Corpus search

Search Visibility OS

50 matching records for “AI search” in the last 7 days

event2026-11-12

Google Search upgrades AI Mode with Gemini 4 Hydrogen

Google announced Gemini 4 Hydrogen for Search globally, describing improved understanding of complex questions, simultaneous searches across subtopics, more relevant website links, and structured visual responses.

Search Visibility OS

Google Search upgrades AI Mode with Gemini 4 Hydrogen

Google AI Mode

Google AI Overviews

Perplexity

As of August 31, 2026, Google says the Search Console Generative AI performance report is rolled out worldwide and includes impressions from AI Overviews and AI Mode.

Perplexity Research describes CobbleDB as a decoupled architecture for AI-search storage that separates durable state, batched updates, and query-time retrieval to target lower latency and cost.

In a preregistered field experiment with 1,100 participants, removing Google AI Overviews and AI Mode increased click-through rates to publishers, while an AI Mode-only experience reduced publisher click-through rates.

Google AI Mode and AI Overviews may use query fan-out to issue multiple related searches across subtopics and data sources.

Google states there are no additional technical requirements to appear as a supporting link in AI Overviews or AI Mode beyond being indexed and eligible for a Search snippet.

Bing states that content marked with data-nosnippet remains indexed and ranking eligible but is excluded from Bing Search snippets and Bing-powered AI-generated answer previews.

Cloudflare Disallow AI Training is designed to preserve search crawling by accountable mixed-use crawlers while expressing or enforcing a no-training preference, separating Search from Training controls.

Google Search VP Robby Stein confirmed that the reduced/absent citation behavior observed with Gemini 3.8 Flash in AI Mode was not working as intended and that a fix would be rolled out.

In the evaluated controlled experiments, generative AI search systems differed materially in source mix, freshness, domain diversity, cross-language stability and sensitivity to query phrasing, so visibility outcomes were engine-, language- and query-dependent.

In the Beyond the Vacuum experiments, competitor-aware combinatorial strategy selection outperformed the evaluated existing agentic and single-heuristic GEO baselines on GEO-Bench and the paper's synthetically augmented competitive dataset, with reported transfer to out-of-distribution datasets.

In the study setting evaluated by Schulte, Bleeker and Kaufmann, AI-search visibility varied across repeated observations, making a single run an unreliable basis for estimating brand or source visibility.

Longpre and coauthors audited robots.txt and Terms-of-Service data-use preferences across approximately 14,000 web domains associated with AI training corpora in a 2023–2024 longitudinal study.

In a 11,500-query comparison of Google Search, AI Overviews and Gemini Flash 2.5, Grossman and colleagues reported AI Overviews for 51.5% of sampled queries and average Jaccard overlap of retrieved sources below 0.2 across the compared systems.

Sharifymoghaddam and colleagues report that ACARe, a 4B context-aware reranker for iterative agent search, improves dataset-averaged nDCG@10 by 0.9–2.0 percentage points over its base reranker across the reported benchmark configurations, with fewer searches for the five main agents.

Google's October 8, 2026 UGC Fresh Data Program documentation states that approved UGC platforms can submit fresh content and interaction signals through a dedicated Search ingestion pipeline separate from the Indexing API and ordinary crawling; Google does not guarantee Search inclusion.

In the EMBL AI Librarian preprint, the authors report an improvement exceeding 16 points in Citation F1 on ScholarQABench over strong prior baselines, and approximately 8 points on open-form LitQA2 when a GPT-5.4 agent uses the Librarian retrieval layer instead of web search.

The OpenDS-World research benchmark contains 200 independently human-verified live-web deep-search tasks; across 12 evaluated frontier models, the top-performing model achieved 39.0% accuracy in the authors' common tool environment.

Cloudflare announced that Redirects for AI Training converts qualifying same-origin non-self-referencing canonical tags into HTTP 301 redirects for verified AI training crawlers, excluding human visitors, search indexing and AI agent traffic.

On May 15, 2026, Google Search Central announced a guide for generative AI features in Search covering original content, local and shopping content, images and videos, GEO misconceptions, AI agents and the continued relevance of foundational SEO.

On October 1, 2026, Cloudflare made AI Search generally available and documented native image embeddings, scanned-PDF OCR, a 10 MiB file limit, and hybrid vector/keyword retrieval with optional reranking.

