What happened

Dots are new AI helpers from OpenAI, the company behind ChatGPT. You hand one a task and it goes online, finds information and fills in forms to get it done. Even if you close the conversation, it keeps working in the background, and it comes back only when something needs your decision. AI assistants like this, which go online for people to look things up, compare prices and place orders, are what the industry calls AI agents. OpenAI unveiled dots on 29 September 2026 at DevDay, its annual developer conference in San Francisco.

OpenAI calls dots “always-on agents”, meaning they don’t stop when the conversation does. Each dot (that is, each of these helpers) has its own cloud computer, a machine set up for it online rather than the one on your desk, and its own browser. It can also connect to the apps you authorise, more than 4,000 of them according to the company. When you’re not talking to it, it looks things up in the background, which OpenAI calls “proactive research”, but in that mode it can only look, not act: it can’t send messages, change content or control the browser.

Most of OpenAI’s examples are about work. One early tester’s dot noticed he’d forgotten to invoice a publication, prepared the invoice and sent it once he approved. During the keynote, booking a plane ticket also came up as one of the small jobs a dot can take care of along the way.

Why it matters to you

For the past couple of years people have used AI as question and answer: they ask, the AI answers, and they click through to whatever sites the answer mentions. Dots turn that into delegation. You hand off a task and go do something else, while the agent opens the browser, finds the information, fills in the forms and reports back with a result.

The tech newsletter Platformer gave a concrete example from testing. An insurance broker had sent its author a form asking how many floors his new office had and whether it had sprinklers, and he didn’t know the answers. His dot dug out the lease, looked things up on city websites, checked the budget documents and came back to him with a few questions. In the end only two questions were left blank, and he spent that whole time on his own work.

Now picture it in your business. Say you run a packaging company. A buyer tells their agent: “Find three packaging suppliers that can deliver next month and take small orders, and list their contact details and terms.” Then they head into a meeting. The agent goes through the suppliers’ websites one by one, so the first one to read your site isn’t a person, and all the buyer ever sees is the summary the agent hands back.

Suppose your website gives the lead time only as “call us for details” and doesn’t mention a minimum order. A buyer reading it might pick up the phone. But the agent is there to compile information, and on your pages it can’t find a lead time or terms. Shops selling to ordinary customers are in the same boat: if someone asks an agent for a birthday cake that can be delivered this weekend, and your website doesn’t say where you deliver or the last day to order, the agent can’t find that either. Going by how OpenAI describes dots, our inference is that a site like that is likely to be passed over before the agent hands back its summary, and the person who sent it never even sees your name.

Nor does it run just once. OpenAI’s own suggestions include having a dot go through its user’s calendar every morning and run regular checks. Follow that design through, and an agent keeping an eye on packaging suppliers for a buyer will come back to the same set of sites again and again, and the one it can’t read gets passed over again on every round.

What to do about it

Start preparing, but there’s no need to panic. Not many people can use dots yet: only two paid ChatGPT plans, Pro and Business Premium, have them, and access is still arriving in stages; Enterprise customers can try a beta, but only once a company admin switches it on. Taiwan is among the places they’re available.

But OpenAI says it plans to bring dots to more users soon. It has also opened its tools for building AI agents to other companies’ developers, and has now added the ability for AI to operate software by itself, so others can build agents that operate websites too. The agents visiting your site won’t all be dots.

The trouble is that the way you check your own site isn’t the way an agent uses it. When you click around, you bring a person’s habits: you can guess which page the information is on, a vaguely labelled button still makes sense to you, and if nothing looks wrong you move on. An agent arrives with an errand to finish, and the points where it can’t read a page or gets stuck won’t show up however many times you browse like a person. OpenAI hasn’t published any rules for how agents choose sites, either; beyond dots, every company’s agents work their own way, and all of them keep getting updated.

So your site needs to keep adjusting to how these machines read websites. If you work through a checklist once and call it done, your site will fall behind as soon as these AIs change how they work. That ongoing adjustment is something you can hand to us: our team fixes your site’s content, brings what other sites say about you into line, and checks regularly how each AI describes you. You just review the results. We call this our managed GEO service.

Whether you make it onto the list an agent hands back starts with your own website. Send us your URL and we’ll set up a demo: a time with one of our consultants where we show you everything live. We’ll start by running a GEO audit. That’s geoweb.tw’s free website audit, which scores your site across 12 areas on how well AI can read it, and the consultant will walk you through what the score means. Then we’ll test ChatGPT live with a few questions that don’t mention your company, to see whether it thinks of you on its own: you’ll see on the spot who it recommends and whether you’re among them.

Further reading