AGI Arrived Last Week. Here's What It Means for Your Website.

Jensen Huang says AGI has arrived. OpenAI shipped GPT-6 Astra on September 3, and three days later the CEO of Nvidia posted four words on X that no AI lab has been willing to say out loud. Whether he's right is an argument for the researchers. What it means for anyone who owns a website is not an argument at all. The thing researching your business is now a model with a browser, and it either reads your site or it doesn't.

This post covers what actually shipped, why the AGI label is worth about as much as the tweet it came in, what Astra does that matters, and the three fixes every business needs to make now that the model is the customer's research assistant.

What actually happened, in order

Here's the timeline, because the press coverage has scrambled it.

On September 3, OpenAI released GPT-6 Astra, calling it “the most intelligent and aligned model in the world.” It rolled out first to a limited set of organisations and is now going out to ChatGPT Plus, Pro, Business and Enterprise users, plus the API, Microsoft Azure and AWS Bedrock.

On September 6, Nvidia CEO Jensen Huang replied to a thread on X: “GPT-6 Astra, trained on ~100K+ NVIDIA Grace Blackwell NVLink72. From ChatGPT to o1 to Astra in 4 years. AGI has arrived.” He added that 400,000 more GPUs are coming online next. His first version of the post said 300,000 chips trained the model. He deleted it and reposted with the smaller number, which Yahoo Finance picked up.

On September 8, Fox Business ran the story under the headline that has been everywhere since: “Nvidia CEO Jensen Huang declares ‘AGI has arrived.’”

So the claim didn't come from OpenAI. OpenAI's own president, Greg Brockman, told reporters it was “not unreasonable to feel that we are now in the AGI era,” which is about as far as a company can lean without saying the words. The words came from the man who sells the chips, in the same breath as an announcement that he's selling 400,000 more of them.

Is it AGI? Depends who's selling

AGI has no agreed definition, and that's the whole problem with the headline.

Nvidia's definition, as Fox Business reported it, is a system that can pass professional certifications and standardised tests across a range of fields with top-tier scores. By that bar, Astra qualifies. So did a couple of models before it.

OpenAI's definition is “highly autonomous systems that outperform humans at most economically valuable work.” That's a much higher bar, and OpenAI itself hasn't claimed Astra clears it.

Gary Marcus, who has spent a decade being right about what AI can't do yet, called Huang's post an attempt to settle a scientific question by corporate fiat. He points out that Huang gave no evidence and no definition, that Astra is roughly on par with Anthropic's Claude Fable 5.1 in real use rather than a leap ahead of it, and that Huang already declared AGI once before about an earlier model. Marcus also quoted the creator of the ARC-AGI benchmark, who was blunt: saturating ARC-3 is not proof of AGI, and they're not making that claim.

That last point matters because ARC-AGI-3 is the number OpenAI leads with. Astra matched the human action-efficiency baseline on 96% of levels. It's a real result. It's also a benchmark, and the people who built the benchmark are telling you not to read it as a finish line.

The explainx.ai breakdown put it well: a capable model shipped on an enormous hardware footprint, and the company that sold the hardware called it AGI while announcing more hardware. Both halves are true. Only one is a technical claim.

My take: I don't care what we call it. I care what it does. And what it does is the part nobody in the AGI argument is talking about.

What Astra actually does

Strip away the benchmark table and Astra is a model that operates a computer. That's the product.

OpenAI's own list of what it handles: filling out online forms, updating customer records, organising calendars, researching online, drafting documents, building spreadsheets and presentations, installing software, and troubleshooting what it sees on screen. The Fox piece quoted the company calling these “tedious” tasks. A developer who tested it on launch day made the sharper point: that sounds boring until you remember most work is boring.

A few facts from the OpenAI API documentation and OpenRouter's listing, because the numbers tell you who this is for:

• 1,050,000-token context window, with output up to 128,000 tokens. It can hold an entire website, every page, in memory at once.
• $10 per million input tokens, $50 per million output tokens. Web search is billed at $10 per 1,000 calls. Every time it browses, someone is paying for it, which means it's browsing with intent.
• Reasoning effort from low to max. It can be told to think hard about a task, and it will.

