Issue 12 — 23/06/2026

AIO Impact on Google CTR: 2026 Update — CTR Decline Levels Off Across 53 Brands

Seer Interactive · 24/04/2026

A full-year analysis of 5.47M tracked queries and 2.43 billion organic impressions across 53 brands found that the AI Overview-driven CTR decline, which had been forecast to continue into 2026, instead levelled off in Q1 2026 actuals. Per the study, Q4 2025 data had projected continued decline, but early 2026 figures reversed that trend. Practitioners should treat this as a directional signal rather than a confirmed inflection point, and continue monitoring CTR at the query-intent level rather than drawing brand-level conclusions.

What It Takes to Rank in Google's AI Overviews in 2026: Correlation Study Across 8,500 Keywords

Seer Interactive · 28/05/2026

Across a stratified sample of 8,500 keywords covering 30 industries and 9 intent types, Seer found that AI Overviews now appear on nearly 65% of question-based searches and that the first citation slot correlates with standard SEO authority signals — consistent with Google's own published guidance on generative AI features. The study also found that paid search spend, TikTok presence, and SERP features showed no statistically significant correlation with AI Overview citation. Practitioners can use this to deprioritise speculative GEO tactics in favour of domain authority and content-quality fundamentals.

2026 AI Search Traffic Report: ChatGPT's B2B Referral Share Falls from 89% to 63% in Eight Months

Goodie · 21/05/2026

Wave 2 of a longitudinal AI referral study — drawing on GA4 data from an anonymised brand panel, triangulated against SimilarWeb traffic across 25.77 billion visits (January–April 2026) and SensorTower iOS rankings — found ChatGPT's share of measurable B2B AI referrals at 62.6%, down from 89.1% in May–August 2025. Claude reached 18.5%, Gemini 10.6%, and Perplexity 7.3% in the same period. Practitioners running GEO programmes optimised solely around ChatGPT citation should audit their prompt-sampling coverage to include Claude and Gemini, which are now material referral sources.

Ahrefs Updates AI Overview Citation Study: 38% of Citations Now Pull From Top-10 Organic Results

Ahrefs · 23/06/2026

Analysing 863K keyword SERPs and 4M AI Overview URLs — more than double the scale of Ahrefs's 2025 study — the updated research finds that 38% of AI Overview citations come from pages ranking in Google's organic top 10. The corollary is that 62% of citations go to pages outside the top 10, a finding consistent with independent data cited elsewhere suggesting AI retrieval runs a materially different selection pipeline from standard ranking. Practitioners should not assume organic rank is sufficient for citation eligibility, and should audit citation-earning pages for structural and authority signals independent of their SERP position.

Conductor Seven-Month Study: Each AI Engine Has a Distinct and Persistent Source Preference by Intent

Conductor · 07/05/2026

Tracking citation behaviour across ChatGPT, ChatGPT Search, Perplexity, Google AI Overviews, Google AI Mode, Gemini, and Claude from September 2025 through March 2026 (1,056 data points), Conductor found that each engine maintains a consistent editorial source preference by intent type: Perplexity and Gemini favour YouTube across most intents, ChatGPT and ChatGPT Search are the only engines surfacing Wikipedia, and Google AI Mode routes users back to Google properties. A single unified AEO content strategy cannot efficiently cover the full engine ecosystem. Practitioners should map content formats and source types to specific engines rather than optimising for a generic 'AI search' target.

Microsoft MAI-Thinking-1 Preprint Reveals 24.2 Billion Common Crawl Pages in Training Corpus

TechTimes / The Decoder · 05/06/2026

At Build 2026, Microsoft AI CEO Mustafa Suleyman described MAI-Thinking-1 as trained exclusively on 'enterprise-grade, clean and commercially licensed data.' The model's own 109-page technical preprint, first surfaced by developer Simon Willison, describes a pipeline that includes Common Crawl — an open web-scrape repository with no licensing guarantees — comprising 24.2 billion pages after filtering and deduplication. The gap between the marketing claim and the technical disclosure is confirmed. Practitioners advising enterprise clients on AI-generated content provenance should treat vendor training-data claims as unverified until corroborated by published preprints.

