When OpenAI shipped GPT-6 Astra this month, everyone asked about reasoning and coding. The question that matters if you run a brand is narrower: when Astra answers a buying question, whose pages does it actually read?
Astra, the newer and more capable model, searches the web 40% less than the one it replaced. Fewer queries per answer, fewer pages opened to write each one. And the searches it keeps almost never touch the open web. More than eight in ten go straight to a specific brand's own pages.
Astra has stopped asking the internet what it thinks of you. It reads you directly now. Here's what that changes about getting cited, and it continues a pattern we've watched across every ChatGPT model this year.
GPT-6 Astra runs 40% fewer searches than the model before it
For every question, GPT-6 Astra fires 6.1 search queries on average. GPT-5.6 Sol, the model it replaced, fired 10.3 on the same questions in the same environment. That's a 41% drop in how much the model reaches for the web.
It reads less, too: 50 web results per answer against Sol's 82. Fewer searches, fewer pages opened, and yet the answer a user gets is about as citation-dense and points to brands about as often. Astra is doing the same job with a much smaller search footprint.
The obvious question is where the searches it keeps are going. They're not spread across the web.
83% of its searches go straight to brand-owned pages
Three out of every four searches Astra runs are scoped to a single brand's domain, using Google's site: operator to pull pricing, plans, and feature pages from one company at a time. Another 8% skip search entirely and open a brand URL directly, https://www.hubspot.com/pricing/sales, because the model already knows the address.
Put those together and 83% of Astra's search budget goes to brand-owned pages. Only about 17% is open-web querying: the reviews, roundups, and forum threads that used to make up most of what ChatGPT read.
This didn't start with Astra. Domain-scoped searching has risen every generation, 59% two models ago, 71% on Sol, 75% now. It's the same shift that brought the site: operator back on GPT-5.6 Sol, and Astra pushes it further and faster.
ChatGPT reads your own pages now, not what reviewers say about you
For years, your visibility inside ChatGPT ran through other people's pages. If a G2 reviewer didn't rank you in the top five, if Reddit had no thread about your tool, if TrustRadius scored you low, that's what the model saw and repeated. You optimized by getting mentioned in the sources ChatGPT trusted.
Astra's behavior quietly dismantles that. The model reads less like an analyst summarizing what the web thinks of you and more like a buyer who opens your pricing page, scans the plan cards, and quotes them back. When 83% of its retrieval is your own pages, the middleman thins out.
That moves the whole game onto your own site:
Your pricing, product, and plan-comparison pages are now the highest-leverage real estate you own for AI visibility. They are the pages Astra actually opens. If your pricing hides behind a "contact sales" button, or buries the plan-vs-plan numbers under scroll and marketing copy, Astra quotes the wrong thing or moves on.
Your URLs became a ranking factor. When the model opens a page by typing the address from memory, it has to know that address. Clean, canonical, semantic URLs get fetched. Ones buried under redirects and tracking parameters get missed.
Review content still earns its keep, but less of it. G2, Capterra, and the rest still get read on the ~17% of open-web queries. They're no longer the main event. Budget accordingly.
Astra cites brands as often as before, with half the searching
None of this means Astra is cutting corners. On the numbers that matter to a brand, it holds steady with the model before it: it cites brand-owned pages about as often, and packs a similar number of citations into each answer. The answer is roughly 12% shorter, which reads as tighter rather than thinner.
So the story isn't that a weaker model started guessing. It's that a stronger one learned to get to the same place with half the searching, by going straight to the source instead of canvassing the web first.
How to get cited when ChatGPT reads you directly
Make your pricing page quotable in one read. Put the plan-vs-plan comparison, the actual prices, and the core feature differences above the fold and in plain text. Astra opens the page, lifts what's clear, and leaves. Clarity is the whole game.
Keep your URLs clean and stable. Direct fetches mean the model is working from memorized addresses. Short, canonical, semantic URLs get opened. Churn, redirects, and parameter soup cost you the fast path in.
Write for the quote, not the rank. Astra reads ~50 results and quotes ~5 in a 305-word answer. Short claims, real numbers, and minimal hedging survive that compression. Brochure copy gets dropped.
Rebalance away from pure review plays. Third-party review sites still matter for the shrinking open-web slice. They shouldn't be where most of your AI-visibility budget goes anymore. If you want to see which of your pages ChatGPT actually reaches and quotes, Writesonic's AI Visibility Tracker maps it against every major model.
Every metric, side by side
Across the same 50 prompts at matched effort, GPT-6 Astra issues 6.1 fan-out queries to Sol's 10.3, reads 50 web results to Sol's 82, runs site: on 75.1% of its searches, and lands a 71.1% first-party citation rate in a 305-word answer. Less searching, tighter output, and a much heavier lean on brand-owned pages.
Questions we're still chasing
- Does the efficiency gain hold at higher effort? We want to run the same test on the High tier for both models.
- Do the other models in the GPT-6 family show the same "search less, fetch brands more" pattern, or is it specific to Astra?
- How stable is it? Running the same prompts five times each would tell us whether Astra's brand-page picks are consistent enough to optimize against with confidence.
- Does it hold outside English? The same prompts in Hindi, Spanish, and German would show whether this is a property of the model or of the English web it reads.
Methodology
50 prompts across 16 categories, held constant, run at matched effort. We captured three conditions so we could separate what the model changed from what the environment changed: GPT-5.6 Sol in the classic Chat surface, GPT-5.6 Sol in the newer Work surface, and GPT-6 Astra in Work (the only surface exposing it on Plus). The clean model comparison in this piece is Sol and Astra at matched Work conditions, so the deltas we report reflect the model, not the surface. Payloads pulled from ChatGPT's /backend-api/conversation/{cid} endpoint; URL classification by Claude Haiku 4.5 with prompt caching; site: detection patched to catch Astra's site.brand.com variant.
Data captured September 7-8, 2026. Astra: 50 verified conversations. Sol-in-Work: 49 verified. One run per prompt per condition, single US Plus account. Directional by design.
The switch between ChatGPT's Chat and Work surfaces also changes citation formatting and answer length across every model routed through them, including how structured citations are stored. That's a separate finding with its own implications for measurement tools, and we'll cover it in a follow-up.
Every model release moves the citation map. This one moved it toward your own front door. The brands that make their pricing and product pages easy to quote will win the citation; the ones that gate or bury the details will watch Astra quote someone else. See how your pages read to AI today at https://app.writesonic.com/signup, or book a demo and we'll walk your domain through it.
Data & Infrastructure @ Writesonic
Tarsh builds the data pipelines behind Writesonic's AI visibility research, including the AI Ads Index. His work focuses on real-time measurement of how AI platforms cite, mention, and rank brands.
Growth Marketer
Tanay covers the intersection of AI and marketing at Writesonic.
His work focuses on how LLMs and AI Agents are reshaping search, and how marketing teams can adapt their SEO and content strategies for an AI-first search landscape.
