Keyword Research With GPT-6 Astra: The Real Setup

GPT-6 Astra landed on September 3, 2026, and the first thing a lot of people tried was asking it for keyword volumes. It answered. The numbers looked plausible. They were not real.

That is not a knock on the model. It is a category error. Search volume lives in a search database, not on the web pages a model trained on, and Astra’s knowledge cutoff is April 30, 2026 regardless. A more intelligent model produces a more convincing invented number, which is worse for you, not better.

What Astra genuinely changes is how much of the job fits in one session. That is the part worth rearranging your workflow around, and it is not the part most launch coverage talks about.

What Astra can and cannot do on its own

Split the job in two and it stops being confusing.

What Astra is genuinely good at. Expanding a seed into themes, grouping terms by the job the searcher is doing, reading a SERP and telling you what kind of page wins it, spotting that four of your existing posts are circling the same intent, and turning a shortlist into a brief. That is judgment work over text, and it is what the model is for.

What it cannot do at all. Tell you that a term gets 480 searches a month. Tell you its difficulty. Tell you who ranks today. Tell you what your own site already gets impressions for. None of that is inferable from training data, and the confident tone it uses for the first list does not change when it moves to the second.

That gap is the entire reason this post exists. The workflow below is just the discipline of keeping the two halves separate.

The context window changes the shape of the job

Astra has a 1,050,000-token context window, takes up to 922,000 tokens of input, and returns up to 128,000. Those figures are on the official model page. Numbers that large stop meaning much, so here is what it buys in practice.

A normal keyword research session is a series of amputations. You research 200 terms, then summarize them down to 40 because that is what fits, then lose the detail that would have told you two of them were the same page. Then you paste in a few SERPs, then drop them to make room for your content inventory.

With a window this size, you stop summarizing. The full expansion, the difficulty numbers, the live top ten for your head terms, and a list of every post you have already published can all sit in context while the model decides what to write next. The decision gets better because nothing was thrown away before the decision was made.

That is a genuine change, and it is the one thing that makes Astra worth a different workflow rather than the same workflow with a new model name.

The pricing cliff nobody mentions

Here is the part that will surprise you on the first invoice.

OpenAI’s standard text pricing for Astra is $10 per million input tokens, $1 per million cached input tokens, and $50 per million output tokens. Fine. But prompts above 272,000 input tokens cost more than the standard rate.

Read that next to the section above. The workflow that makes Astra worth using is the one that fills the context window. The context window has a price step in it at roughly a quarter of its capacity. So the exact job this model is best at is the job that walks you over the line.

None of that makes it a bad tool. It makes it a tool with a shape you should know before you point an agent at it and walk away. Two practical consequences follow.

• Load the data once and let it cache. Cached input is $1 per million against $10. A long session that keeps referring back to the same keyword table is far cheaper than ten fresh sessions that each re-paste it.
• Use Batch or Flex for the bulk passes. Both run at half the standard rate. Clustering 800 terms does not need to happen interactively.

The workflow, step by step

1. Connect the data before you ask anything. An SEO MCP server gives Astra a tool it can call mid-sentence. Without one you are pasting exports, and every follow-up question means going back to a dashboard. The point of doing this inside an agent is that the loop stays closed.

2. Give it the commercial context first. Name the site, the country, who you sell to and what a conversion is. A keyword can be easy to rank for and worthless to you, and the model has no way to know which is which unless you say so. This one paragraph does more for the output than any prompt trick.

3. Expand once, from one seed. One good seed expansion beats six. Six gives you six competing tables and no way to reconcile them.

Copy this prompt

Expand this seed into keyword themes for my site: [seed].
For each theme, pull volume, difficulty and intent from the SEO tools.
Group them by the job the searcher is doing, not by word similarity.
Then flag any theme my existing posts already cover.
Stop there. Do not draft anything yet.

4. Read the real SERP before you commit. Difficulty scores are a shortcut. The live top ten is the evidence. Ask Astra what kind of pages rank, who owns them, and whether anything in there is weak enough to displace. If the answer is a wall of established publishers, walk away and keep the credit.

