Latest AI search research

Google Cloud published reference implementations including an enrichment agent for producing OKF documents from BigQuery datasets with citations and schema documentation.

In 224 of 323 citations (69%), Google cited a brand's self-promotional listicle but excluded that brand from AI Overview recommendations, instead recommending competitors.

Google states that converting HTML to text for LLM consumption is trivial, contradicting the premise that markdown simplification is necessary for AI crawling.

92% of ChatGPT citations are not cited by Claude for the same query, indicating significant algorithmic divergence between AI systems.

Reddit is cited with authoritative medical sources 36% of the time in ChatGPT but only 6% of the time in Google AI Overviews, indicating AI engines assign different contextual roles to the same source.

Meta launched AI Mode in Facebook Search that generates answers powered by Meta AI and uses public social content from Groups, Reels, and other Meta apps as sources.

Fable 5 achieved the highest score among frontier models on Cognition's FrontierCode evaluation for difficult coding tasks.

Expert survey of ~50 local SEO professionals found that complete and current business information is among the most important signals for local search rankings.

Google confirmed that hyphenated domain names carry no negative SEO signal, and domain names can contain up to 61 hyphens.

300 marketing executives from large enterprises were surveyed regarding AI search adoption and organizational response strategies.

Social search adoption among Americans increased from 30% in 2020 to 35%, representing a five-percentage-point shift in information discovery behavior.

Reddit was cited alongside medical authority sites 36% of the time in ChatGPT but only 6% in Google AI Overviews, representing a 6x difference in how the same source is contextually positioned.

SISTRIX visibility analysis across US and UK markets (May 26 - June 2) showed sites matching query intent and result type gained visibility while mismatched sources lost ground.

Only 2.2% of citations remain stable across three identical ChatGPT prompt runs from a dataset of 815,000 prompt-page pairs.

Google rolled out a new Generative AI report section in Search Console showing impressions data for content appearing in AI Overviews and AI features, with filtering capabilities by pages, countries, devices, and dates.

Reddit's top 3 SERP share increased to 10.24% after May 2026 core update, up from 8.56% after March and 9.19% after December, across 100,000 tracked keywords.

Brands with presence on four or more platforms showed a 2.8x increase in likelihood of appearing in ChatGPT responses compared to those with fewer platform mentions.

Google officially stated that third-party SEO tools are not endorsed by Google and lack access to Google's internal ranking data.

61% of AI users cite speed and ease as primary reasons for using AI tools.

Google announced that users can follow creator profiles to increase visibility of that creator's content in the Discover feed.

Study used three AI-detection tools to classify content across a 6-year publication span from January 2020 to March 2026

Traditional search outpaces AI tool growth in absolute volume despite faster growth rates for AI tools.

AI Overviews compress time-on-SERP behavior across search intents: without AIOs, navigational vs. local searches show 20-point variance (12% vs. 32%), but with AIOs all five intent types cluster between 41.9-48.5% at 21 seconds.

AI-referred traffic converted 42% better than non-AI traffic in March 2026, representing a sign flip from worst-performing to best-performing channel in U.S. retail within 12 months.

Google CEO Sundar Pichai acknowledged that AI Overviews require further tuning and refinement in how they answer certain queries, as some responses contain excessive opinion.

Google rolled out new AI performance insights in Merchant Center to track product performance across AI-powered shopping experiences, including share of voice benchmarking, funnel performance metrics, conversational query insights, and product attribute completeness tracking.

Adding JSON-LD schema to 1,885 pages resulted in a 4.6% decline in AI Overview citations relative to control pages, with AI Mode and ChatGPT showing statistically insignificant changes of +2.4% and +2.2% respectively.

Google Research developed ALDRIFT framework that refines generative models toward lower-cost answers while maintaining likelihood under the model.

Search results layout has changed significantly over seven to eight years, with AI overviews and commerce links now dominating above organic results.

A controlled experiment tracking AI bot visits to webpages with LLMs.txt files over a 3-month period found zero AI bot visits to the test pages.

Web Bot Auth is in experimental phase with limited rollout; not all Google user agents currently implement the protocol.

AI search engines collectively drive approximately 1% of publisher traffic.

Citation overlap of sources varies from 16-59% across AI engine pairs, but brand overlap remains consistently narrow at 35-55%, indicating AI engines cite different sources but recommend the same brands.

GPT-5.5 achieved an 84.9% win-or-tie rate against human professionals across 44 tested occupations in the GDPval benchmark.

Year-over-year LLM referral traffic increased 72.7%, with ChatGPT referrals growing 66.7% and Claude referrals growing 23x.

