Search Visibility OS
Research separates citation selection from citation absorption in AI search
A 2026 preprint proposes a two-stage GEO measurement framework separating citation selection from citation absorption and evaluates 602 controlled prompts across ChatGPT, Google AI Overview/Gemini and Perplexity using a public dataset.
DeepTRACE audits citation reliability across generative search and deep-research systems
DeepTRACE introduces an eight-dimension audit framework spanning answer text, sources, and citations, and applies it to generative-search and deep-research systems. The study reports substantial unsupported-statement rates and citation accuracy ranging roughly 40–80% across evaluated systems.
Mechanistic study separates citation-marker generation from evidence use in RAG
van Dort and Heuss use activation patching in Llama-3.1-8B-Instruct to identify distributed components involved in deciding whether to emit inline citations, then test interventions on PopQA-derived prompts and HotpotQA.
ReCite evaluates intent-aware retrieval and verification for faithful academic citations
Huang and colleagues introduce ReCite, a peer-reviewed agentic academic citation system that identifies citation locations, plans searches from citation intent, verifies claim-evidence support and retries retrieval after mismatches.
EMNLP study traces deep-research citation errors to agent orchestration
Peer-reviewed research by Hirsch and colleagues evaluates citation faithfulness in three open-source multi-agent deep-research systems using local diagnostics of agent invocations. In AI-Q, the authors attribute 84.7% of final-report errors to the orchestrator; roughly 31% of those orchestrator errors are hallucinations and the remainder citation mistakes.
CITECHOICE separates citation allocation from source admission in agentic search
CITECHOICE uses controlled replay of authentic multi-turn agentic-search transcripts to test how document presentation and order affect citation allocation when competing sources support the same fact.
Peer-reviewed audit finds AI Overview citation selection differs by document provenance after rank controls
Kakimov and colleagues published a peer-reviewed observational framework for auditing citations in Google AI Overviews using YMYL queries drawn from MS MARCO. Their publisher abstract reports that documents classified as AI-generated were cited more frequently than human-authored documents after controlling for observed retrieval rank, particularly among citations outside the observed retrieved results.
Study finds model-specific concentration in AI search news citations
Kai-Cheng Yang analyzed 24,069 AI Search Arena conversations and 366,087 citations from 12 OpenAI, Google and Perplexity search models in March–May 2025, finding distinct model-family patterns and concentration among news-source citations.
Bing AI Performance adds Intents, Topics, Citation Share and Compare previews
Microsoft introduced four preview AI visibility measurement capabilities in Bing Webmaster Tools: Intents, Topics, Citation Share and Compare.
AgentGEO research diagnoses and repairs citation failures across simulated generative-engine stages
An original research preprint introduces a citation-failure taxonomy, an iterative AgentGEO diagnosis-and-repair framework, and the MIMIQ document-centric benchmark with 204 webpages and 12,240 generated queries. In controlled simulated generative-engine evaluation, the authors report higher citation rates after targeted repair.
MCite-RL research improves visual citation grounding through iterative evidence refinement
Peer-reviewed EMNLP 2026 research introduces MCite-RL, which combines iterative visual evidence retrieval and cropping with citation-aware reinforcement learning to improve localized citations in multimodal RAG.
MiRAGE introduces claim-level information and citation metrics for multimodal RAG
Martin and colleagues' peer-reviewed EMNLP 2026 research introduces MiRAGE, a claim-centric evaluation framework for retrieval-augmented generation from text and audiovisual sources. It separately measures information factuality and coverage (InfoF1) and citation support and completeness (CiteF1), and compares human and automated evaluations against other RAG metrics.
Research benchmarks citation-ranking models across ten LLM engines under controlled retrieval
Bajemon and Rochet benchmarked citation allocation on 229 GEO-Bench queries across ten LLM engine configurations given five fixed passages per query, reporting stronger predictions from comparative decision models than from generic relevance rerankers and releasing a distilled 4B ranker.
Research finds AI news citations more concentrated than audience traffic across countries
An October 2026 original preprint measures digital-news ownership across 77 countries and compares citation concentration from AI systems with audience traffic, finding higher concentration in the studied citation panels.
Research separates retrieval, content matching and model knowledge in AI citations
Chen and Koudas report a study of 158,099 AI answers and 468,037 citations in two commercial markets, combining within-website analysis, API search replay and controlled experiments. The study finds that retrieval and semantic question matching explain citation behavior better than several common page-format signals in its tested settings.
Research demonstrates citation-governance risks from low-barrier web publishing
Liu and colleagues map 17,211 citation instances across 6,356 domains and conduct controlled publication experiments across 10 AI-search platforms. Their study reports that 8 of 10 platforms cited a fabricated concept within seven days, revealing potential weaknesses where favored source domains permit low-barrier publishing.
