This month in AI: agents go Christmas shopping, ads flood AI answers, and the rollback list grows

The assistants went shopping. Meta's Muse passed 10 million users with shopping its biggest use, Anthropic gave Claude a wallet, and Shopify wired nine AI agents into its checkouts four weeks before Black Friday. The same month showed what happens when nobody checks the machines' work: voter guides built on an AI-written website, a ban on AI shop signage at Walmart, and a Big Mac priced $1.20 differently two miles apart.

The releases that matter

1. The first agentic Christmas is four weeks away

Shopping is now the biggest thing people do with Meta's assistant. Zuckerberg told investors on 21 October that more than 10 million people have set up Muse, with over a quarter of weekly users shopping through it at an average purchase above $200. Anthropic launched Claude Duet on 21 October, a US-only research preview that buys through Stripe Link and Shop Pay across more than 6,000 connected apps, and refuses a short list of actions like share trading even when the user approves. Shopify's 28 October edition wired nine agents into its stores, ChatGPT, Claude, Gemini and Grok among them, put Shop Pay inside personal agents, and shipped machine-readable storefront tools so agents can read policies and check out without scraping. The money is following: Adobe measured AI-referred retail visits up 71% year on year in September, spending markedly more per visit than ordinary traffic, and Shopify is calling this the first agentic holiday season. The caution came from Toronto, where Muse, left on always-allow, accepted a Marketplace offer and gave the buyer the seller's home address without asking.

2. Ads flooded AI answers, and organic is winning so far

Google expanded ads in AI Mode to ten more countries on 28 October, started testing ads inside follow-up questions in the US, and put the UK on the list for later this quarter. The detail that matters: existing Search, Shopping and Performance Max campaigns are automatically eligible, with no separate opt-out for the AI surface, so plenty of brands are already advertising inside AI answers without having chosen to. The early scorecard favours the unpaid route. One agency measured ChatGPT ad click-through at 0.91% against its 6.4% Google benchmark, and SE Ranking found roughly 14% of ChatGPT ads showing against unrelated prompts.

3. Chatbots built voter guides on an AI-written site, then dropped it when caught

The Washington Post asked six chatbots to build voter guides for an address in every US county ahead of the midterms and analysed more than 3,000 results. The bots couldn't name a congressional district 48% of the time and named the wrong one in 22% of the cases where they answered. Worse, several leaned on Ballot Brief, an obscure and largely AI-produced website whose mapping errors they repeated: Microsoft's Copilot cited it in three quarters of tests. The repair was as quick as the damage, though. After the Post flagged it, ChatGPT's use of the site collapsed to 4% in retests and Amazon paused Alexa's voter guides entirely. The same week, Wesleyan's researchers counted over $137m of AI-made midterm advertising with only 27% carrying a disclosure label, and Wikimedia reported unauthorised edits it believes came from OpenAI agents.

4. The rollback list: Walmart's signs, McDonald's prices, an NHS scribe

The list of companies walking back customer-facing AI got specific. Walmart banned AI-generated signage from its stores, telling staff that signs must come through approved corporate channels after AI-made deli menus and posters spread. A Reuters investigation found McDonald's AI pricing engine, used across nearly 14,000 US restaurants, produced a $5.69 Big Mac two miles from a $6.89 one, with franchisees describing pressure to adopt it; McDonald's says the tool suggests rather than sets prices, and a class action followed within weeks. In the UK, Healthwatch reported an NHS AI scribe dropped the word "null" from "null demyelination", turning a clean scan into a suggestion of nerve damage; it was caught because the patient happened to be an NHS clinician. And Reuters reported Meta cancelled the second wave of its AI-replacement layoffs after the agents underperformed and staff revolted.

5. Google pays its sources now, and Anthropic's are suing

Who each assistant can read is coming apart. Google's pilot programme now pays more than 400 publications to feed AI features in Google News, with CNN, the FT, Axios and Forbes joining in October on flat licensing deals and a separate programme covering AI Overviews and Gemini. The same month, four of Canada's biggest publishers, Postmedia, Torstar, The Globe and Mail and The Canadian Press, sued Anthropic in Ontario over training data, claiming up to C$20,000 per work across millions of articles; the same group already has a live case against OpenAI. One engine is signing its sources while another gets sued by them, and no two assistants end up reading the same web.

6. The money got picky

Anthropic filed for its stock market listing on 26 October under the ticker ANT. The headline number, an $80.7bn half-year loss, is less dramatic than it looks: $72.4bn of it is a non-cash accounting effect, the operating loss underneath was $693m, and figures shown to investors put third-quarter revenue at $16.8bn, up about 45% on the quarter. OpenAI cut the valuation on its latest raise from $1.4tn to $1.25tn after investor resistance, per the Financial Times. And the fix for expensive AI turned out to be cheaper AI: legal AI firm Harvey, whose margins reportedly went sharply negative as agent usage exploded, built its own model on an open-weight base and says costs recovered, while AT&T now runs about 40% of its AI workloads on open models.

The undercurrent

Smaller signals worth filing away:

• Free ChatGPT quietly moved to GPT-6 in early October: the model answering most of your customers changed. Any AI-visibility baseline from the summer is stale.
• Shopify shipped shop.md: a file any AI agent can read to search over a billion products, pay with Shop Pay and track orders. A robots.txt for shopping is forming, and it covers brands off Shopify too.
• AI found two 14-year-old flaws in PHP, the language behind WordPress and most of the web's content systems. Councils including Edinburgh and Oxford took sites offline, and roughly nine in ten measurable hosts were still unpatched in late October.

What to do this month

Concrete, in order of urgency:

1. Get agent-ready before Black Friday. Clean product feeds, schema mark-up and live stock data are what shopping agents read; Shopify merchants should switch on the Fall edition's agent connectors. AI-referred visitors spent markedly more per visit in Adobe's September data, so this is the highest-intent traffic of the season.
2. Check your Google campaigns for AI Mode. Ads expanded to ten more countries on 28 October with existing campaigns auto-enrolled and the UK weeks away. Know where your spend is showing before the paid lane's weak early benchmarks spend it for you.
3. Re-baseline what AI says about you. Free ChatGPT switched to GPT-6 in October, so answers recorded in the summer no longer describe what customers hear. Run a fresh report and compare.
4. Patch anything running PHP this week. That includes most WordPress sites. US authorities confirm active exploitation, and a site that's offline or hijacked is invisible to customers and AI crawlers alike.