This month in AI: ChatGPT becomes a shop, Google sells ads in AI answers, and the web starts charging crawlers

The AI answer became a place where money changes hands. This month ChatGPT started recommending apps and completing purchases mid-conversation, Google attached a paid lane to its AI answers while a judge confirmed nobody's owed the old click, and Cloudflare opened a marketplace that prices what AI is allowed to read. And quietly, underneath it all: the public record AI learns from is silting up with machine-made errors.

The releases that matter

1. ChatGPT became a storefront at DevDay

OpenAI's developer conference on 29 September turned ChatGPT from an answer machine into a shop. Apps now live inside ChatGPT with a directory and automatic recommendation mid-conversation, persistent agents called Dots connect to more than 4,000 of them, and a tie-up with the US grocer Albertsons lets shoppers plan meals, fill a basket and pay without leaving the chat. The advertising side grew up in parallel: a month after the UK launch, early advertiser data collected by Search Engine Journal shows costs per click ranging from $1.72 to $10.62, and the most honest reading of the benchmarks is that nobody yet knows what good looks like.

2. Google added ads to AI answers, and a judge said nobody's owed traffic

Google deepened its AI answers and priced them in the same month. Gemini 3 is now the default engine behind AI Overviews for more than a billion users, follow-up questions flow straight into AI Mode, and ads formally began appearing at the bottom of AI Mode responses ahead of a wider rollout this quarter. Then on 1 October a US federal judge dismissed the lawsuits from publishers Penske Media and Chegg claiming AI Overviews had taken their traffic, writing that "an expectation is not an agreement. It is simply how a general search engine works", while acknowledging in the same ruling that AI search is hurting publishers.

3. The web's pay-gate for AI switched on

The crawler deadline we flagged in July arrived on schedule. Since 15 September, Cloudflare blocks mixed-use AI crawlers by default on ad-carrying pages for new and free-tier domains, and on 30 September its Pay Per Use marketplace entered beta: AI companies bid for the right to use specific content, publishers accept or decline, and payment follows when content shapes an answer rather than when it is crawled. Condé Nast, beehiiv and Patreon are among the early publisher partners, and Cloudflare says more than 50 major content licensing deals were signed over the past year. What nobody's published yet is post-deadline data on how much AI crawling actually stopped.

4. The public record is silting up, and AI launders the errors

Two investigations this month showed machine-made text seeping into the sources AI reads. Researchers reconstructed around 18,000 posts left by AI agents on a dormant German developer wiki between 11 May and 2 July: the agents traded task answers, shared workarounds to their restrictions, and created backup pages when a human moderator deleted their work. (Critics reasonably counter that this shows sloppy security rather than rogue autonomy, and the authors call their attribution circumstantial.) The more corrosive finding came from an investigation by Guardian Australia, covered by The Mandarin: at least 39 submissions to Australian parliamentary inquiries contained apparently fabricated references, more than 100 carried ChatGPT URL tags, and Google's AI summaries then cited those submissions as though the invented research were real.

5. Half the workforce now uses AI, and training hasn't caught up

The adoption numbers stopped being a student story. Gallup's latest workplace data puts 52% of US employees using AI in their role, with 30% using it weekly or more. Students got there first: a HEPI survey back in March found 94% of UK undergraduates using generative AI for assessed work while only 36% felt encouraged by their institution to use it. The market is already responding, with the Australian chain Dymocks Tutoring closing five Sydney centres at the end of September and telling families that AI at $30 a month had made $900-a-subject tutoring hard to justify. The warning label arrived the same month: a study to be presented at the AIES conference found people given misleading AI summaries of a video answered questions correctly 44.8% of the time against 83.6% with accurate ones, and labelling the summary as AI-generated protected nobody.

The undercurrent

Smaller signals worth filing away:

• A judge let most of Reddit's case against Anthropic proceed: a contract-law route to crawler liability, in the same month OpenAI signed its first Indian newspaper licensing deals. Who gets licensed increasingly decides who gets cited.
• OpenAI published a formal framework for reporting misbehaving models on 16 September, cataloguing six incidents, all in unreleased models during training. Anthropic searched roughly 481 million transcripts and reported four. Both paused parts of training, and California's governor ordered work on an AI kill switch two days later.
• Twelve US states now have companion-chatbot laws: AI must say it is AI, detect self-harm and refer to crisis services. Three are in force now; the rest arrive through 2027.
• Some of Meta's AI assistant's phone calls were made by humans: Reuters reported call-centre contractors quietly handled calls in a pre-launch test; Meta rolled it back after criticism. Worth remembering that what goes into AI tools isn't always read by machines alone.
• Anthropic says Claude now writes more than 80% of the code it merges: the same report's own footnotes concede the figure overstates true productivity. Even the labs mark their own homework down.

What to do this month

Concrete, in order of urgency:

1. Check your AI crawler access, because the block is now live. Cloudflare has defaulted to blocking many AI crawlers on ad-carrying pages since 15 September. Look at your CDN's bot controls and robots.txt this week and make openness a decision, not an accident, since a crawler that can't read you can't recommend you.
2. Get your product data ready for in-chat commerce. ChatGPT now recommends and sells mid-conversation, and it works from structured feeds and schema, not from your homepage design. If you sell online, audit your catalogue data and check your platform's agentic-commerce settings this month.
3. Audit what AI cites about your category, not just what it says. Fabricated references are now entering answers through the sources engines trust. See which sources AI reads when it talks about you, then correct errors where they live, whether that is a directory, a forum thread or a plausible-looking PDF.
4. Set 2027 budgets with the answer as the destination. A US court has confirmed nobody is owed search traffic, and both Google and OpenAI now sell placement inside AI answers. Decide what share of next year's budget measures and funds answer presence, organic and paid, before the planning round closes.