The chat window is turning into a billboard. This week we learned that advertisers can point ChatGPT ads at specific people, using a list of phone numbers or email addresses, and that OpenAI also matches ads to what users type. Meanwhile Meta quietly gave itself permission to read your shop's website, and Google built a shopping cart that lives above every store. Four stories today. The first one is the one you can't ignore.
ChatGPT ads can now be aimed at your exact customer list
404 Media ran its own tests and found that anyone buying ChatGPT ads can upload email addresses, phone numbers or phone device IDs and show ads to those people inside their ChatGPT sessions. On top of that, ad selection looks at the context of the conversation itself. Ask about running shoes, get a running shoe ad. Users on the Free and Go plans see ads; paid plans don't.
If this sounds like Facebook's custom audiences, that's because it is the same idea, moved into the place where people now ask what to buy. And it's growing fast. The Information reports OpenAI went from roughly 20,000 advertisers in late August to about 60,000 now, after cutting prices. One agency told the outlet that clicks for some clients cost around $5 on ChatGPT versus $30 to $50 on Google. That's one agency and one kind of client, so don't treat it as a rule.
Why it matters to you: we wrote last month about ChatGPT's ad cookie following users to store websites. Put that next to customer-list targeting and the picture is clear. The AI assistant your customer asks for advice is becoming a paid shelf, and the brands with the best customer data will rent the best spot on it. Your bigger competitors are already uploading their lists.
My take: you don't need to buy these ads to feel them. A customer who arrives in your DMs may have just been shown three rivals by a chatbot. That makes the first reply you send more important, not less.
Meta's new terms let it read, and train on, your shop's website
Here's the quiet one. Meta's updated commercial terms take effect on October 30. Nieman Lab had three legal experts read them, and their reading is blunt: if you keep a business account on Facebook, Instagram or WhatsApp, you agree that Meta can crawl websites "associated with" that account and use what it collects to train its products. The experts said they found no opt-out. One attorney said it's unclear whether a court would enforce this, but businesses should assume it would.
Nieman Lab wrote it up as a news publisher problem. It's just as much a seller problem. Your product descriptions, your size guides, your carefully written FAQ page: if your Instagram business profile links to your site, those pages are now fair game under Meta's reading of its terms.
What to do this month:
- Read the update in Meta Business Suite, where Meta says it posted the notice in early October.
- Know that blocking one crawler isn't enough. Meta runs several crawler names, and blocking one doesn't necessarily stop the others.
- Decide what's public on purpose. Supplier prices, wholesale terms and anything you'd hate to see in an AI answer belong behind a login, not on an open page.
I wouldn't delete your business account over this. For most sellers, Instagram is the shop. But walk into October 30 knowing what you're signing.
Google ships Universal Cart: one basket for every store
On October 15 Google turned on Universal Cart in the US, inside Search and the Gemini app. It's a single cart that holds products from many different retailers. It watches for price drops and restocks, finds promo codes, and suggests which card to pay with. At big partner retailers like Walmart, Target and Sephora, shoppers can pay right there with Google Pay. Everywhere else, the cart hands them off to the store's own checkout. Google previewed it at I/O in May and says Canada, Australia and the UK come next.
Iranian shops aren't in this, and won't be soon. So why care? Because it shows where shopping is going. The cart used to belong to the store. Now it belongs to the assistant, and the assistant's favourite trick is comparing your price with everyone else's and pinging the shopper when someone is cheaper.
The defence is the same everywhere: be the seller who is easy to buy from. That means accurate stock, clear prices, and product data an AI can actually read. If your own assistant still guesses about sizes and colours, start with training it on your real product catalog. Clean data is what every shopping agent, Google's or yours, runs on.
Meta opens up Muse Spark, a giant model anyone can download
Also on October 15, Meta released the weights of Muse Spark Open, a version of the Muse Spark model it has kept closed since April. The numbers are big: 816 billion parameters, with only 42 billion working on each word, and room for up to a million tokens of context. It handles text, images, video and audio. The weights are on Hugging Face and in the model catalogs of Amazon, Google and Microsoft.
Two catches. First, it's "open weights," not open source. It ships under Meta's own Muse license, which limits commercial use for larger companies. Second, you won't run this on a laptop. Reports say two top-end Mac Studios can manage it, which tells you who it's for: developers and companies, not shop owners.
Why sellers should still notice: every time a strong model becomes downloadable, the tools built on top of it get cheaper and easier to run locally. That pressure lands, eventually, in the price of the AI tools you pay for.
Quick takes
- ChatGPT's voice got new voices and better accents in French, German, Hindi, Mexican Spanish, Korean and Japanese. OpenAI says the API version comes later. Persian isn't on the list yet, so don't plan around it.
- OpenAI and Meta told US judges that deepfakes can't be reliably caught. An OpenAI lawyer said she doesn't expect any method to spot every AI-made image. For sellers: a screenshot or a voice note "from the bank" proves less every month. Confirm payments in your actual account.
- The White House's new superintelligence task force started its 120-day clock, with a report due on cyber risk, biosecurity and model control. Big-picture, slow-moving, worth one eye.
- Microsoft finished training MAI-Base-2, a huge in-house model it says beat open models in its private tests. More big players means more competition on price, which is good news for anyone paying for AI.
What this week means for your shop
The thread running through all four stories: the big platforms want to stand between you and your customer. ChatGPT wants to show them ads before they reach you. Meta wants to learn from your website. Google wants to hold the cart. None of that is going to stop.
What you still own is the moment a customer actually writes to you. Answer fast, answer correctly, and that conversation is yours. Remember that the five-minute reply window decides more sales than any ad slot.
That's the part Vardast is built for: an assistant that answers your customers on Instagram, WhatsApp, Telegram and your website, from your own product data, any hour they write.