GPT-6 Astra: what actually changes for digital marketing

OpenAI announced GPT-6 Astra on 3 September 2026, with a phased rollout that began with a small set of organisations and extends to ChatGPT's Plus, Pro, Business and Enterprise plans, as well as to its API and to AWS.

You are going to read a lot of headlines over the next few days. This piece is not about those. It is about what changes, if anything, in how a brand gets found and gets bought.

What was announced, without the trimmings

OpenAI presents Astra as a leap in computer use, browsing, software engineering, science and professional work. The concrete figures it has published:

• Computer use close to twice as fast as the previous version.
• Scores of 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3 and 100% on ExploitBench.
• Training on more than 100,000 GPUs at the Stargate facility in Texas.
• A new reasoning technique, recurrent depth, which hides some or all of the model's reasoning.

Two qualifications worth not skipping. First: those scores are published by OpenAI about its own models. There is no independent verification yet, and until there is, they are manufacturer claims rather than established facts. Second: the opacity of the new reasoning has already drawn objections from safety specialists, who question how far you can supervise a model that does not show how it reaches its conclusions.

What matters to a brand is not the model

It is one thing: the assistant no longer answers, it acts.

Until now the conversation about AI and visibility was about answers. Someone asked, the model replied with what it had read, and the argument was whether it cited you. A model that browses, uses a computer and completes multi-step processes changes the ground: it no longer just summarises pages, it goes into them, compares, fills in forms and sees things through.

OpenAI gives explicit examples of the kind of task: preparing taxes, ordering food, job hunting. Translated into what matters to a business: some of your visitors are about to stop being human, and they will arrive with a specific task they intend to finish.

Why this accelerates something already under way

This is not a turn, it is an acceleration. The data already pointed here: AI Overviews appear in roughly 16% of searches according to Semrush, and SparkToro's analysis puts 58.5% of US searches at producing no click at all.

We wrote about it months ago in SEO vs GEO vs AEO and the diagnosis has not changed. Astra does not change it either: it pushes it along.

What to do, and what not to

Worth doing

Make your site machine-readable. Correct structured data, headings that mark real hierarchy, prices and availability in the HTML rather than inside a script that has to run first. An agent comparing options needs to extract data, not interpret an animation.

Make your processes completable. If a form only works through three visual steps and no labels, an agent gets stuck. What makes a site usable for someone on a screen reader is almost exactly what makes it usable for an agent: real labels, buttons that are buttons, links that are links.

Publish something citable. Your own data, a case study with figures and a period, a view that is not already on twenty other sites. A generative model does not cite the generic: it cites what it cannot find elsewhere.

Earn brand mentions off your own site. That is what gets you named, and it does not depend on your domain.

Not worth doing

Do not throw away your SEO. The same work — structure, authority, useful content — is what makes you legible to an agent. Nobody has announced the end of search.

Do not rewrite your whole blog as questions and answers. Turning forty articles into FAQs to "optimise for AI" produces worse content for people, and models do not reward the format: they reward substance.

Do not buy anything this week out of fear. An announcement from twenty-four hours ago, still rolling out in phases and with no independent verification of its figures, is not a basis for changing a budget.

What we still do not know

And it is worth saying out loud: we do not know how this will affect real traffic. There is no data yet. Nobody has it, because the model was announced the day before yesterday and its rollout is not finished.

Anyone telling you this week exactly how much traffic you are going to lose is making it up. The honest position is that the direction has been clear for a while and that the speed, now, is higher.

What it comes down to

If your site leans on informational traffic that answered simple questions, that part has been eroding for a while and will erode faster. If it leans on being found when someone already wants to buy, the work is different: making sure an agent can understand your offer, compare it and complete the process.

That is a technical and editorial review, not a campaign. And it is the same one we were recommending before Astra existed.

---

*Sources: OpenAI's announcement of 3 September 2026 as reported by CNBC, 9to5Mac and Forbes. The AI Overviews and zero-click figures are from Semrush and SparkToro. The benchmark scores are those published by OpenAI and do not yet have independent verification.