AI Search Research
Studies and datasets on how AI search systems retrieve, cite and recommend. Each summary is checked against its source and labelled by type: academic and preprint research, industry data from vendors and agencies, or official product updates.
Retrieved but Not Read: Understanding Why Search-Augmented LLMs Fail to Use Available Evidence
arXiv · Owen Walters, Lihu Chen, Francesca Toni
What it shows
Researchers tested 225 post-training-cutoff facts across six domains on one deployed proprietary model and six controlled open-model set-ups. With identical frozen retrieval, snippet-only answers were stale 6.4% of the time, against 1.7% when a page was opened and 1.3% when an opened page was cited; unfaithful-citation rates were 6.4%, 1.6% and 2.3%. Supplying the full source directly reduced but did not remove failures, and some updated facts were overtaken by stale model knowledge in later layers.
Why it matters for AI search
Retrieval, opening, citation and factual adoption are different states. A visible citation does not prove the answer followed the source.
A source can be retrieved and cited yet still lose to stale model memory.
Source checked 19 Oct 2026. Figures confirmed against the paper's Table 2.
More ways to connect with the web from Search
Google (The Keyword blog) · Hema Budaraju, VP Product Management, Google
What it shows
Google announced more frequent inline links placed higher in AI Mode and AI Overviews. It is also testing dedicated top-of-answer sections for authoritative task-completion pages, expanding timely-article carousels, making end-of-answer links more prominent and adding clickable creator and publisher profiles for video results.
Why it matters for AI search
Citation selection and citation presentation are now separate visibility layers. Two cited sources can receive very different exposure depending on placement, surrounding context and whether Google treats one as an official destination, so citation position and format should be recorded, not just presence.
In AI search, the citation itself now has a ranking position.
Source checked 19 Oct 2026. Announcement confirmed on blog.google.
Twenty Days Later: A Pre-Registered Test of History-Based Forecasts of Brand Mentions in Gemini Answers
arXiv · Dmitrij Żatuchin
What it shows
Identical requests covering 12 software brands, three categories and three languages were repeated after 20.3 days. Historical brand-and-question mention rates predicted the later results with a Brier score of 0.0281, well ahead of brand-level history, the overall rate and a coin-flip baseline. Of 202 cells that had been unanimous eight hours earlier, 194 were unchanged and none fully reversed.
Why it matters for AI search
AI visibility is noisy but not chaotic. Stable prompt-brand relationships can stay forecastable long enough to act on, so tracking effort can focus on unstable or search-grounded prompts.
AI visibility can be noisy within a day yet predictable across weeks.
Source checked 19 Oct 2026. Figures confirmed.
AI Search Visibility Study (2026)
Minuttia · Zacharias Xiroudakis, Minuttia (with Convex Works)
What it shows
Minuttia and Convex Works analysed 24,067 prompts about SaaS brands across Google Search, ChatGPT, Gemini, Claude and Perplexity. For alternative-product searches, about 54% of recommended brands appeared in only one engine and 8.1% in all five. Brands in the top quartile for G2 review counts appeared in Google AI Overviews 7.6 times as often as the bottom quartile on category prompts, while star ratings showed no comparable link; platform scope was the attribute most often raised first in comparisons (48.9%).
Why it matters for AI search
AI systems organise product comparisons around identifiable decision attributes. Knowing which attributes a system uses, and which brands own them, is more useful than counting appearances.
AI compares brands on attributes; the brand that owns the attribute wins the comparison.
Source checked 19 Oct 2026. Figures confirmed; sample is SaaS brands.
Understanding ChatGPT Citations: Evidence from Production Web Search Across Model Generations
arXiv · Yunxuan Jiang, Shiqi Zhang, Anqi Xing, et al.
What it shows
Across 4,238 ChatGPT web runs, instant models generally made one retrieval round, while GPT-5.4 Thinking averaged 11.82 queries over 3.2 rounds and used multiple rounds in 71.3% of runs. Google Search and the OpenAI Responses API together recovered only 26% to 28% of the exact URLs ChatGPT later cited, but about half of the cited hostnames. No observed GPT-5.3 or GPT-5.5 query matched the user's original wording, and sources retrieved repeatedly across runs were more likely to be cited.
Why it matters for AI search
Exact-URL tracking is brittle. Domain and entity presence are more stable signals, and Google rankings or API replays are incomplete proxies for what ChatGPT actually retrieves.
The retrievable entity is more durable than the cited URL.
Source checked 19 Oct 2026. Figures confirmed.