5. Shortlist, then stop. The discipline that saves the most time is ending the session before drafting. A target that survives a night’s thought is worth a post. One that does not was cheap to discard.

Four things the big window actually unlocks

Generic advice about context windows is useless. These are the four jobs that were awkward before and are not now.

1. Deduplicate against your own site before you write. Load every published title, description and target keyword alongside the new expansion. Ask which shortlisted terms are already covered by something you shipped. This is the check that stops you writing your fourth post about the same intent, and it was previously too much text to hold next to the research.

2. Cluster on live SERPs instead of on words. Two keywords belong on one page when the same kind of result ranks for both, which you can only know by looking. With the SERPs for thirty terms in context at once, the model can group by what Google actually returns rather than by string similarity.

3. Read a whole Search Console export in one pass. Your impressions data is the best keyword source you own and it is usually too big to paste. Export the last three months from Search Console, drop it in whole, and ask which terms you rank 8th to 20th for with real impressions. Those are cheaper wins than anything a keyword tool will suggest, because you already rank.

4. Keep the brief and the evidence together. When the model writes the brief in the same session that holds the SERPs and the volumes, the brief can cite them. You get “target this because the top three are all thin listicles” instead of an unsourced instruction.

None of these need a new tool. They need the research and your own data to be in the room at the same time, which is the thing that changed.

What makes a shortlist survive

The model will happily hand you thirty targets. Four is a better answer, and these are the tests that get you there.

Does the current top ten leave room? Not “is the difficulty score low” but whether anything on that page is beatable by a better version. Forum threads, a five-year-old post, a thin listicle, all beatable. Four established publishers and a documentation page, not beatable this quarter.

Would the person searching this buy anything? Informational terms build an audience, commercial terms build revenue, and you need both, but you should know which one you are picking each time rather than discovering it later.

Do you have anything real to say? The terms where you can write something nobody else can are the ones worth the week. Everything else is a summary of other people’s pages, competing with the pages it summarizes.

Can you point at the page it becomes? If you cannot describe the finished post in one sentence, the target is a topic rather than a keyword, and it will turn into something unfocused.

A shortlist that passes all four is short. That is the point of it.

Where it still falls short

It will agree with you. Ask “is this a good keyword?” and you have already told it the answer you want. Ask “what is the strongest reason not to target this?” instead. Same model, better information.

It cannot see your own Search Console unless you connect it. Your impressions data is the best keyword source you own, and it is invisible to a model by default. Terms you already rank 12th for are almost always a better use of a week than a term you rank nowhere for.

Volume is not demand for you. A 2,000-a-month term in a market you do not serve is worth less than a 40-a-month term your buyers type. The model optimizes for whatever you told it to optimize for, so tell it revenue, not traffic.

And it does not know it is guessing. This is the one that costs people real time. There is no tonal difference between Astra reporting a number a tool returned and Astra producing a number that feels right. If you did not watch a tool call happen, treat the figure as fiction.

Running it with ContextBolt SEO

ContextBolt SEO is a hosted MCP server, so connecting it to Astra is one URL rather than an install. Once it is connected, keyword research, difficulty, live SERPs, competitor rankings and your own Search Console data are all things you ask for in plain English, in the same session where the decision gets made.

Pricing is a flat $35 a month with 1,000 credits, and keyword, difficulty and SERP calls are one credit each whatever comes back. So a broad expansion does not cost more than a narrow one, and you are not pricing each question before you ask it. If a lookup comes back empty, it costs you nothing, because charging for a number the product did not return is indefensible.

There is no dashboard to learn. The research collects on an SEO Board you never had to build, so the session you ran last Tuesday is still there when you come back to it.

Is Astra the right model for this?

For a long research job that spans browsing, files and several tools, it is the strongest OpenAI option available and the context window genuinely changes what fits in one pass.

For clustering 500 terms or generating title variants, it is overkill and a smaller model does it for a fraction of the cost. The honest version of this advice is that the model matters less than whether it can see real numbers. Claude runs the same workflow perfectly well, and so does ChatGPT once you close the data gap.

Pick the model you already work in. Then give it a live source, because that is the half neither of them can invent.