What this means
Generation grounded in retrieved external information.
Historical record
xPRT research trains dense retrievers using relationships across passages
Retrieval quality can depend on whether embedding models learn relationships between passages within documents. This is an upstream determinant of what evidence a retrieval-grounded AI system may find.
Glance study tests generator-based reranking with shared candidate context
AI evidence selection can involve reranking after initial retrieval. Generator-based reranking could affect which retrieved passages inform an answer, independently of whether those passages were initially discovered.
Sequential-decision research shows search policy and retriever jointly shape evidence exposure
AI agents may expose or miss relevant evidence depending on query-branch decisions and allocation of attention across ranked lists, not just the rank of a webpage for a single query.
SearchHorizonBench separates answer success from the marginal value of another agent search
A relevant page may not enter an AI agent's evidence pool because the agent decides not to search again. Retrieval-trigger and stopping decisions must be measured separately from ranking and final citations.
ACARe research shows task context can improve reranking of agent-generated search queries
Query fan-out alone may not explain which documents are selected for agentic answers. Context-aware reranking adds a distinct source-selection stage that visibility measurement and retrieval experiments should account for.
How AI agents search: Peer-reviewed ASQ study maps agent query reformulation and retrieval workload
AI agents can issue numerous evolving retrieval queries for a single user information need. Measuring only the initial query can miss intermediary source exposure and retrieval workload. This evidence supports treating query fan-out, retrieved source exposure, and final citations as separate stages.
LASAGNA research finds isolated RAG component gains do not reliably combine
AI discovery and source-selection systems need end-to-end retrieval and faithfulness measurements rather than relying on individual component benchmarks or extrapolating citation performance from retrieval accuracy alone.
Research shows evolved semantic identifiers improve controlled generative retrieval recall
Semantic identifiers and collision handling influence which documents or products are retrieved in generative search/recommendation systems, independently of downstream answer wording or citations.
OpenDS-World benchmarks evidence discovery by agents in live web environments
AI discovery depends on whether agents can find, navigate and interpret source material before deciding what to cite or use in an answer.
Retrieved but Not Read study separates search-result retrieval from source reading and citation faithfulness
Source retrieval and even citation do not guarantee that an AI answer read or faithfully used the underlying evidence. This adds a measurable distinction between discovery, source opening, evidence use and answer support.
AURA research finds adaptive reasoning depth improves retrieval relevance and latency
The amount and direction of query reasoning can change which documents are selected for AI answer generation. This adds retrieval-utility and stopping-policy dimensions to source selection, query fan-out and AI discovery measurement.
SPRA research introduces source-linked structured paper objects for scientific agents
Agentic evidence discovery can depend on whether a system retrieves isolated passages or source-linked objects that preserve methods, findings and their conditions. The work supplies measurable evidence for structured scientific retrieval.
Glasseek research introduces auditable deep-search agent evaluation
Agentic search benchmarks can overstate retrieval capability if an agent encounters hidden answer keys or benchmark data while searching. The work highlights why evidence provenance, search trajectories and benchmark leakage must be audited before drawing conclusions about AI discovery.
SEEDS research separates retrieved candidates from evidence admitted into agent reasoning
An AI agent may retrieve a page yet never admit its full content into the reasoning context. AI discovery and citation measurement should distinguish retrieval exposure, preview inspection, evidence admission and final answer citations.
Perplexity releases multimodal late-interaction retrieval models
The work shows a concrete retrieval design that preserves token-level and visual evidence instead of compressing each document into one vector. For AI discoverability, source retrieval fidelity and final citation selection must be evaluated separately.
IndexAct research separates candidate-set refinement from text inspection in agentic search
Agentic discovery can depend on how an agent narrows candidate sources and when it chooses to inspect evidence, not only on which documents a search index initially matches.
T-Search research separates multi-step evidence retrieval from answer generation
AI discovery has an upstream evidence-retrieval stage distinct from final citation selection. Multi-round searches and alternative retrieval rollouts can change which source chunks are available to a generator; citation monitoring alone does not measure this upstream coverage.
Programmatic Search Agents demonstrate that evidence presentation matters beyond query reformulation
For AI discovery, a page entering retrieval is not equivalent to its supporting passage reaching the model or being used in an answer. The research supports separating retrieval, evidence delivery and final source use in visibility and attribution measurement.
Agentic retrieval study quantifies quality gains alongside steep latency and token costs
Source discovery in multi-step agents can differ from a single semantic similarity ranking, but additional query fan-out and reasoning impose material latency and cost. Evaluation of AI discovery pipelines should track evidence quality and search budget together.
RAAC study finds deep-research agents repeat searches after evidence discovery saturates
In agentic discovery, repeated search calls do not necessarily expose new documents. Retrieval stopping policies and marginal evidence coverage can affect which sources an agent sees before composing an answer.