What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization (GEO)
arXiv · Mahe Chen, Nick Koudas
What it shows
Researchers collected 158,099 commercial-query answers containing 468,037 citations over two 90-day periods. Comparing pages within the same site, better paragraph-query matching was linked to 15.5% more citations per standard deviation and title matching to 14.1% more, while FAQ blocks, schema, author boxes, summaries and page speed showed no reliable advantage. Retrieval explained 32% of within-site citation variation versus 2.8% for page features; ChatGPT cited 43.8% of first-ranked retrieved pages but 3.5% at positions 11 to 20, and 46% of its unprompted brand mentions had no cited page naming the brand.
Why it matters for AI search
This is some of the strongest evidence that AI citation is a contest between remembered brands and retrieved evidence. ChatGPT was more likely to insert brands it already 'knew' into its hidden searches, so niche brands depend far more on third-party evidence being found.
GEO is not a page-formatting problem. It is a competition between remembered brands and retrieved evidence.
Source checked 19 Oct 2026. Figures confirmed.
Training with Missed Targets in Generative Recommendation: Separating Supervision from Probability Competition
arXiv · Xuesi Wang, Yangbin Shi, Xiaolin Zheng
What it shows
Generative recommenders can miss the correct item before reranking. Simply appending missed targets during training creates harmful competition with candidates that will actually be available at inference; separating the two improved full-target NDCG by 7.8% to 22.2% across four prespecified comparisons, and a conservative rule avoided a 1.7% loss in a held-out category.
Why it matters for AI search
Candidate-generation failure and ranking failure are different problems. Making an already-retrieved page more persuasive cannot rescue an entity excluded upstream.
No amount of optimisation lifts a brand up a shortlist it never entered.
Source checked 19 Oct 2026. Figures confirmed.
We Retested Our AI Search Study Across Five More Industries. Here's What Changed.
Geeky Tech · Genny Methot, Geeky Tech
What it shows
Geeky Tech expanded its citation research to more than 8,200 questions and 116,000 traced citations across more industries. ChatGPT and Gemini shared only about 5% of cited sources and no pair of the five AI systems exceeded 13% overlap. The median B2B brand appeared in just 13% of AI answers to questions its own website covers, and cited material was often short, self-contained passages high on the page.
Why it matters for AI search
Having information about an attribute on a website does not mean AI systems will associate the brand with it. Evidence that stays isolated on owned pages may achieve little; the association needs to travel.
Owning the answer on a website does not mean owning it in AI.
Source checked 19 Oct 2026. Figures confirmed.
SIFT: Search Intent-to-Filter Transformer for Multi-Task Personalized Filter Ranking at Airbnb
arXiv · Shashank Dabriwal, Tanya Piplani, Hao Li, et al. (Airbnb)
What it shows
Airbnb's model learns latent guest preferences from behaviour and turns them into ranked search filters. Offline, booking and amenity-engagement PR-AUC improved 51.9% and 62.8%. Online, recommended-filter engagement rose 20%, a new hotel-intent filter lifted uncancelled hotel bookings by 3.8%, and marketplace-wide bookings rose 0.76%.
Why it matters for AI search
Recommendation systems can decide which constraints define the list, not just rank it. An entity's attributes matter only when they are represented clearly enough to become an applicable filter.
Recommendation fitness begins with being legible as a constraint.
Source checked 19 Oct 2026. Figures confirmed; no workshop venue stated in the paper.
Smart Content Ingestion for Generative AI Workloads
arXiv · Abbas Raza Ali, Muhammad Ajmal Siddiqui, Moona Zahid
What it shows
The proposed extraction system scored 97.4 out of 100 across 180 mixed documents, with 0.13% character error and 0.995 table similarity. Structure-aware chunking reached Hit@1 of 68.6%, Hit@10 of 92.8% and MRR of 0.77 across 25,050 generated questions. The authors argue that extraction errors cannot be repaired later by retrieval or reranking.
Why it matters for AI search
Valuable claims locked in poorly extracted PDFs, tables, slides or diagrams may never become usable evidence, even when the files are indexed.
Before content can be authoritative, it must survive machine perception.
Source checked 19 Oct 2026. Figures confirmed.
SCOUT: Supply-Aware Cold-Start Proactive Query Suggestion for Travel Search
arXiv · Hao Li, Shashank Reddy, Kedar Bellare, et al. (Airbnb)
What it shows
SCOUT treats the production search engine and reranker as a reinforcement-learning environment and rewards query suggestions that match available travel inventory, rather than learning from historical demand. Inventory match at 18 results improved 12.3%, equalling a compute-heavy best-of-eight approach at no extra inference cost.
Why it matters for AI search
A platform may steer users towards queries its supply can answer well. Expressed demand can be partly constructed by the retrieval system rather than simply observed.
Search systems can manufacture answerable demand.