Perplexity Integrates Deep Research Into Computer, Routing Across 20+ AI Models

OpenTools · 17/06/2026

Perplexity has moved its Deep Research capability into its Computer product, routing complex research tasks across more than 20 AI models. The architectural shift means Perplexity's research-mode answers are now synthesised from a broader and more dynamic model pool. For GEO practitioners, this increases the citation surface area — content must be structured for retrieval across heterogeneous model stacks, not a single underlying model's preferences.

Perplexity Architecture: End-to-End Retrieval Pipeline Built Around the Sonar Model

The AI Engineer (Substack) · 13/06/2026

A detailed technical breakdown of Perplexity's search architecture describes a pipeline comprising a proprietary crawler, an index of hundreds of billions of pages, a hybrid ranking funnel terminating in cross-encoder rerankers, and the Sonar model — which generates answers with a citation on every sentence and is restricted to content it has retrieved. Per the author, a 2025 study of 576,000 code samples found nearly one in five package recommendations from memory-based models were hallucinated — a failure mode Perplexity's retrieval-first design is explicitly built to avoid. Practitioners optimising for Perplexity citation should prioritise passage-level extractability: the Sonar model cites what it retrieved, not what it inferred.

MAI-Thinking-1 Dataset Breakdown: 30 Trillion Pre-Training Tokens, 54.6% Code, No HuggingFace Data

Kili Technology · 05/06/2026

An independent review of Microsoft's 109-page MAI-Thinking-1 technical report documents a pre-training corpus of 30 trillion tokens (54.6% code), with no open-source or HuggingFace datasets included. The report names every tool used in the pipeline but discloses no data vendors — including those supplying human preference data used in safety alignment. The 'no synthetic data' claim holds for pre-training but breaks in reinforcement learning, where both SWE problems and tool-use environments are synthesised. Practitioners building content strategies around AI training-data provenance should note that 'no synthetic data' is a scope-limited claim requiring context to interpret accurately.

The Training Data Problem: Quality Signals Diverge From Volume as AI Bottleneck Shifts

Swarm Signal · 07/02/2026

As high-quality human-generated text becomes harder to source and synthetic content spreads across the web, the field faces increasing pressure on training data quality rather than data volume. The analysis argues that what models learn from is now the binding constraint — ahead of architecture and compute. For GEO practitioners, this reinforces the strategic value of publishing original research and primary-sourced content: as synthetic web text proliferates, human-authored, verifiable content carries a proportionally larger signal weight in future training corpora.

Knowledge Graph Enhanced Memory-Augmented Retrieval for Long Context Modeling (KGERMAR)

ACL ARR 2026 March Submission (authors anonymised at review) · https://openreview.net/forum?id=qCI2vlnLSI · 01/03/2026

(Pre-publication / ACL ARR under review) KGERMAR constructs dynamic, context-specific knowledge graphs from input text at inference time, integrating graph-structural embeddings with textual semantics across three memory banks — contextual, semantic, and structural — fused via learned weights. For GEO practitioners, the framework demonstrates that entity clarity and explicit relationship tagging in published content can improve retrieval precision in long-context modelling pipelines, providing further empirical grounding for the structured-entity markup recommendations already circulating in the field.

CacheRAG: A Semantic Caching System for Retrieval-Augmented Generation in Knowledge Graph Question Answering

Yushi Sun, Lei Chen · https://arxiv.gg/abs/2604.26176 · 30/04/2026

(Pre-publication / arXiv) CacheRAG addresses a structural flaw in LLM-based knowledge graph question answering: existing systems act as stateless planners, regenerating retrieval plans from scratch for every query — leading to schema hallucinations and limited coverage. The proposed architecture transforms the retrieval planner into a continual learner by caching semantic query patterns. For GEO practitioners, the paper reinforces the signal that consistent, schema-conformant content — particularly content using structured data and explicit entity definitions — is more reliably retrieved across repeated query patterns than unstructured prose.