Two more things from the safety overview that didn't make the business press. Astra is the first OpenAI model to reach the Critical threshold for cyber capability under their Preparedness Framework, meaning it can find previously unknown security flaws without a person guiding each step. And it's significantly more resistant to prompt injection than the previous model, which is the first thing that matters if you're going to let it loose on the open web. Al Jazeera framed the launch around exactly that scrutiny.

Then there's the part Sam Altman said impressed him most. He asked Astra to research a chip supply chain, and it came back with sections he hadn't asked for, saying in effect, you didn't think to ask me about this part. The model isn't waiting for instructions. It decides what's relevant, goes and looks, and reports back.

Hold that thought, because it's the whole reason this post exists.

Why this is a website problem, not an AI problem

Think about who researches your business.

Two years ago it was a person. They typed into Google, scanned ten blue links, clicked three, read your homepage, compared you to two competitors, and made a call. Your job was to rank and then convert.

Today it's a model. Someone asks ChatGPT, or Perplexity, or Google's AI mode, “who does X in Y,” and the model does the research. It fetches twenty sites in the time it takes you to read this sentence, decides which three are worth mentioning, writes up an answer, and hands it over. The person never sees the other seventeen. They don't scroll. There is no page two of an AI answer.

Astra pushes this one step further. It doesn't just answer the question, it runs the task. Book the appointment, fill the intake form, shortlist the vendors, draft the email. The model is now the customer's research assistant and, increasingly, their hands.

So the question for every business owner has changed. It used to be “do I rank.” Now it's “does the model read me, does it trust me, and does it say my name.” That is what GeoTest measures, and that is what this post is about from here on.

The three ways your site fails an AI model

We've scored a lot of sites. The failures fall into three buckets, and most sites have at least two of them without knowing it.

01 The model can't read your site at all

This is the big one, and it's the one nobody sees because it's invisible in a browser.

If your site was built on Lovable, v0, Bolt, Replit, or a standard client-rendered React or Vite setup, it may ship as an HTML shell that only becomes a complete page after JavaScript runs. It looks perfect in a browser. But crawlers with limited or no JavaScript rendering can receive little more than the shell: no useful text, headings, prices, or contact details.

Google's own documentation on JavaScript SEO has warned about rendering dependencies for years. AI crawlers differ in what they render and how long they wait. The safe standard is simple: important content should be present in the server response, not dependent on a crawler executing your app correctly.

We check the raw server response as well as the rendered page because the difference exposes this failure. Browser-only audits can miss it: they load the page like a human would, JavaScript runs, and everything looks fine.

I've watched clients spend months on content while their most important text was absent from the initial response. Fix this first. Nothing else counts until it's done.

02 Your page doesn't answer the question

Models cite pages that state the thing plainly. What you do, who it's for, what it costs, where you are, how long it takes. They skip pages that tease, pages that make you fill out a form to find out the price, pages that are 800 words of “we're passionate about” and zero facts.

This isn't a style preference. A model building an answer is looking for a sentence it can lift and attribute. “Emergency plumber, Podgorica, 24 hours, callout from €40” is citable. “Your trusted partner for all your plumbing needs” is not. The first one can end up in the answer with your name next to it. The second one is easy to discard.

Real numbers beat adjectives. Real pricing beats “contact us for a quote.” A page that answers the question a buyer would actually ask is worth more than ten pages that talk around it.

Structured data belongs here too. Schema.org markup tells machines what your page represents, whether it's a service, product, FAQ, or local business, and provides clean fields to parse. It's boring work. It's also 18 points out of 100 in the current GeoTest rubric because it provides an explicit machine-readable description of the page.

03 Nobody else mentions you

Models can use what other sites say about you to corroborate what you say about yourself. That's not a flaw, it's how you'd do it too. If the only place on the internet that says you're the best is your own homepage, the model has no independent reason to believe it.