Cross-engine GEO audit finds little exact-URL citation overlap in a fixed-prompt sample
Benjamin Tannenbaum reports a June 6, 2026 observational audit of 15 fixed commercial prompts across ChatGPT, Microsoft Copilot, Google and Perplexity. The study reports 589 citation observations representing 528 distinct URLs and 356 domains; mean pairwise same-prompt exact-URL Jaccard overlap was 0.0079, and 84.9% of compared engine pairs shared no cited URL.
Gemini 3.8 Flash rollout briefly disrupts citations in Google AI Mode
Early testing of Gemini 3.8 Flash in Google AI Mode showed sharply reduced or absent source links for some queries. Google Search VP Robby Stein confirmed the behavior was unintended, and links were observed returning the following morning.
Bing Webmaster Tools launches AI Performance citation reporting in public preview
Microsoft launched an AI Performance public preview in Bing Webmaster Tools for measuring publisher citations across Microsoft Copilot, Bing AI summaries and select partner integrations, including cited URLs, sampled grounding queries and citation trends.
ALCE introduces an end-to-end benchmark for LLM answers with citations
Gao and colleagues introduced ALCE, an evaluation benchmark for LLM systems retrieving evidence and generating cited answers, with measures for fluency, correctness and citation quality.
Large-scale study separates citation-pool presence from entity exposure in generative search
A large-scale empirical study of Chinese-language generative search analyzed eight web/app interfaces, 614 queries and 160,860 cleaned citation-level records, finding that entities present in citation pools were selectively surfaced and that source sets differed systematically between app and web interfaces.
Gemini hotel citation audit finds query intent shifts OTA versus non-OTA source shares in Tokyo
An original March 2026 audit of 156 paired Tokyo hotel prompts to Gemini 2.5 Flash with Google Search grounding classified 1,357 cited pages as OTA or non-OTA. Experiential queries produced a higher share of non-OTA citations than transactional queries (55.9% versus 30.8%).
Retrieved but Not Read study separates search-result retrieval from source reading and citation faithfulness
A 2026 research preprint tests how search-augmented LLMs use current evidence across 225 changed facts and reports that snippet-only answers correlate with more stale answers and unsupported citations than answers citing opened pages.
EBG study finds complete evidence retrieval and citation coverage are distinct RAG bottlenecks
Peer-reviewed EMNLP Findings research introduces evidence-based generation (EBG), requiring line-level citations and a joint answer-and-evidence metric. Controlled evaluations identify incomplete multi-passage retrieval and loss of document-level context as limitations of similarity-based chunked RAG.
Controlled SIGIR study isolates competitive first-citation selection across six LLMs
Vishwakarma, Kumar and Jamidar report 252,000 controlled paired-source trials across six LLMs and 18 content factors. In a two-source simulated retrieval context, topical relevance and source order most strongly affected which source received the first citation.
FeatGEO research evaluates feature-level optimization for citation visibility and content quality
Liu and Xu introduce FeatGEO, which optimizes interpretable webpage structural, content and linguistic features before generating revised text. Their GEO-Bench experiments across three generative engines report stronger citation visibility than token-level rewrite baselines while maintaining or improving content quality.
GEO-SFE tests how content structure shapes citation behavior
Researchers proposed GEO-SFE, separating document structure into macro, meso, and micro levels, and reported experiments across six generative engines.
GEO-Flag research measures GEO-like web content and citation-source verifiability
An original research preprint introduces GEOFlagBench and an intervention-paired detector for GEO-optimized webpages. Applied to 10,095 usable pages from released Google Search and Gemini-grounded retrieval results, the model flagged 8.90% as GEO-like; the authors explicitly characterize this as an estimate without ground-truth labels.
ACM health-search audit finds limited overlap between Google AI Overview citations and organic sources
Sünkler and colleagues published an ACM WebSci Companion audit of 900 English health queries on Google US, collected February 2–4, 2026. AI Overviews appeared for 87.7% of this constructed query set, and the average observed domain-level overlap between AI-cited sources and the organic top ten was 46.2%. YouTube represented 6.8% of AI citations versus 0.6% of organic results in the sample.
Research traces ChatGPT Web retrieval-to-citation behavior across model settings
An original observational study compares four ChatGPT Web settings from GPT-5.2 to GPT-5.5, examining query rewriting, retrieved candidates and cited sources at URL, hostname and entity levels, with external Google Search and OpenAI Responses API proxies.
Deep literature search study separates citation coverage from topical relevance
Original research on a 250-paper computer-science benchmark finds that breadth-first bibliography expansion greatly improves scholarly retrieval recall and that reference-list overlap alone does not capture topical citation relevance.
Research models GEO citation competition and tests verifiable-content rewards
Xu, Guo and Xiong proposed a repeated-game model of GEO content rewriting and generative-engine defenses and evaluated a verifiable-content reward (VCR) mechanism that credits checkable factual substance while penalizing manipulative citation-seeking rewrites. Their preprint reports improved defense-utility scores on controlled benchmarks.
XBCP research finds language mismatch weakens agentic retrieval and citation fidelity
Lu and colleagues introduce Cross-lingual BrowseComp-Plus (XBCP), keeping questions and answers in English while varying the languages of supporting evidence. They evaluate four deep-research agents and sparse/dense multilingual retrievers; the multilingual condition distributes documents across 12 languages. Authors report lower evidence recall, answer accuracy and citation reliability with language-mismatched evidence.
