OpenAI put Sponsored Agents inside ChatGPT ads and wired HubSpot and Shopify into the buy flow#
Sponsored Agents beta: a ChatGPT user who clicks an ad can start a conversation with an agent the business sponsors. OpenAI began testing with a group of US advertisers on September 16. TNW
“A clearly labeled conversation”: the thread is distinct from ChatGPT’s independent answers and separate from the conversation the user started. Users can ask follow-ups, then follow a link to the advertiser’s site. TNW
Beta eligibility: participation can carry conditions such as creative automation and automatic bidding, plus a minimum daily ad spend for 28 days. OpenAI
Ads Manager plugin: advertisers create campaigns, update them and analyse performance with plain-language prompts. The tool reads a landing page and a campaign objective, suggests copy and images, and leaves the edits to the advertiser. OpenAI · Reuters
AI copy customisation: existing headlines and descriptions get rewritten for the context of a conversation and translated into the user’s preferred language. TNW
HubSpot is the first CRM partner, Shopify the first ecommerce partner: from September 16, a HubSpot account can hold a ChatGPT Ads account, so campaign creation and performance reporting sit next to lead follow-up in one place. TNW
The Shopify app and the market schedule: US merchants run ads straight from products already connected through Shopify Catalog; the app reaches other ChatGPT Ads markets on September 23. TNW
How the channel got here: cost-per-click pricing in April, a $1bn annualised run rate on August 31, and an Amazon advertiser pilot on September 11. TNW
When the click lands in a conversation instead of a landing page, your product’s front door moves#
Attaching a channel got cheaper: a Shopify seller has the catalog connected already, so the app is the whole integration. Nobody rebuilds a stack to test one more channel.
Copy stops being the bottleneck: drafts of headlines and images arrive from the landing page, which moves the work from writing to reviewing and measuring.
The first screen becomes the first answer: if a conversation opens after the click, the hero headline is not what decides the conversion.
Tools you already run become ad surfaces: campaigns and reporting inside HubSpot and Shopify mean the team already living there never leaves it to buy ads.
Product data an agent can answer from turns into an asset: an agent that reads clean product data answers better, so specs and shipping details decide how the ad performs.
What to try now: design the conversation that sits behind the click#
Build one agent that answers for your product: load product data and your most common questions, then run it. What the ad promises but the agent cannot answer shows up fast.
Run generated copy against your own: take the variants that come out of one landing page and put them next to yours, then compare click-through and conversion.
Fix three questions you will always get: pick the ones buyers ask every time. Dimensions and compatibility are the usual ones, so put the answers in the agent.
Test a conversational landing page: point one ad at a screen that opens with a few questions and another at the current page, then compare where people drop off.
The line between a labeled ad and ChatGPT’s own answer is still untested#
The label may not carry the trust: a clearly marked sponsored thread can still be read as ChatGPT’s own answer.
Performance data is thin: where a conversation ends and whether a lead justifies the spend needs advertiser data that does not exist yet.
Click prices can climb: when advertisers pile into a new surface, cost per click rises and the cheap window closes.
The schedule outside the US is unsettled: the Shopify app arrives in other markets on September 23, so sellers elsewhere should check their own date.
Anthropic merged Claude chat and Cowork into one interface: documents and slide creation came with it, rolling out to Pro and Max first. TechCrunch
Zed opened the Delta public beta: a multiplayer environment for coding with agents and reviewing their work; its own repository runs without pull requests. Zed
Google Home opened MCP server early access: agents such as Claude and ChatGPT can control smart home devices and read event history. TechCrunch
Mozilla’s Firefox Smart Window beta runs on Mistral models: France and North America first, with the UK and Germany later this year. Mistral
Arcee AI closed a Series B above a $1bn valuation: it says the open-weight Trinity family cost about $20m to build. TNW
Cohere and Aleph Alpha signed a definitive combination agreement: the combined company keeps the Cohere name, two headquarters and more than 1,000 staff. TNW
Google DeepMind launched the DeepMind Institute: Shane Legg, Demis Hassabis and James Manyika direct it, and Legg called an AGI declaration premature. TNW
The EU Commission president will invite the frontier labs in: Ursula von der Leyen said she wants to support industry efforts to pace the frontier. Politico
The EU competition chief would entertain an antitrust exemption for safety coordination: Teresa Ribera said no formal request has arrived. Politico
California made AI ad disclosure law: Governor Gavin Newsom signed SB 1050, which requires explicit labeling when AI-generated performers appear in video or audio ads. Governor of California
Bipartisan AI safety talks picked up in the Senate: Ted Cruz wants a Commerce Committee markup this month, and Amy Klobuchar is negotiating a bill on model approvals and imminent-harm alerts. Semafor
Vishal Sikka’s Hang Ten Systems added $53M to its seed: five weeks after the first $32M, for $85M total. TechCrunch