GPT-6 Astra reads websites like an agent, not a search engine

GPT-6 Astra, OpenAI's new model released on September 3, 2026, does not read your site the way a search engine does: it operates it like a person. It clicks, fills forms, opens several tabs and runs 40-minute tasks, scoring 72.6% on OSWorld 2.0 against 65.7% for GPT-5.6 Sol. In the coming days, this is the model generating ChatGPT's cited answers for paying users. When the reader of your content stops being a crawler and becomes an agent, the content that wins changes.

This is not a loose prediction. The sequence is verifiable: Astra launched September 3 in limited access for organizations in OpenAI's Daybreak program and reached paying users the next day (Axios, VentureBeat). Greg Brockman closed the briefing with "Welcome to the AGI era." OpenAI reports 98.6% on ARC-AGI-3 (run on its own harness, a caveat worth remembering) and positions Astra as "the world's best computer use model." You can argue about whether that is AGI, and people are. What is not arguable is what the model does with a browser.

What changed on September 3

Until last week, a cited answer worked roughly like this: the user asks, the search layer retrieves documents, the model summarizes and links. Your GEO job was to get into that retrieved batch.

Astra breaks that pattern in one place: on top of retrieving, it executes. OpenAI's demo shows Astra preparing tax returns, building product listings and assembling game scenes by operating real software. The benchmarks back it up. On OSWorld 2.0, a test of operating computers, Astra scores 72.6% in about 40 minutes per task, versus Sol's 65.7% in 75 minutes, roughly 47% faster. As VentureBeat put it, "Astra is designed to navigate software much as a person does, across browsers, spreadsheets, websites and desktop applications".

Translated to your site: when a user asks ChatGPT something like "compare marketing plans for a clinic in Guadalajara," the model answering does not settle for three extracted fragments. It opens your page, maybe the competitor's, checks tables, scrolls, compares figures. Your content is being read by a hired researcher, not an indexer.

The fourth reading mode: agentic browsing

Search traffic has been splitting into distinct reading modes for a while. Astra makes the split explicit. Each mode reads differently and rewards different things:

We have covered the first three separately: reading AI Mode queries in Search Console, measuring citation volatility and getting cited by ChatGPT search. The fourth layer does not replace them, it stacks on top. A page that is not indexable gets read by nobody, agent or not.

What our own data shows

This is not theory. We track the traffic mintec.co receives from AI engines every week, and this week it was 49 sessions in 7 days: 38 from ChatGPT, 5 from Claude, 4 from Gemini and 2 from Perplexity. The peak came on September 3, 15 sessions in a single day, the same day Astra was born. The most cited page, a post on Meta's customer lifecycle strategy, took 7 sessions from ChatGPT alone.

The pattern across our cited pages repeats: a direct answer in the first paragraph, numbers with dates, and a comparison or table. That is not luck. AI engines cite what they can verify fast, and they verify fast what is structured and dated. When the model also operates the browser, that pattern stops being convenient and becomes mandatory: an agent comparing data across two sites will notice when your number has no date or your table lives inside an image.

Verifiability becomes your advantage

There is one Astra detail almost nobody in marketing is discussing, and it matters most for GEO: its reasoning technique, "recurrent depth," hides part of its chain of thought, and safety researchers are already worried about how hard that makes it to monitor. OpenAI itself gated its most advanced cybersecurity capabilities behind a trusted-tester program (VentureBeat).

Why should that matter to you? A model that cannot fully explain why it picked a source is under more pressure to pick sources that need no explanation: verifiable data, named authors, clear dates. The model's opacity shifts the burden of proof onto content. Sites that make verification easy will be preferred by the agent, and by the humans auditing it.

Five checks to run this week

1. Every page answers its query alone. If a question needs three clicks, an agent with a 40-minute task budget might make them. A human will not. Write as if paragraph one is all anyone will read.
2. The full journey lives inside your domain. "How much does it cost?" and "What is included?" should sit two clicks apart, or on the same page. The agent follows the complete path if it exists.
3. Numbers carry dates and sources. "Our analysis of 49 emails" is not a citation. "49 LLM sessions in 7 days, internal report, September 6" is.
4. Key data lives in tables and schema. If your important number is inside an image, it does not exist for the agent. Markdown tables, FAQ schema, dates in the HTML.
5. Monitor per engine. Search Console now mixes AI Mode queries into the main report, and we published the method for reading them. Split ChatGPT, Claude, Gemini and Perplexity in your analytics and watch which pages they cite.

Opinion: GEO is not dying, it is stacking

The GEO playbook we have is not disappearing. The citation layer still works, as our real traffic numbers show. What changes is the ceiling: an agent that operates your site will remember whether your content survived the full visit, and the models trained on user behavior downstream will learn from that.

Watch the other direction too. The "agent visibility" sellers are already showing up, pitching magic services with invented metrics and guaranteed citations. This week's lesson is the same as always: nobody guarantees a citation from a model, least of all one that reasons opaquely. What stays guaranteed is verifiable, structured, honest content. That survives any model change. September 3 changed the reader, not the rule.