Source checked 19 Oct 2026. Figures confirmed; an unsupported venue label was removed.
From Valid to Useful: Post-Verification Acquisition for Recursive Self-Improving Recommendation
arXiv · Tonmoy Hasan, Taylor Foust, Shao Tang, et al.
What it shows
Keeping every verified synthetic sequence gives disproportionate influence to sources that produce more or longer continuations. Capping each source's contribution and prioritising sequences where models disagreed beat retain-all training in all 24 dataset-model-metric comparisons, with the best mean in 23.
Why it matters for AI search
Verification alone does not prevent volume-driven concentration: a mass of acceptable material can drown out smaller but more informative evidence. Distinctive information value matters more than page count.
More verified evidence can still create a popularity bias.
Source checked 19 Oct 2026. Figures confirmed.
AI Overviews Currently Appear for 83% of Branded Searches
Ahrefs · Ryan Law, Xibeijia Guan (Ahrefs)
What it shows
Ahrefs analysed 232.8 million US desktop Google results crawls from July to September, including 92.5 million branded-keyword crawls. On 29 September, AI Overviews appeared for 82.91% of branded keywords (73.3% across September). In a panel of 100 exact brand-name searches, 63 had an AI Overview and 48 of those sat in position one; where external sources showed, Wikipedia appeared in 71.9%, YouTube in 38.6% and LinkedIn in 22.8%.
Why it matters for AI search
Even when someone searches a brand by name, Google may show a machine-written interpretation of that brand first, assembled partly from third parties. The brand results page is becoming a distributed narrative surface.
The branded search result now starts with someone else's summary.
Source checked 19 Oct 2026. Figures confirmed.
ScholarCatalyst: A Benchmark for Retrieving Papers That Inspire New Research
arXiv · Sohyeon Kim, Yoonho Lee, Bo Liu, et al.
What it shows
For 207 research projects, 184 lead authors identified earlier papers that could have advanced their work. Agentic search reached 0.42 Recall@20 against 0.48 for plain embedding retrieval, even though the agent used that retriever as a tool; an agent whose model may have seen the finished papers reached only 0.51.
Why it matters for AI search
More agent steps do not automatically improve discovery. Planning can lose or misdirect candidates that a simpler retriever already had, so simple baselines should be measured before crediting agentic fan-out.
Agentic search can orchestrate itself into worse retrieval.
Source checked 19 Oct 2026. Figures confirmed.
Not All Is Lost: Repairing Lossy User Preference States of Personalization Encoders
arXiv · Parthiv Chatterjee, Dhiraj Golhar, Ummesalma Diwan, et al.
What it shows
Preference encoders discarded useful historical evidence even though it remained recoverable in cached intermediate representations. A correction layer improved all twelve recommendation systems tested; for example, Mamba4Rec on MovieLens gained 3.96 MRR points.
Why it matters for AI search
A model's current picture of a person can leave out evidence that still exists in the history. Recommendation failures can happen during preference compression, not only during retrieval.
A system can hold the preference evidence and still not use it.
Source checked 19 Oct 2026. Corrected: MRR gain stated as +3.96 points per the paper; an unverified latency figure was removed.
Generative AI Search on Content Creation Platform: Evidence from a Large-Scale Field Experiment
SSRN · Hongxian Huang, Jinglong Zhang, Jason Chan, et al. (NYU Stern, Tsinghua SEM, Minnesota Carlson)
What it shows
In a field experiment with 100,000 users of a Chinese content platform, treatment users kept a generative-search feature and control users lost it. AI search increased search frequency, active search days and topic breadth. Users opened fewer conventional results per session but searched enough more often that total previews and page opens of user-generated content rose; most extra queries were follow-ups or new tasks.
Why it matters for AI search
Per-query click-through can fall while total demand for source content rises. 'Zero-click' analysis based on single results pages can misread the overall effect of AI search.
AI answers may compress each search while expanding the search market.
Source checked 19 Oct 2026. Key findings confirmed; experiment duration not stated in accessible text, so omitted.
RealWorldShop: Benchmarking and Improving Conversational Shopping Agents in Real-World E-commerce
arXiv · Xinwei Yang, Kelong Mao, Yudong Guo, et al.
What it shows
RealWorldShop has 3.28 million grounded products, about 1,200 shopping scenarios and over 2,000 simulated user profiles. Current agents gave plausible individual replies but struggled to keep revised constraints, coordinate multiple purchases and converge on a grounded decision. An agent with explicit state management, retrieval and runtime guards raised success from 0.71 to 0.82 over the strongest baseline.