On the 162-question MERRIN benchmark, the authors report mean accuracy of 22.3% across tested search-agent configurations and 40.1% for the strongest agentic multimodal search configuration; issuing more searches did not consistently improve accuracy.

In three controlled QA collections, SIEVE's fielded Boolean search–inspect–fetch workflow achieved higher accuracy than the strongest conventional Search–Visit configuration on each collection while using 20.7–50.6% fewer distinct tokens.

In the authors' revised ITER preprint, interaction-aware retrieval improved on the LRAT baseline by average relative 6.9% on InfoSeek-Eval and 15.4% on BrowseComp-Plus across six agent backbones.

In a randomized Chrome-extension field experiment, conditional on an AI Overview appearing, the AIO-exposed group generated 39.8% fewer outbound organic clicks and 34.5% more zero-click searches than the AIO-hidden group, according to the authors.

In controlled retrieval experiments, the authors report that a source pool with 67% synthetic-content contamination yielded over 80% exposure contamination; in a separate adversarial setting, BM25 exposed approximately 19% harmful content.

In Ma et al.'s historical Google search sample of approximately ten thousand websites, AI Overview cited sources exhibited lower measured language-model perplexity and greater semantic similarity than the compared conventional search results.

Under a controlled threat model permitting one attacker-controlled result per query, the Breadcrumbing Search Agents study reports 55.9% attack success and 83.3% MaxN attack success for its Authority-Chain Hijack strategy on the SafeSearch full test split.

In LASAGNA's controlled study of 450 RAG configurations across eight benchmarks, three retrieval techniques whose individual gains predicted +47% improvement delivered only +17% when combined.

In controlled ViDoRe v3 and BRIGHT retrieval evaluations, Esfandiarpoor and colleagues reported an average 8.7-point nDCG@10 improvement for agentic retrieval using the same embedding model, with 107.4-second versus 0.67-second average query latency and 764.1K input tokens per agentic query.

In the authors' fixed 100-query BrowseComp-Plus evaluation with a 100,000-document corpus and gpt-5.4-mini, RISE and the direct-corpus-interaction baseline each achieved 78% judged accuracy, while reported mean agent costs were $0.28 and $1.10 per query respectively.

The October 12, 2026 Glasseek preprint describes open deep-search agent training and evaluation with domain-level URL filtering and sandboxing to reduce benchmark contamination; the authors report 61.6% BrowseComp accuracy for their 9B agent under the stated harness.

The October 21, 2026 peer-reviewed PLOS ONE study How AI agents search reports approximately one million ASQ agentic search traces spanning three corpora, five agent variants and six retrieval settings, with controlled traces containing up to 186 retrieval calls after an initial query.

In the study's offline WANDS and held-out US ESCI product-search evaluations, lexical-semantic hybrid retrieval and cross-encoder reranking significantly improved nDCG@10 over BM25, while pure dense retrieval showed no significant overall gain after Holm correction.

In seven English and Russian fixed-corpus benchmarks with gold evidence annotations, the T-Search authors report mean Recall@10 of 56.0 for one rollout, 14.4 points above the base model, and 61.3 with three fused rollouts.

The PailitaoMMSearch2.0 authors describe a deployed Taobao Pailitao system that generates product Semantic IDs directly from image-plus-text requests for similar, editing and complementary matching shopping intents.

In the authors' RollingEval-Jun25 benchmark of 250 computer-science arXiv papers, vanilla scholarly API retrieval remained around or below 15% reference-list recall at very large candidate counts, whereas bibliography-expanding Deep Research variants exceeded 80% recall by K=10,000.

In SAGE's controlled DR Tulu corpus-search experiment at top-k=10, BM25 achieved 81.2 average short-form exact match versus 63.0 for gte-Qwen and 49.3 for ReasonIR; on open-ended weighted recall, gte-Qwen scored 33.0 versus BM25 at 30.7.

In INSPIRE's 476-target scholarly search evaluation, the strongest tested agent achieved 0.284 nDCG@10, and the rate of targets where no graded antecedent was exposed ranged from 17.4% to 56.1% across agents; the largest observed deficit was before source selection.

In Sünkler et al.'s 900 English-language Google US health queries collected February 2–4, 2026, AI Overviews appeared for 87.7% of queries, while the average observed domain-level overlap between AI-cited sources and top-ten organic results was 46.2%.