching incumbents and narrowing viewpoint diversity.
The literature also underscores epistemic and governance risks. Audit work on public-importance topics emphasizes that GEs are becoming “arbiters of public knowledge” because they do more than retrieve: they frame, synthesize, and legitimate information (Li et al., 2024). Commentary on post-ChatGPT search warns that search engines now generate, index, and distribute AI-generated content, creating recursive feedback loops in which provenance becomes harder to recover and errors can compound across the web ecosystem (Memon et al., 2024). Synthetic-source audits reinforce that concern by showing that cited inputs can themselves be AI-generated at non-trivial rates (Allaham et al., 22 May 2026). A plausible implication is that governance for GEs must address not only answer-level accuracy but also source provenance, citation standards, and downstream concentration effects.
The cumulative picture is that GEs constitute a new class of information system whose distinctive properties are synthesis, citation-grounded answering, adaptive retrieval, mixed internal/external knowledge use, and strong dependence on source selection and presentation. Early behavioral evidence shows a shift toward knowledge work and higher cognitive complexity (Suri et al., 2024). Comparative system studies show that GEs differ from traditional search in source breadth, retrieval depth, and conceptual framing (Kirsten et al., 13 Oct 2025). Audit and security work show vulnerabilities in authority construction, robustness, provenance, and source manipulability (Li et al., 2024, Hu et al., 2024, Mochizuki et al., 8 Oct 2025, Allaham et al., 22 May 2026). Optimization research shows that visibility in GEs is stage-dependent, engine-sensitive, and increasingly treated as answer influence rather than retrieval rank (Kim et al., 12 Feb 2026, Wu et al., 21 Apr 2026, Wu et al., 13 Oct 2025, Chen et al., 6 Sep 2025). Taken together, these findings indicate that generative search is best understood not as a better search box, but as an infrastructural reorganization of retrieval, synthesis, citation, and exposure.