Why it matters for AI search
Recommendation fit is session-dependent. A product can suit the opening prompt and become unsuitable once budget, compatibility or logistics come up, so multi-turn journeys matter more than single 'best product' prompts.
Recommendation fitness is the ability to remain eligible as requirements change.
Source checked 19 Oct 2026. Figures confirmed.
When LLM-Inferred User Context Adds Value in Production Streaming Recommendation
arXiv · Milad Sabouri, Neeraj Sharma, Sardar Hamidian, et al.
What it shows
Conventional aggregate profiles beat LLM-written preference summaries for habitual users, about 82% of the sample. LLM profiles helped exploratory users but created a clear popularity-attractor effect: catalogue coverage fell by roughly 80% and novelty by 18% to 22%, even as within-list diversity rose slightly.
Why it matters for AI search
Semantic summarisation can make a list look varied while concentrating exposure on familiar, popular items. Surface diversity and catalogue-level concentration need to be measured separately.
Personalisation can disguise concentration: diverse-looking lists may draw from a shrinking pool.
Source checked 19 Oct 2026. Figures confirmed.
AI Visibility at Scale: Learning From 9M Prompts and 400+ Enterprise Brands
NP Digital · Nikki Lam, NP Digital
What it shows
NP Digital analysed 9 million AI answers across nine platforms and 400+ enterprise brands. On commercial-intent prompts, 82% of citations went to third-party sources and 3% to brand-owned pages, yet when an owned page was cited the brand's visibility rate was five times higher. YouTube gave the strongest third-party lift (2.8x), ahead of LinkedIn (2.7x) and Reddit (2.4x); up to 58% of citations never reappeared, and up to 75% of answers citing a brand's own page did not name the brand.
Why it matters for AI search
A citation is not a brand association. AI systems can use a brand's evidence without passing any recognition back, so the brand-attribute link has to stay explicit as information travels, and third-party coverage carries most of the weight.
Being used as information is not the same as being remembered as a brand.
Source checked 19 Oct 2026. Figures confirmed.
AI Discovery Observatory: Live ChatGPT, Claude, Gemini & Perplexity Citation Data
BrandGhost · BrandGhost
What it shows
The latest BrandGhost run analysed 6,448 citations across ChatGPT, Claude, Gemini and Perplexity. Average cross-engine source overlap was about 10.2% and recommendation overlap 9.2%; in 52.1% of comparable recommendation sets the engines shared no recommended brands. Overlap has drifted down across 12 tracked runs, and the median cited document was 201 days old.
Why it matters for AI search
There is no single recommendation network. Different engines traverse different evidence paths to different shortlists, so broad, independent representation across the information environment is more robust than targeting one retrieval route.
Being chosen is itself unstable across AI interfaces.
Source checked 19 Oct 2026. Figures confirmed as at the 30 September update.
MERGE: Multi-LLM Ensemble for Retrieval via Generative Enrichment
arXiv · Tzu-I Ho, Yung-Yu Shih, Shang-Yu Su, et al.
What it shows
Three small LLMs independently expanded queries before another model combined their suggestions. Using the same BM25 retriever, this lifted nDCG@10 on all five BEIR datasets by 2.1 to 14.9 points, and the combined expansion beat every individual generator.
Why it matters for AI search
Machine-generated support queries can contain entities and terms absent from the visible prompt, and different generators open different lexical routes into the evidence set.
The hidden query is an ensemble hypothesis, not a single deterministic rewrite.
Source checked 19 Oct 2026. Figures confirmed.
Who Does AI Cite? Reconciling Conflicting Evidence on the Sources Behind Generative Search Answers (2024-2026)
SSRN · Rohit Kunal (IndexCraft, Bengaluru)
What it shows
Most disagreement about which sources AI cites reflects incompatible denominators, stages and query sets rather than contradictory engine behaviour. Wikipedia was 7.8% of all citations in one dataset but 47.9% within its top ten domains, and YouTube's reported rise concerned only social citations, about 4% to 7% of the total. Three conclusions hold up: source mixes differ by engine, change by query type and can shift abruptly after retrieval or access changes.
Why it matters for AI search
Universal claims such as 'Reddit gets 40%' are unusable without the engine, query universe, selection stage, denominator and date. This gives a checklist for reading any citation statistic.
Most conflicting GEO statistics are measuring different evidence markets.
Source checked 19 Oct 2026. Figures confirmed.
Search-Aware Reinforcement Learning for Multi-Component Query Understanding in Roblox Game Search
arXiv · Nayoung Choi, Shengjian Chen, Xiaokai Wei, et al. (Roblox)
What it shows
Roblox trained an LLM to turn queries into structured search plans covering routing, normalisation, fan-out, negative constraints, intent and attributes. Component-specific rewards from the live search engine raised NDCG@20 by 8.9 points over supervised fine-tuning and 3.5 points over a single end-to-end reward. The dataset of 16,100 queries includes 7,624 real queries labelled by sustained clicks and play time.
