OpenAI released GPT-6 Astra on 3 September, first to a small group of organisations and then, a day later, to ChatGPT Plus, Pro, Business and Enterprise plans and the API. Within hours the SEO corner of the internet had opinions. One widely shared post declared that GPT-6 Astra "should be illegal for SEO". Other early testers found it a worse writer than the model it replaced. So which is it?

Reading across what has been published in the first week, and remembering that the same argument played out for GPT-5.5 in the spring, the honest answer is that both camps are describing real things. They are just describing different jobs. Here is how it splits.

Where it is genuinely better: work with a right answer

Astra is a big step up at long, multi-step tasks where the result can be checked. Its context window is 1.05 million tokens, and OpenAI's own numbers show it staying coherent through longer jobs than GPT-5.6 Sol. In practice that means an entire site crawl, sixteen months of Search Console data, a competitor's pages and your existing content can all sit in one session while it reasons across them, instead of being summarised and forgotten every few messages.

The most useful published test so far is FloxoLab's audit of their own codebase, run the day after release. Astra found ten confirmed defects, including eleven BlogPosting schemas pointing at an image that did not exist, a redirect quirk and an accessibility failure in a modal. Every one held up under verification. The author's caveats are worth as much as the findings: it had full access to the site, a strict brief, and "every finding still had to survive source checks, tests, and production verification". Without those conditions, they note, the same model "could still produce a confident list that nobody should ship".

That is the pattern across the enthusiastic reviews from other testers. Keyword clustering, grouping terms by the job the searcher is doing, reading a set of results and saying what kind of page wins, spotting that four of your posts are circling the same intent, turning a shortlist into a brief: all of that is reasoning over material it can see, and it is very good at it.

Where the reviews split: writing

Some marketers report copy with less of the usual AI smell and a better feel for the intent behind a prompt. Others ran it through an editorial-voice test and placed it eleventh, below its predecessor, and found that AI-writing detectors were not fooled by paragraphs written in a named author's style. The fairest summary from that review is that Astra excels where there is a verifiable right answer and stays mediocre where the standard is taste.

For SEO content that changes little. A model that can draft a page in seconds was already the norm. The page still has to be checked, given real specifics, and published by someone who knows the business. What Astra improves is the brief that the draft is written from, because it can now hold all the evidence at once.

The thing everyone agrees on

Read past the headlines and the excited posts and the sceptical ones say the same thing: the model is not the data. Astra's knowledge cutoff is 30 April 2026. It has never seen a search-volume database, a rank tracker or your Search Console. Ask it how many people search for a phrase each month and it will give you a number, because giving you a number is what it is built to do. One of the more careful write-ups on keyword research with Astra puts it well: a more intelligent model produces a more convincing invented number, which is worse for you, not better.

This is not a niche worry. On the AA-Omniscience benchmark, which rewards factual recall and penalises wrong answers, GPT-5.5 posted the highest accuracy of any model and one of the highest hallucination rates at the same time. The hallucination rate there measures how often a model states a wrong answer rather than admitting it does not know. The OpenAI models are, by temperament, answerers. That is a wonderful quality in an analyst and a dangerous one in a data source.

So the practical rule for using GPT-6 Astra for SEO is the same one we wrote about when people asked whether ChatGPT can do their SEO, only more so. Give it the numbers and it is the best analyst most small businesses have ever had. Ask it for the numbers and it becomes an eloquent guesser.

Closing the gap

The way to give it the numbers is a connector. Astra supports the Model Context Protocol natively, which means it can call tools that return real measurements mid-conversation rather than relying on what it remembers from training. Our MCP server exposes thirty of them: daily positions in the top 30 for every tracked keyword, competitors, sixteen months of Search Console history, monthly search volumes from the database the model does not have, site audits, backlinks, security checks, and whether ChatGPT, Gemini and Perplexity are mentioning you at all. It never touches your website, and the only things it can set up are tracking inside your plan's allowance.

Connected, the questions that were guesswork become reports. Things worth asking it this week:

  • Which of my keywords are close enough to page one to be worth a fortnight of effort, and why?
  • Mine my Search Console history for queries I get impressions on but have no page for, and turn them into a content plan.
  • Did last week's drop line up with a confirmed Google update, or with something we changed?
  • Where do I get cited in AI answers today, and which questions name my competitors instead?
  • Write my web developer a prioritised brief from the latest audit, with the evidence attached.

Two things still to be careful with. Do not trust a number it has not just fetched, however confidently it is stated. And do not let it write and publish in one go: its own reviewers say the draft is where a human earns their keep.

The other side of the same coin

Astra is not only a tool you use. It is also now the model answering a growing share of your customers' questions inside ChatGPT. Every model change reshuffles who gets recommended, and it is measurable: when GPT-5.5 Instant became the ChatGPT default in May and brand links became clickable in answers, referral traffic from ChatGPT to websites rose by more than a third in a month. Nobody was told. The businesses that noticed were the ones tracking their mentions.

If you only do one thing this month, put the questions your customers actually ask to ChatGPT on a schedule, record who gets named, and watch what the new model does with your name. Then ask it why.