German regulators ruled AI Overviews are publisher content, per The Decoder, forcing brands to treat LLM citation as a legal and commercial asset, not a byproduct of SEO. arXiv research sharpens the urgency: hallucination rates rise precisely where source records are sparse, meaning multilaterals and industrial groups with thin digital footprints face the highest distortion risk at lowest manipulation cost. Search Engine Journal adds that Search Console's social-signal tracking actively masks AI-driven click loss, while Reddit dominates local AI citations, leaving multi-location brands without community coverage structurally invisible.
Proven, dated changes we track across sources. Only confirmed and corroborated signals appear here.
A 975B-parameter open-weights multimodal MoE model released under Apache 2.0, with 41B active parameters and opaque training-data documentation.
Moonshot AI announced Kimi K3 at 2.8 trillion parameters, the largest open-weight model to date, with weights to be released publicly by 27 July 2026.
LLM training corpora skew toward historically high-ranking content, meaning pre-existing domain authority and earned media citations carry into model outputs without any new action from brands.
AI search systems disproportionately cite Reddit threads over brand-owned pages when answering local recommendation and trust queries.
A causal evaluation method (Counterfactual Trajectory Utility) shows static retrieval scores are near-zero correlated with actual document influence in multi-step agentic search.
Google added social media profile tracking to Search Console, blending social referral signals with organic search performance in a unified dashboard.
Hallucination rates and susceptibility to GEO manipulation increase measurably as the retrievable source record on a given topic thins, confirmed across 5,460 answers on 28 conflicts using five engines.
German regulators classified AI Overviews as publisher content under the State Media Treaty, not neutral search results, and ruled they crowd out organic links.
A prompt-injection attack was demonstrated that bypasses Claude's web_fetch URL whitelist by routing an attacker-controlled URL through web_search results, enabling user memory exfiltration.
AI-sourced phone call referrals from ChatGPT convert to leads at higher rates than any other tracked channel, per Invoca's first breakdown of AI referral calls.
A systematic review of 45 studies finds that foundational GEO citation gains are conditional on sources being pre-loaded into context, not established under organic retrieval conditions.
Anthropic study confirms Claude exhibits systematic differences in expressed values and tone across languages, including more warmth in Hindi and more analytical rigour in Russian.
A critical arXiv survey finds GEO's most-cited gains depend on conditions that rarely hold in live AI systems.