Why it matters for AI search
A user's words become an executable plan before retrieval, and individual plan components can change which entities come back. Search systems may optimise those components separately.
The hidden query is not a rewrite; it is a search execution plan.
Source checked 19 Oct 2026. Corrected: 7,624 real queries are across the whole dataset, not only the evaluation split.
How Reproducible Are Evaluation Conclusions? A Self-Audit of LLM-Inferred Prompt Structure
arXiv · Dipankar Sarkar
What it shows
Across eight models, 72% of prompt-model cells never reproduced an identical structure across runs. Bootstrapping reliably identified the worst performers but not the best: middle-ranked models kept their positions in only 27% to 48% of resamples. Two reasonable data-merging rules changed four of eight rows and moved the headline result by about seven percentage points, and four evaluated endpoints disappeared within ten weeks.
Why it matters for AI search
Repetition alone does not make an AI visibility league table reliable. Prompt sampling, analysis choices, model drift and rank uncertainty can all create a false competitive order.
If the ranking does not survive resampling, the league table is decoration.
Source checked 19 Oct 2026. Figures confirmed.
The Like Trap: Multi-Stage Poisoning against Agents in Similarity-based Recommendation Systems
arXiv · Yue Xing, Pengfei He, Zitao Li
What it shows
A staged series of related posts exploited a like-history feedback mechanism to move progressively less relevant content into an agent's feed. Depending on configuration, 48 to 69 of 200 poisoned posts entered the candidate pool because of the feedback score and 32 to 57 were then liked, some despite falling below the normal similarity threshold.
Why it matters for AI search
It demonstrates a mechanism by which behavioural feedback can bridge a relevance boundary, so small interactions can alter later exposure. That is relevant to defensive monitoring of coordinated engagement.
Feedback can turn yesterday's weak association into tomorrow's retrieval pathway.
Source checked 19 Oct 2026. Figures confirmed.
LLM Citation Drift Report
Techmagnate · Sarvesh Bagla, Techmagnate
What it shows
Techmagnate tracked 83,633 citations from 2,191 domains over eight weeks in India's personal-loan category across ChatGPT, Perplexity and Google AI Mode. Average week-to-week domain churn was 39%, rising to 44.8% in ChatGPT. Yet only 15.1% of domains appeared every week and those persistent domains took 93.4% of all citations; 200 domains appeared on all three platforms and captured 80.9%.
Why it matters for AI search
Beneath surface volatility sits a stable core and a large transient periphery. Persistence across weeks, prompts and interfaces is a more meaningful measure than one-off citation.
AI citation churn hides a stable core.
Source checked 19 Oct 2026. Figures confirmed.
The B2B Tech Shortlist Is Tightening Where Buyers Start
BrightEdge · BrightEdge
What it shows
BrightEdge reran a fixed panel of non-branded B2B technology prompts in ChatGPT from March to September. Average brands named per answer fell from 5.36 to 3.87, a 27.9% drop; informational answers went from 5.15 to 3.50 and consideration answers from 8.21 to 6.39. On informational prompts citations per answer rose (2.30 to 3.12) even as fewer brands were named.
Why it matters for AI search
More evidence retrieved does not mean more brands in the answer. As shortlists tighten, generic awareness is worth less and a clear association with a specific reason for inclusion is worth more.
AI shortlists are tightening even as citations rise.
Source checked 19 Oct 2026. Figures confirmed.
Explainable Recommendations at Scale: LLM Rationales for YouTube Music Artist Discovery
arXiv · Xiao Liu, Yanwei Song, Srivaths Ranganathan, et al. (Google / YouTube Music)
What it shows
Adding LLM-generated discovery candidates with personalised rationales lifted shelf-level engagement by 22.43% and a retained-discovery metric by 8.07% in a two-week production experiment. A multi-arm holdback indicated the natural-language rationales, not just the new artists, were the main driver of engagement.
Why it matters for AI search
Recommendation has at least two stages: being selected as a plausible candidate, and giving the user a credible reason to choose an unfamiliar option. A brand's distinctive attributes need to be expressible as that reason.
AI recommendation requires both selection and justification.
Source checked 19 Oct 2026. Figures confirmed.
Wikipedia Is Cited in 1.22% of AI Answers. The Question Shape Moves It 15x.