Research compares generative-search citations with Google source overlap and freshness
Chen and colleagues compared Google top-ten results with citations from four generative search configurations using 1,000 ranking-style prompts and found low average domain overlap, alongside intent- and vertical-specific source differences.
Research shows repeated UGC retrieval can contaminate deep-research citations
A Cornell Tech original preprint demonstrates that frequently retrieved user-generated pages can become a cross-query poisoning surface for open deep-research agents, influencing both retrieved evidence and report citations.
Study compares language predictability and semantic similarity in generative-search citations
Ma and colleagues studied approximately ten thousand websites across Google conventional and generative search and reported differences in language-model predictability and semantic similarity of cited content. Additional controlled Gemini RAG and participant experiments examined LLM content polishing.
Longitudinal audit measures AI Overview activation, citation overlap and claim grounding
An original longitudinal audit issued 55,393 trending queries across 19 categories and examined when Google AI Overviews appeared, which domains they cited, and whether cited text supported generated claims.
Medical AI Overviews study separates source selection, citation grounding and accuracy
An original preprint audits Google AI Overviews for 3,611 benchmark health questions, comparing cited sources with traditional Google results, source credibility, statement grounding and physician-written accuracy rubrics.
Controlled research finds location channels change AI citations and business recommendations
Chen and Koudas report a 31,200-answer controlled audit of how declared and network location influence sources cited and businesses recommended by tested ChatGPT and Gemini configurations.
Randomized publisher experiment finds AI answers reshape citation exposure and reading behavior
An original randomized field experiment with 37,561 readers of The Washington Post archive compared conventional archive search with an interface also showing cited AI answers. The authors report more overlap in consumed topics, shifts toward cited articles and follow-up searches, and a small increase in article consumption per reader despite reduced article consumption per search.
MAGEO research tests reusable GEO strategies with fidelity-aware citation evaluation
An original research preprint introduces MAGEO, a multi-agent GEO framework that learns reusable, engine-specific content-editing strategies. It also introduces twin-branch evaluation with a frozen retrieval list and DSV-CF metrics combining semantic visibility with attribution fidelity.
OpenAI WebGPT demonstrates browser-based answering with citations
OpenAI published WebGPT, a GPT-3-based system that searches the web, follows links and cites sources while answering questions.
EvidenT study tests source verification before enterprise RAG answer generation
Peer-reviewed EMNLP Industry research describes a lightweight source-grounding pipeline that checks extracted evidence against retrieved documents before generation. On roughly 500 enterprise queries, the authors report improved gold-source hits and no citations to non-retrieved URLs.
Two structured GEO experiments test earned-source effects on AI visibility
Two practitioner-led GEO experiments tracked 15 commercial-intent queries across multiple AI platforms and logged 775 citation events, finding that a small number of third-party placements accounted for much of the measured citation visibility.
Bing Webmaster Tools introduces Topics Portfolio (Beta) for AI Performance
Microsoft Bing introduced Topics Portfolio (Beta) within Bing Webmaster Tools AI Performance, showing personalized thematic clusters of grounding queries with approximate ecosystem metrics and site-specific citation opportunities.
Research re-tests historical GEO levers on modern generative engines
A 2026 study validates an engine-free GEO measurement protocol and re-tests influential 2023-era content interventions against ten modern engine families. Under its paired, volume-controlled harness, quotation addition, statistics addition and cite-sources do not increase citation on the evaluated modern engines.
Peer-reviewed study measures attribution gaps between retrieved and cited AI-search sources
A peer-reviewed audit of 13,929 LMArena search-log observations across 11 model configurations measures the difference between relevant URLs disclosed as retrieved and URLs cited in AI answers.
Research finds AI-search source selection and answer synthesis diverge across systems
A March 2026 original preprint compared 11,000 real-world information queries across GPT-4o-mini API configurations, Google AI Overviews and Google organic results. It reported low overlap between top cited domains of the OpenAI API web-search condition and Google, and uneven use of cited source content in generated answers.
Cloudflare–OpenAI pilot uses network signals to expand ChatGPT Search index discovery
Cloudflare and OpenAI reported early results of a pilot begun in July 2026 in which Cloudflare network metadata helped OpenAI discover candidate pages for its search index. The partners said a cohort covering about 1% of Cloudflare's network helped discover and index hundreds of millions of previously unknown URLs; OpenAI separately evaluates and ranks pages.
T-Search research separates multi-step evidence retrieval from answer generation
Tsymboi and colleagues introduce T-Search, an open-weight multi-round agentic retriever over fixed corpora that returns ranked evidence chunks and justifications to a separate answer generator. Across seven English and Russian benchmarks with gold evidence labels, the authors report mean Recall@10 of 56.0 in one rollout, 14.4 points above their base model, and 61.3 with three fused rollouts.