Real-time answer fetching accounted for 56.9% of crawler activity (primarily ChatGPT), training 28.8%, and discovery/indexing 14.3%.

Reddit accounts for 6.6% of Perplexity citations versus 1.8% for ChatGPT across 680 million citations analyzed — platform-specific citation preferences differ substantially, requiring off-site presence across multiple platform types for cross-platform AI visibility.

86% of AI citations across ChatGPT, Gemini, and Perplexity come from brand-managed or brand-influenced sources: 44% from brand websites, 42% from listings/directories, and 8% from reviews/social, based on 6.8 million citations from 1.6 million queries.

YouTube receives 200x more AI citations than any competing video platform (Vimeo, TikTok, and Dailymotion each below 0.1%), accounting for 20% average citation share and 29.5% of all Google AI Overview citations — ranking as the #1 cited domain in AI Overviews.

AI brand recommendations are highly inconsistent: fewer than 1-in-100 chance ChatGPT or Google AI returns the same brand list across two runs of the same prompt, based on 2,961 prompt tests with 600 volunteers across 12 categories.

Government (.gov) websites represent 6% of sources linked in Google AI Overviews versus 2% in standard search results — a 3x overrepresentation — based on analysis of 68,879 real searches by 900 US adults, indicating AI Overviews heavily favor authoritative government domains.

Perplexity failed to read JavaScript-rendered pages, returning Access Denied errors, while server-side rendered comparison pages were successfully parsed — demonstrating consistent AI crawler inability to process client-side rendered content.

Bing-indexed content is included in Bing Chat (Copilot) answers by default; sites must use the NOARCHIVE meta tag to opt out — confirming that Bing indexing status directly controls inclusion in Copilot AI answers.

Microsoft 365 Copilot uses the Bing search service for web grounding; URLs not indexed by Bing cannot be used as knowledge sources — making Bing indexing a hard prerequisite for Copilot visibility.

ChatGPT Search for Enterprise and Edu workspaces uses Bing as its sole third-party search provider — confirming Bing indexing is a prerequisite for ChatGPT visibility in enterprise contexts.

Anthropic's ClaudeBot crawls web content for potential use in AI model training; site owners can block it via robots.txt using the 'ClaudeBot' user agent string, with IP verification available at Anthropic's published endpoint.

Perplexity officially confirms PerplexityBot surfaces and links websites in Perplexity search results and is not used for AI foundation model training — distinct from training crawlers.

OpenAI operates two separate crawlers: GPTBot for training data and OAI-SearchBot for ChatGPT search results — each requires independent robots.txt configuration; allowing one does not enable the other.

Google officially recommends including author name and a unique author URL as Article structured data properties, with author.url specified as 'a link to a web page that uniquely identifies the author.'

Core Web Vitals are a Google ranking signal but act as a floor rather than a ceiling: Google will surface sub-par page experience content if it is the most relevant result available.

Google's core ranking systems require LCP under 2.5 seconds, INP under 200ms, and CLS under 0.1 as Core Web Vitals thresholds; achieving these is described as necessary for success in Google Search including AI features.

48% of AI citations originate from community platforms (Reddit, YouTube), and 85% of brand mentions in AI answers come from third-party pages rather than brand-owned domains.

Consistently pairing a brand name with a descriptive expertise label ('unoccupied home insurance experts') across BBC News and The Times coverage produced a 194% increase in GEO traffic and 382% rise in ChatGPT referral visits — demonstrating the context wrapping technique.

Microsoft reported a 357% year-over-year increase in AI referrals to top websites in June 2025, totaling 1.13 billion visits.

Brands cited in Google AI Overviews receive 120% more organic clicks per impression than non-cited brands on the same AIO-present SERP (2.07% vs 0.94% organic CTR for informational queries), based on 5.47 million queries and 2.43 billion impressions across 53 brands.

Person schema appears in 58.9% of AI-cited pages (70.4% for ChatGPT specifically), making it the most prevalent schema type in citations, likely because it signals author authority to AI systems.

ClaudeBot crawls 20,583 pages for every single referral it sends back — compared to Google's 5:1 ratio — meaning AI crawlers consume content heavily but return almost no traffic, making citation-based visibility the only meaningful outcome of AI search indexing.

SE Ranking's XGBoost model trained on 300,000 domains found that removing llms.txt as a variable improved prediction accuracy — indicating llms.txt has no measurable positive effect on AI citation frequency as of late 2025.

Comprehensive content (10,000+ words, Flesch readability score ~55) received 62x more AI citations than thin equivalents (187 vs 3 citations) — indicating content depth and readable comprehensiveness are major citation drivers.