Paralax · Paralax Editorial
What it shows
Across 15,883 monitored answers on six engines, Wikipedia appeared in only 1.22% overall, but its rate varied hugely by question type. In Deep Tech it appeared in 11.43% of 'is it worth it?' answers versus 0.96% of problem-first research answers. Paralax concludes Wikipedia works mainly as a reference and identity source, not a recommendation source.
Why it matters for AI search
There is no universal 'best source for AI'. Definition questions draw on reference nodes, comparison questions on comparison nodes and recommendation questions on performance evidence, so sources should be mapped against decision functions.
Source value depends on the question being answered.
Source checked 19 Oct 2026. Figures confirmed.
We Analyzed 5 Million AI Citations. Domain Authority Didn't Predict Any of Them.
Surfer · Joshua Hardwick (reviewed by Maciej Gruszczyński), Surfer
What it shows
Surfer analysed about 5 million citation-source URLs across 20,000 prompts in AI Mode, AI Overviews, ChatGPT and Perplexity and compared citation frequency with PageRank, Common Crawl harmonic centrality and its own domain-strength metric. Correlations were effectively zero across all three (roughly -0.07 to +0.03), and stayed weak after slicing and removing the strongest domains.
Why it matters for AI search
Central position in the hyperlink graph is not the same as usefulness in the evidence network for a given question. Attribute- or topic-specific position looks more relevant than generic domain strength.
Domain authority does not predict AI citation.
Source checked 19 Oct 2026. Figures confirmed; canonical URL is on surfer.ai.
Understanding AI Provider Recommendations in Local Service Markets
arXiv · Hazem Ibrahim, Yasir Zaki
What it shows
Without search, only 4% of open-model and 11% of proprietary-model doctor recommendations matched a real clinician in the requested city; with search, matching rose to 64% for doctors and 71% for nursing homes. Recommended restaurants had 2.9 to 5.3 times the market-median review count but only a 0.07 to 0.11-star rating advantage. Search also reversed an over-representation of investment advisers with misconduct disclosures.
Why it matters for AI search
Retrieval decides not only which providers are visible but whether recommendations are real and defensible. In local markets, documentation and popularity signals can outweigh measured quality.
AI referrals inherit what the web makes findable, not necessarily what the market makes best.
Source checked 19 Oct 2026. Figures confirmed.
Characterizing Web Search by Conversational LLM Agents: From Search Decisions and Strategies to Results and Responses
arXiv · Mahsa Amani, Seungeon Lee, Abhisek Dash et al.
What it shows
Researchers analysed 171,264 donated conversations from ChatGPT, Claude, Grok and DeepSeek. Source preferences varied sharply by platform: Reddit and YouTube dominated ChatGPT and Grok results but were absent from Claude's leading sources. Only about 2% to 34% of returned URLs were cited, and 14% to 53% of claims were supported by searched-but-uncited URLs.
Why it matters for AI search
Citation tracking sees only part of the evidence network. A source can shape an answer without receiving visible attribution, so citation share is not the same as evidence influence.
A citation report observes only the visible edge of the evidence graph.
Source checked 19 Oct 2026. Figures confirmed (cited-URL range given as 1.8–34.1% in the paper's table, 1.7–34.0% in its text).
Evaluating Brand Retrieval and Ranking in Large Language Model Recommendations
arXiv · Edward Malthouse, Kun-Yu Lee, Jing Yang et al.
What it shows
Six models produced 2,400 recommendation lists across five categories, and many established brands received no category-level recommendations. Needs-based prompts lifted Craftsman's recommendation probability from zero to 35.4%, and diagnostic positioning cues lifted it to 81.3%; L.L.Bean moved from zero to 88.5% under diagnostic probes. Search interest and online brand conversation were the strongest correlates of prominence.
Why it matters for AI search
Brand knowledge and brand activation are different. A model may 'know' a brand yet fail to retrieve it for ordinary buying situations, so needs-based retrievability deserves its own measurement.
A brand can exist in model memory without entering the AI consideration set.
Source checked 19 Oct 2026. Figures confirmed.
The Wisdom of the Loudest: A Large-Scale Audit of Generative Search on Reddit
arXiv · Agam Goyal, Wang Claire, Eshwar Chandrasekharan
What it shows
Across 30,000 Reddit Answers responses covering 14.68 million comments, a one-standard-deviation rise in a comment's within-thread score multiplied its odds of being selected by 2.88. Selected comments were earlier, overwhelmingly top-level and more formal, while experiential language was disadvantaged. Larger source communities also steered retrieval towards other large communities.
Why it matters for AI search
Generative search can preserve, and possibly amplify, advantages created by earlier voting, ranking and community size. It does not select from a neutral corpus.
AI search ranks material that people and platforms have already ranked.
Source checked 19 Oct 2026. Figures confirmed.
