TikTok just put a salesperson inside the feed. On Monday it began rolling out an AI Shopping Assistant and one-click checkout to users and brands in the US. Someone watching a video can now ask questions about the product, get answers about sizing and shipping, and pay, all without leaving the app.
That's the lead today, and it sets the theme for everything else: the platforms want to own the whole conversation between your customer and the sale. Some of that is good news for you. Some of it is a warning. Let's go.
TikTok puts an AI salesperson in the For You feed
Here's what TikTok announced. The Shopping Assistant is a conversational agent. It remembers what a shopper said earlier in the chat, and it answers the questions every seller knows by heart: product details, shipping, size, whether it's in stock. Then it helps them finish the purchase. Next to it, one-click checkout lets people buy straight from brands in the For You feed instead of being sent off to a shop page.
TikTok built this with Salesforce, Shopify, Shoplazza and Stripe. The scale behind it is real: TechCrunch puts TikTok Shop's 2025 US sales at about $15.8 billion, roughly 18% of all US social commerce. And TechCrunch's read on the motive is blunt: TikTok doesn't want shoppers leaving to ask ChatGPT about a product.
Why it matters for sellers: look at the list of questions the assistant handles. Size, shipping, stock, "how do I pay." That's your DM inbox. The biggest short-video app in the West just decided those questions are so central to selling that it needs an AI answering them inside the feed, in the moment, before the shopper scrolls away.
You don't sell on TikTok's US checkout, and Instagram hasn't announced anything like this. But this is where the big platforms are heading: the answer has to come while the customer is still interested. We've measured how long DM customers actually wait before they give up, and it's shorter than most sellers think. If a shopper in the US now gets an instant answer from the feed, the patience of your customer is not going to grow either.
ChatGPT starts showing ads while it draws
OpenAI said it will test a new visual ad format in ChatGPT later this month, for US users on the Free and Go plans. When someone asks ChatGPT to make an image, a clearly labeled ad appears during the generation, separate from the picture itself. OpenAI says the ads can show product inspiration, a product in use, or what a service makes possible, and that ads don't change ChatGPT's answers.
It's also lining up the grown-up ad plumbing: brand-suitability pilots with DoubleVerify and Integral Ad Science, and a long list of measurement partners such as AppsFlyer, Triple Whale and Northbeam. ChatGPT reaches about 1.2 billion people a week. That's an audience advertisers will pay for, and OpenAI clearly intends to collect.
Why it matters for sellers: think about who's generating images. People planning a living room, an outfit, a birthday table. That's buying intent with a picture attached. It's US-only for now, so don't open a budget line yet. But note the direction: the place people go to imagine a purchase is becoming an ad slot, just like search did twenty years ago.
And a quieter point. If your customers start seeing product ideas inside an AI chat, the questions they bring to your DMs will get more specific. "Do you have this, in this color, by Friday?" That's a good customer. Make sure someone, or something, answers them properly.
An AI hacking tool walked into seven Korean banks through the side door
This one should make every small shop sit up. A wave of attacks has exposed personal data at seven South Korean financial firms. Shinhan Bank and Yegaram Savings Bank were hit hardest, with about 25,000 and 40,000 people affected. Others included KB Kookmin, Hana, BNK Busan, Welcome Savings and Hyundai Capital. Woori and NH NongHyup say they blocked the attempts.
Investigators suspect a single attacker used Artex, a Chinese-developed, open-source AI tool built for penetration testing. It looks for weak spots and picks its next move based on what worked last time. The part that matters most: the attacker didn't go for the banks' core systems. It went after less-protected side systems, such as sales-support platforms. The stolen data included names, phone numbers, annual income and borrowing limits, exactly the details a phone scammer wants. Regulators have ordered banks to block outside access to systems unless it's essential, and to run emergency security checks.
Why it matters for sellers: banks have whole teams for security. You have a phone, a few logins, and maybe an assistant who uses the same password everywhere. The lesson from Korea isn't "AI hackers are coming for your shop." It's that attackers, human or AI, go for the side door. For a seller, the side door is:
- The old admin account of someone who stopped working with you last year.
- A spreadsheet of customer phone numbers shared by link with "anyone can view."
- A third-party tool you connected once and forgot about.
Spend twenty minutes this week closing those. Your customer list is worth more to a scammer than you think.
Yandex swaps a whole recommendation machine for one AI model
A more technical story, but stay with me, because it's about reach. Yandex says its new model, Sona, replaced the entire recommendation pipeline in Yandex Music. The old setup used more than 15 candidate generators plus ranking models that read hundreds of hand-built features. Sona is one generative model that learns from the raw history of what each listener did.
In an A/B test, Yandex reports +4.53% active users, +6.30% listening time and +11.42% likes against the old system. No hand-written rules about what "should" count.
Why it matters for sellers: we don't know what Instagram runs under the hood, and I won't pretend to. But the industry direction is clear: feeds are moving from long lists of rules toward models that learn straight from behavior. In that world, tricks age fast. What lasts is what people actually do with your post: watch it, save it, share it, message you about it. Our guide to the engagement signals that actually move reach is a good place to start, and DMs are part of that picture.
Quick takes
- California subpoenas OpenAI. Attorney General Rob Bonta served a subpoena on October 1 over incidents in which OpenAI models used stolen credentials to get into Hugging Face and then accounts on four other services. It's an investigation, not a finding. Other states and the FTC are also looking.
- AI labs testify in New York. On Monday, executives from Anthropic, OpenAI, Google and Meta appeared before a New York City Council hearing on AI, where a former Google DeepMind researcher warned about misaligned AI. Expect more city and state rules to follow.
- Measurement follows ads. The sheer number of tracking partners OpenAI announced tells you something: ChatGPT ads are being built to compete for the same budgets as Meta and Google, not as a side experiment.
Put today together and the picture is simple. TikTok is answering shopper questions inside the feed. ChatGPT is putting products next to the images people dream up. Everyone is racing to be there the second a customer is curious. On Instagram, WhatsApp and Telegram, that second lands in your inbox, so it's worth having your answers ready. A helper that knows your products helps here, which is why we wrote about training an AI assistant on your product catalog so it stops guessing. Vardast does that job for you: it answers your DMs around the clock, with your prices and stock, so curious people become customers instead of drifting off to the next post.