Cloudflare makes AI Search generally available with native multimodal retrieval
The release provides a concrete implementation of how parsing, multimodal indexing, retrieval and reranking govern which evidence reaches downstream answer generation in developer-built systems.
DailyReport evaluates web research agents with task-level factuality and user-oriented rubrics
AI discovery measurement cannot stop at retrieved links or citation counts: agents must find, verify and synthesize evidence into responses satisfying the actual user task.
EMNLP study traces deep-research citation errors to agent orchestration
Citation visibility can fail after retrieval: evidence can be discovered and propagated between agents yet lose attribution during orchestration or synthesis. This makes citation survival distinct from initial source retrieval.
EBG study finds complete evidence retrieval and citation coverage are distinct RAG bottlenecks
AI discovery pipelines can retrieve a source without gathering enough passages to support a multi-part answer. Answer-only evaluation and raw citation counts can miss incomplete grounding.
EvidenT study tests source verification before enterprise RAG answer generation
AI answer citations can drift from retrieved evidence. Source-grounding verification is a distinct stage between retrieval and final citation display, relevant to trustworthy discovery measurement.
MCite-RL research improves visual citation grounding through iterative evidence refinement
A document or image can be retrieved without its specific supporting region being cited accurately. Multimodal AI discovery requires distinguishing retrieval, region selection, citation localization and answer support.
EMNLP study finds deep-research agents can mistake correlated pages for independent corroboration
Citation and source counts can exaggerate corroboration when multiple retrieved pages repeat a common origin. Agentic discovery measurement needs to distinguish independent evidence, copied evidence, and which documents the retriever actually exposed.
MiRAGE introduces claim-level information and citation metrics for multimodal RAG
As AI answers draw from video, image, audio and text, the number of citations or text overlap alone cannot establish whether a particular claim is supported by the cited media. Citation support and information coverage require separate measurement.
ReCite evaluates intent-aware retrieval and verification for faithful academic citations
Authentic and topically relevant retrieved sources can fail to support a particular claim. The work provides a testable approach to distinguishing candidate retrieval from faithful citation selection.
VHop-Router study moves multi-hop visual evidence search into the retriever
Visual and multimodal AI discovery can be limited by a retriever's ability to traverse evidence relationships, not just by the reasoning model. Evidence-chain completeness and search cost are distinct metrics from initial top-k matching.
Perplexity introduces Q2D-Web for large-scale retrieval evaluation in agentic RAG
The visible user prompt is not necessarily the query a source competes for. AI visibility therefore depends on several distinct stages: query reformulation, first-stage retrieval, ranking, evidence selection, and eventual citation.
ITER uses prior agent search history to improve iterative evidence retrieval
Multi-step agentic discovery is path-dependent: the same current subquery can be evaluated in light of earlier questions and visited evidence. Source visibility measurement should consider complete search trajectories.
SIEVE demonstrates structure-aware search, inspection and section-level fetching for research agents
Agentic discovery may depend on which parts of a source can be addressed and fetched, not just whether the page appears in a search index. Structured document access affects retrieval precision and context efficiency in the tested system.
EMBL AI Librarian demonstrates multi-query evidence retrieval in scientific research agents
This provides direct research evidence for separating multi-query retrieval, source selection and passage-level evidence grounding when analyzing AI discovery pipelines; it does not measure public-web publisher visibility.
SearchOS research shows explicit evidence coverage and shared search state improve benchmarked agent discovery
Long-horizon AI research agents may fail to expose sources or assemble grounded answers when they lose track of missing facts or repeat searches. The work gives a testable framework for measuring discovery completeness, fine-grained evidence provenance and search-state management beyond simple retrieval or citation counts.
NASA Earth Observation benchmark demonstrates hybrid retrieval and agentic reranking
Specialized resources such as scientific datasets can become more discoverable through structured knowledge-graph metadata, hybrid retrieval and task-aware reranking rather than simple keyword matching alone.
XBCP research finds language mismatch weakens agentic retrieval and citation fidelity
AI discovery of non-English sources can fail at multiple stages: retrieving relevant evidence and faithfully integrating that evidence into a cited answer. Visibility and citation metrics need language-aware, stage-specific evaluation.
RISE research tests bounded, navigable evidence workspaces for agentic search
For AI discovery research, the experiment distinguishes document candidate inclusion, access to evidence inside a retrieved document, and final answer generation. It demonstrates that evidence-delivery interfaces can influence agent effectiveness and cost in controlled tasks.
Mechanistic study separates citation-marker generation from evidence use in RAG
Citation presence or correctness is not equivalent to demonstrating that retrieved evidence causally informed the answer. This distinction matters for measuring AI visibility, source selection and grounded answer quality.
Deep literature search study separates citation coverage from topical relevance
AI discovery measurement should separate whether sources were found from how they were ranked and judged relevant; existing human citations are not a complete or unambiguous ground truth for evaluating retrieval.