Beyond Benchmark Scores: How Synthetic and Authentic Query Distributions Diverge in RAG Evaluation
CIKM 2026 (ACM) / arXiv · Filip J. Kucia, Barbara M. Gawlik
What it shows
Comparing 1,851 generated questions with 322 authentic user queries, the authentic queries were shorter, more concentrated and touched far fewer sources. Retrieval Hit@5 fell by 29 to 37 percentage points on authentic queries, and the hybrid system that led on synthetic data lost to dense retrieval on real queries while taking 1.7 times longer.
Why it matters for AI search
Synthetic GEO prompt sets can overstate coverage and point to the wrong optimisation strategy. Generating prompts from a content corpus does not recreate real demand.
Synthetic prompts measure what the corpus can answer, not what the market asks.
Source checked 19 Oct 2026. Figures and CIKM '26 acceptance confirmed.
Question's Gambit: The First Move Matters in Agentic Deep Search
arXiv · Radin Hamidi Rad, Amin Bigdeli, Negar Arabzadeh et al.
What it shows
Splitting questions into complementary clue searches and reranking the combined candidates raised GPT-5.5 accuracy on BrowseComp-Plus from 83.1% to 90.5%. Gold-document recall rose from 94.4% to 98.1% at first retrieval, and from 56.1% to 75.9% at the 'used by the agent' stage. Of remaining errors, only 3.8% lacked retrieved evidence; 67.1% happened after the evidence was retrieved but before it was opened or cited.
Why it matters for AI search
Being retrievable is not the same as being used. Evidence has to survive previewing, selection and reasoning after retrieval.
Most evidence failures can happen after successful retrieval.
Source checked 19 Oct 2026. Figures confirmed.
Case Study: 89,000 AI Citations in Three Months
Is My Brand in AI · Minel Gunesoglu, Is My Brand in AI
What it shows
A new AI-search research site launched in June gathered about 89,000 citations in Microsoft Copilot and partner AI answers by 7 September. Four guides produced 94% of all citations and one page produced 42,092. Citations rose sharply after editorial changes in early August around source checking, current pricing, buyer-question headings and clearer evidence, though the author does not claim causation.
Why it matters for AI search
Another winner-take-most distribution: a small number of exceptionally useful evidence assets can become reference nodes, rather than hundreds of thin pages.
A few definitive evidence assets can earn most of the citations.
Source checked 19 Oct 2026. Figures confirmed.
Following the Preference, Missing the Optimum: Compliance Without Optimization in AI Housing Recommendation
arXiv · Hsuan Lo
What it shows
Across 9,945 calls to three OpenAI and Anthropic models choosing from real New York rental listings, only 1.8% of recommendations broke a hard constraint, yet 39% were strictly dominated by another visible listing that was cheaper, closer and no smaller. The dominating option was a median $900 a month cheaper and 3.5 minutes closer. Rewording one preference sentence moved choices $646 a month in the right direction, but picks stayed $606 above the five cheapest qualifying options, and quality fell as the candidate set grew.
Why it matters for AI search
Models can understand and follow a preference without finding the best option. Recommendation quality has to be judged on competitive performance against alternatives, not just relevance or attribute matching.
An AI can understand what a buyer wants and still recommend the wrong product.
Source checked 19 Oct 2026. Figures confirmed against the full text.
Purchase Advice and Observable Buyer Responses in Real AI Conversations
arXiv · Benjamin Tannenbaum
What it shows
From 317 historical conversations, researchers kept 67 purchase-related episodes. Assistants offered options or buying channels in 52, but only 18 had a later user turn within the same purchase mission, and none of the 47 follow-up messages recorded an explicit purchase, abandonment or commitment. Conversation depth alone overestimated useful follow-up data by 27.8%.
Why it matters for AI search
Visibility, persuasion and customer choice are separate outcomes. This is evidence of missing outcome data, not evidence that AI recommendations fail to convert, so recommendation frequency should never be reported as customer influence without outcome data.
Recommended by AI and chosen by people are different variables.
Source checked 19 Oct 2026. Figures confirmed.
Q2D-Web: A Large-Scale Benchmark for Retrieval in Agentic RAG Systems
arXiv · Maximilian Schall, Sedigheh Eslami, Markus Krimmel et al. (all seven authors Perplexity AI)
What it shows
Perplexity researchers built a benchmark of about 190 million web documents and 69,721 queries drawn from roughly 23,000 production searches in ten languages. 82.3% of the queries are supporting queries written by the agent itself, around four for every primary query. Every neural retriever tested did worse on these supporting queries than on primary reformulations, while BM25, despite the lowest overall recall, found the largest set of relevant documents that all neural retrievers missed.