Reviews on Google and relevant industry directories. Partner pages. Press. A mention in someone else's blog post. Your name in a comparison article. Each one is an outside signal the model can inspect.

This is the slowest fix and the one most people skip. It's also the one that separates a site that gets cited occasionally from one that becomes a dependable source.

What Astra changes about all three

Here's why last week matters more than the launches before it.

A chatbot that answers questions can rely on previously indexed material. A model that runs tasks often needs current information. When Astra is booking an appointment or shortlisting vendors, it may visit your site live and needs the page to work. If the page is blank, the task can fail or another provider can be selected. If the price isn't on the page, the model has less information than it gets from a transparent competitor. If nobody else on the internet vouches for you, it has less corroboration.

And browsing has a metered cost. An agent has a budget; businesses should not assume it will keep retrying an inaccessible page.

The AGI debate is about whether the model is as smart as a person. The business question is simpler: the model is doing work a person used to do, and your website was built for the person.

What to do this week

1. Test the raw response. Run your site through GeoTest. It fetches your pages and tells you what came back. If the answer is an empty shell, everything else waits.
2. Put the answers on the page. Pick the five questions a buyer asks before contacting you: price, location, timeline, what's included, and who it's not for. Answer them in plain sentences. Add appropriate schema.
3. Start collecting outside mentions. Ask recent clients for a review. Get listed in the two directories your industry actually uses. Ask one partner to link to you. Small signals compound.
4. Re-test in 30 days. Scores move when the fixes are detectable. If they don't move, inspect what the crawler actually received.

What's next

Astra is one model. Anthropic, Google, and others are shipping similar capabilities, and the next version of each will browse more, decide more, and act more. The direction is fixed even if the AGI label isn't.

For GeoTest, this is the business we built for. We score whether models can read you, whether they cite you, and what they say when they do. Up until last week I had to explain why that mattered. Now everyone I talk to has already watched it happen.

Short version: the customer's research assistant is a model now. AGI or not, it's reading your site today. Make sure it can.

Frequently asked questions

Has AGI actually arrived?

Depends on the definition. Nvidia's definition, top-tier scores on professional certifications and standardised tests, is met. OpenAI's definition, outperforming humans at most economically valuable work, is not, and OpenAI has not claimed it. Researchers like Gary Marcus say Astra is a real improvement but on trend, not a leap. The claim came from Nvidia's CEO, not from OpenAI.

What is GPT-6 Astra?

OpenAI's flagship model, released September 3, 2026. It is built for computer use: browsing, filling forms, updating records, research, coding and document creation. It has a 1,050,000-token context window, costs $10 per million input tokens and $50 per million output tokens, and is available in ChatGPT Plus, Pro, Business and Enterprise, through the OpenAI API, Microsoft Azure and AWS Bedrock.

Why does an AI model launch matter for my website?

Because models are now doing the research your customers used to do themselves. A model fetches your site, decides whether to mention you, and hands the customer an answer. If the model cannot read your site, or your site does not answer the question, you are not in the answer.

How do I know if AI can read my site?

Check the raw server response, not only the rendered page. Sites built on client-rendered React setups can send a nearly empty HTML shell to crawlers. GeoTest fetches your pages the way an AI crawler does and shows you what came back.

What is GEO?

Generative Engine Optimisation. It is the work of making your site readable, citable and trusted by AI models like ChatGPT, Perplexity, Claude and Google's AI mode, so that when someone asks a question in your category, your business is in the answer.

What does GeoTest score?

Six categories out of 100: content quality and depth, structured data, citation worthiness, AI crawler access, authority signals, and FAQ and knowledge coverage. Scores of 80 or above are excellent, 65 to 79 good, 50 to 64 moderate, 35 to 49 weak, and below 35 poor.

What's the first thing I should fix?

Crawler access. If the model gets a blank page, nothing else you do counts. Fix the raw response, then put the answers on the page, then work on outside mentions.

Does this apply if I'm a small local business?

More than to anyone. Local queries such as ‘who does X near me’ are exactly where models are replacing the Google search, and local sites are especially likely to rely on page builders that may send limited content to crawlers.