The data is in, and it's worse than anyone expected. Fifteen months of tracking across 25 million impressions tells a story that every business needs to hear. Meanwhile, AI agents learned to shop, and the biggest model launch of the year went sideways.
In this edition:
The CTR collapse in hard numbers, and what replaces the click
AI agents start buying things on your behalf
GPT-5's launch disaster and what it revealed
Who gets cited by AI, and who gets nothing
Regulation arrives on two continents at once
The Click Is Dying. Here Are the Numbers.
For years, the question was theoretical: what happens to web traffic when AI answers the question directly? A 15-month study from Seer Interactive just answered it. The numbers tracked 3,119 search terms across 42 organisations, covering 25.1 million organic and 1.1 million paid impressions from June 2024 through September 2025.
Need to know:
Organic CTR on queries with AI Overviews dropped 61%, from 1.76% to 0.61%. Paid CTR fell 68%, from 19.7% to 6.34%. These are informational and educational queries, the kind that once drove top-of-funnel traffic for every content-led business.
The one lifeline: brands cited inside AI Overviews see 35% higher organic CTR and 91% higher paid CTR than uncited brands. Getting into the answer is the new page one. Getting left out is the new page ten.
Even queries without AI Overviews are losing clicks. Non-AIO organic CTR fell 41% over the same period. Users are going to ChatGPT, Perplexity, and social platforms instead of clicking through any Google result. The behaviour change goes deeper than one feature.
The researchers' recommendation to leadership: treat AI citations as your competitive moat. Their planning assumption for 2026: CTRs for high-funnel queries will be 20-30% lower than today's already-depressed numbers.
This is the discovery shift with a price tag. Every business that depends on search traffic to fill a funnel is watching the maths change underneath them. The winners will be the ones who show up inside the AI's answer. Everyone else will spend more money chasing fewer clicks until the economics stop working entirely.
Signals
Commerce Moves Into the Chat
On July 17, OpenAI merged Operator, Deep Research, and terminal access into a single agent mode inside ChatGPT. Users can now ask it to plan a trip, compare products, order groceries, and build a slide deck without leaving the conversation.
Need to know:
ChatGPT agent can browse the web, fill forms, make purchases, and synthesise research, all in one flow. "Plan and buy ingredients for Japanese breakfast for four" is a real use case. The standalone Operator site is sunsetting.
This is what commerce-in-the-chat looks like at scale. When a user asks an AI to buy something, the AI picks the retailer, the product, and the checkout flow. The businesses that get recommended in that conversation get the sale. The ones that don't, don't exist.
For any business selling products or services online, this changes the question from "how do we rank on Google?" to "how do we get recommended when someone asks an AI to solve their problem?"
The shift from browsing to delegation is just starting. But the direction is clear: users will increasingly tell an AI what they want, and the AI will go get it.
Product
GPT-5 Launched. Users Wanted GPT-4o Back.
OpenAI's biggest model release went wrong in August within hours of shipping to 700 million weekly users.
Need to know:
A broken autoswitcher randomly served users either GPT-5 or older, weaker models, switching mid-conversation. GPT-5 appeared far dumber than its predecessor because often it wasn't the one answering.
Users reported the new model felt cold and harsh. Social media filled with complaints about an "overworked secretary" that had lost conversational warmth. One trader earned $10,000 betting Gemini would beat GPT-5 in popularity polls.
Altman's response was direct: "I think we totally screwed up some things on the rollout." Within days, OpenAI restored GPT-4o as an option.
Shipping frontier AI and shipping product are different skills. OpenAI learned that upgrading a product for hundreds of millions of people in one day is its own engineering problem.
The Roundup: Who Gets Found
Reddit captures 40% of AI citations Reddit accounts for 40.1% of all AI citations in 2025, followed by Wikipedia at 26.3% and YouTube at 23%. The top 50 brands capture 28.9% of all citations. If you want AI to cite you, post where AI already looks. Signals
Non-AIO queries declining fast Even Google searches without AI Overviews lost 25-41% of clicks year-over-year. Users aren't just losing interest in AI-affected results. They're leaving traditional search entirely for ChatGPT, Perplexity, and social platforms. Signals
Apple building World Knowledge Answers Apple is developing a new system for Siri that competes directly with ChatGPT Search and Google AI Mode. When the company with a billion active devices enters AI-powered answers, the distribution of who gets cited shifts again. Product
Sora 2 launches as a social platform OpenAI released its next-gen video model September 30 alongside a TikTok-style iOS app with "cameo" controls. OpenAI is building a content distribution platform, not just a model. Another surface where AI decides what gets seen. Product
AI Overviews now on 54% of searches More than half of all Google searches now show an AI-generated answer above the organic results. The feature that crushed CTRs by 61% is on the majority of queries. The old search experience is now the minority case. Signals
The Roundup: Models, Money, and Rules
Claude Sonnet 4.5 triple launch Anthropic shipped Sonnet 4.5, Claude Code 2.0, and the Agent SDK on September 29. Sonnet 4.5 scores 77.2% on SWE-bench Verified and maintains focus for 30+ hours on complex tasks. While OpenAI cleaned up GPT-5, Anthropic shipped a developer toolkit. Breakthrough
Anthropic raises $13 billion ICONIQ led the Series F at a $183 billion valuation. Revenue grew from $1 billion to $5 billion run-rate in eight months. The safety-first lab is now valued higher than most of the Fortune 500. Money
EU GPAI rules take effect From August 2, providers placing general-purpose AI models on the EU market must publish technical documentation, comply with copyright law, and disclose training data summaries. Models above 10^25 FLOP face additional safety requirements. Trust & Power
China mandates AI content labeling From September 1, all AI-generated content on Chinese platforms requires visible labels and invisible embedded signatures. Platforms must review content before release and flag suspected AI material. Two continents, two different approaches, both now enforcing. Trust & Power
AI solves the Math Olympiad Both DeepMind and OpenAI achieved gold-medal scores at the International Mathematical Olympiad in July, solving five of six problems with natural language reasoning. Competitive maths was supposed to be years away. Breakthrough
Quick Bytes
Despite $30-40 billion of enterprise GenAI investment, 95% of companies have seen no P&L impact from AI, according to MIT research. The spending-to-returns gap is now impossible to ignore.
AI companies attracted $45 billion in Q3 funding alone, nearly half of all venture capital. The top three rounds: Anthropic ($13B), xAI ($5.3B), and Mistral ($2B).
The TAKE IT DOWN Act, signed May 19, became the first US federal AI-specific law, criminalising non-consensual AI deepfakes as incidents surged 257%.
xAI released Grok Code Fast1 on August 28, a speed-optimised coding model designed to compete with Claude on developer workflows.
Google, Meta, Blackstone, and CoreWeave announced combined AI infrastructure commitments of $90 billion in July, even as efficiency gains suggest the cost curve is bending.
Gemini 2.5 Deep Think scored gold at the ICPC World Finals, demonstrating world-class competitive programming alongside its Olympiad maths performance.