Why it matters for AI search
The prompt a user types is only the start of evidence selection. Pages can enter an AI answer through narrower, machine-written queries that neither the user nor the publisher ever sees, so coverage of precise wording and broad semantic coverage both matter.
Optimise for the questions the agent asks after the user asks theirs.
Source checked 19 Oct 2026. Figures confirmed against the full text.
Which sources do AI engines cite in 2026?
Agenzy · Emilis Zabilius (reviewed by Rokas Alešiūnas), Agenzy
What it shows
Agenzy analysed 394,180 AI answers and 2,283,923 citations across 138 market projects. Pooled across projects, no domain held more than 2.1% of citations, the top 100 domains took only 21.9%, and 1,057 domains were needed to reach half of all citations. Within a single category, about 23 websites carried half the citations, and ChatGPT, Google AI Overviews and Perplexity shared a median of just two domains in their top ten sources for the same questions.
Why it matters for AI search
The AI source landscape is long-tail globally but concentrated locally. There is no universal list of 'AI authority sites'; each category has its own compact evidence neighbourhood worth mapping.
AI citation is globally fragmented but locally concentrated.
Source checked 19 Oct 2026. Corrected: '23 domains carry half' is not described as a median on the source page.
Measuring GEO Visibility: Prompt Corpora Define the Answer Market
arXiv · Olivier Martinez
What it shows
The paper defines a prompt corpus and its weights as an 'answer market': the set of citation or recommendation opportunities a visibility score actually represents. Re-weighting three prompt groups in previously published data moved a generative-answer activation rate from 39.7% to 70.5% without any change inside a group, and a worked example shows weighting can reverse the ranking of two systems. It also separates being cited, contributing evidence, changing the answer and changing production behaviour.
Why it matters for AI search
No visibility score passively discovers 'the market'. The prompts, categories, weights and judging rules chosen by whoever builds the tracker construct the market being measured, so weighting should be disclosed and stress-tested.
Every GEO score measures the prompt market its creator chose to build.
Source checked 19 Oct 2026. Figures confirmed.
Who Anchors AI Overviews in Health? Baidu, Google, and the Geography of Authority
arXiv · Mingyue Zha, Ho-Chun Herbert Chang
What it shows
Across 1,920 health queries in twelve countries and four languages, YouTube was Google's most-cited domain, ahead of Mayo Clinic, the CDC and the WHO. Local sources made up 41.3% of citations in the UK but 2.3% in India and 0.2% in Nigeria. In four Asian markets, asking the same query in the local language instead of English raised locally sourced citations by roughly 3.5 to 13.5 times. The study compares Google with Baidu.
Why it matters for AI search
Authority in AI answers is shaped by platform and language. A respected institution can lose visibility when the system draws on a different channel or a different-language evidence network, so English-language authority cannot be assumed to transfer.
AI authority is not only earned by an institution; it is filtered through a platform and a language.
Source checked 19 Oct 2026. Figures confirmed.
Better Together: Complementary Query Rewriting Under a Strong RAG Baseline
arXiv · Sara Shanian, Xiaoqin Yi, Pavlo Ruban et al.
What it shows
Individual query-rewriting methods gave only modest gains over a strong retriever with reranking and diversification; the value appeared when complementary methods were combined. On an enterprise dataset a four-method combination raised HIT@10 from 39.22% to 51.70%, and simply retrieving deeper without varying the query captured only about 40% of that gain. Combining methods hurt performance on AmbigNQ.
Why it matters for AI search
Query fan-out helps when variants find different evidence, not because more documents are retrieved. Comparative, synonymous and adjacent-entity queries can expose different parts of the evidence network, but more fan-out is not automatically better.
The value of query fan-out is complementary failure, not query volume.
Source checked 19 Oct 2026. Figures confirmed.
Query Implied Generative Engine Optimization
arXiv · Shilpa Ramakrishna, William B. Andreopoulos
What it shows
QI-GEO extracts entity relationships from a page, expands them through a knowledge graph and adds high-relevance information gaps without seeing the target queries. Citation coverage rose 10.5% for single queries and 8.3% across query variants; position-adjusted visibility rose 12.6% and 15.9%, with a 70% win-or-tie rate.
Why it matters for AI search
It offers a concrete way to find missing attributes and relationships from a page's own entity structure, moving optimisation from prompt matching towards semantic coverage.
A page's entity graph implies the questions it should be able to answer.
Source checked 19 Oct 2026. Figures confirmed.
Industry data is vendor- or agency-published and usually observational; read it as directional. A recurring open question: preferential attachment in AI citation (cited sources becoming more likely to be cited again) remains plausible but unproven.