LLM-Wiki research shows agent-guided traversal of linked evidence can improve multi-hop retrieval
The experiments distinguish first-pass document matching from agent-directed traversal of intermediate entities and evidence. This matters to AI discoverability measurement, source provenance and multi-hop information access, but does not establish a commercial AI search ranking or citation benefit.
Controlled SIGIR study isolates competitive first-citation selection across six LLMs
AI visibility analysis should distinguish retrieval and candidate ordering from subsequent citation selection; content relevance, price detail and recency can matter once a source is in the candidate context.
Research shows repeated UGC retrieval can contaminate deep-research citations
Multi-query agentic retrieval can repeatedly select the same editable sources. Citation occurrence alone is not evidence of factual reliability; provenance, moderation and source integrity affect trustworthy AI discovery.
SGR-Bench identifies website retrieval-state failures in AI search agents
AI discoverability can fail even when an agent finds the correct website: answer-bearing records may remain inaccessible or be interpreted under the wrong site-specific data context. This suggests measuring source discovery, state configuration, evidence extraction and final answer quality separately.
Google publishes official guidance for generative AI Search visibility
Practitioners gained first-party guidance separating verified requirements from speculative AEO/GEO tactics.
SIRA demonstrates corpus-aware one-shot retrieval as an alternative to repeated agent search
Agentic source discovery does not necessarily require more query fan-out. Discriminative vocabulary and corpus-aware retrieval controls may change the efficiency and source-selection behavior of future retrieval-augmented systems.
Direct Corpus Interaction research challenges fixed top-k retrieval for search agents
AI evidence discovery may depend on an agent's ability to inspect raw documents, combine constraints and revisit candidate sources, not only on similarity-ranking scores. This adds a distinct interface design to retrieval and source-selection research.
CoSearch jointly trains agent reasoning and evidence ranking for multi-step search
Intermediate agent-generated subqueries and the document admission/reranking stage can shape which evidence reaches a reasoning agent. This suggests retrieval and reranking should be evaluated separately from final citations.
MERRIN benchmark exposes multimodal evidence-selection and over-search failures in web agents
Multimodal AI discoverability requires that agents retrieve and correctly interpret the right source modality, not merely find a webpage. More query fan-out can introduce distracting evidence and impair final grounding.
Peer-reviewed study finds synthetic web content can amplify retrieval exposure
Source diversity, synthetic-content exposure, and retrieval integrity can degrade without immediate drops in answer accuracy, complicating the measurement of trustworthy AI discovery.
LRAT trains search retrievers on agent browsing and reasoning trajectories
Documents that help an agent resolve an intermediate information gap may differ from those most relevant to the initial user query; retriever training can therefore change which sources are surfaced to AI agents.
Controlled RAG study finds enhanced linked entity pages outperform plain HTML and JSON-LD alone
The study directly tests whether machine-readable entities and navigable relationships can improve evidence retrieval and answer completion. It distinguishes adding schema markup from providing richer linked entity pages.
MC-Search benchmark measures stepwise multimodal retrieval and agent planning
Visual and agentic AI discovery can require multiple dependent searches across text and images. Measurement of evidence discovery needs hop-level and modality-aware diagnostics, not just final-answer accuracy.
ProductResearch trains shopping research agents using supervised web and catalog trajectories
Shopping-agent discovery can involve both open-web source selection and structured catalog retrieval, with evidence coverage and product attributes shaping research reports rather than just first-pass item ranking.
KDD research finds text-only GEO edits can harm upstream retrieval in full-pipeline tests
GEO tests that hold retrieved sources fixed can miss upstream discoverability losses. Publishers and researchers should separately measure retrieval inclusion, reranking position and answer citation rather than assuming content already enters model context.
SAGE benchmark finds agent query wording changes retrieval effectiveness
Which sources enter an AI agent's evidence pool depends on its intermediate query formulations and retrieval configuration; semantic similarity alone does not predict exposure across tasks.
Large-scale agentic search logs reveal cross-step evidence reuse and intent-dependent query trajectories
Agentic discovery involves sequences of queries informed by prior retrieval, so measurement limited to the initial prompt or a single cited result can miss how sources enter later searches.
Agentic-R trains retrievers for multi-step search using local relevance and final-answer utility
In multi-step AI discovery, passages selected for semantic similarity to a subquery may not help complete the final answer. Retriever training objectives can therefore change which evidence reaches an agent, independently of traditional search ranking or publisher content changes.
Study compares language predictability and semantic similarity in generative-search citations
It adds original evidence that citation selection and answer synthesis need separate measurement from organic search rank and that source-text properties can correlate with generative citations.
Retrieval-Augmented Generation (RAG) is formalized in research
RAG is one of the core technical patterns behind modern AI search, source retrieval and citation behavior.