When OpenAI shipped GPT-6 Astra on September 3, the questions everyone asked were about reasoning and coding. The question we care about is narrower and more useful if you run a brand: when Astra answers, whose pages does it send people to?
We put it through the same study we've run on every ChatGPT release, and the answer came back higher than any model before it. Astra cites brand-owned sites 71% of the time, a new high across every version we've measured. Then the raw data showed a second change no benchmark captures: Astra dropped structured citations entirely, which is why some brands' ChatGPT numbers look like they collapsed this week even though nothing about their visibility changed. Here's the full picture.
TLDR
- Astra sends 71.1% of its citations to brand-owned sites, up from 58.1% on Sol and higher than any ChatGPT model we've tested.
- It runs Google's
site:operator on 75% of its searches, and writes it with a period (site.brand.com) that breaks naive tracking. - Answers are 54% shorter: 305 words on average, down from 662. This reverses the trend we reported a month ago.
content_referencesis empty on all 50 answers. Citations moved to plain inline links. Dashboards parsing the old field undercount to zero.- Google alignment held. 85% of the domains Astra cites don't rank in Google's top 10 for the same query.
Brand citations hit 71%, and reasoning effort no longer explains it
71% first-party, up from 58% on Sol. A 13-point jump on the same 50 questions.
Where it lands matters more than the size. A month ago Sol's High reasoning tier topped out at 61.4%, and we treated that as the ceiling. Astra clears it by ten points on the default Medium tier. Turning reasoning effort up used to be how you bought brand-heavy answers; the model now does it by default.
One answer up close:
Six inline links in a CRM comparison, all six to a brand's own pricing page: HubSpot, Salesforce, Pipedrive. Astra read review sites along the way; none survived into the answer. Repeat that across 50 prompts and you get 71%.
What to do: Astra quotes your own pages back to buyers, so give it something clean to quote. A plan comparison table, a plain-language value prop, and liftable feature bullets beat another blog post right now. Pricing gated behind "contact sales" can't be cited.
It runs site: on three of every four searches
75% of Astra's fan-out queries now carry Google's site: operator, up from 59% on Sol. It stopped asking open questions like "best CRM for a 50-person company" and started fetching specific pages, one brand at a time.
That shift hides a detail that will quietly break your tracking.
Astra writes the operator with a period, site.hubspot.com, not the usual colon. Google treats them identically, so the search works the same, but our own pipeline read Astra's site: usage as 0% until we caught it and rewrote the pattern. Any extractor matching on the colon has that blind spot today.
Add direct URL fetches, where Astra opens a pricing page instead of searching, and 83% of its search budget goes to targeted brand-page retrieval.
What to do: Widen any fan-out parser to catch both site. and site:. Then run a domain-scoped query against your own pricing pages and see what comes back, because that's the query Astra is running on you.
Answers are half as long, which reverses what we found last month
305 words per answer, down from 662 on Sol. Half the length.
Last month's Sol study reported both reasoning tiers converging around 670 words, and we called it the new baseline. It held for one generation. The drop is behavioral, not noise: Astra reads 50 web results per prompt where Sol read 120, runs fewer searches, cites fewer sources. The shape moved from a research assistant summarizing a dozen sources to a concierge that opens three pages, quotes them, and hands you a paragraph.
What to do: A 305-word answer has no room for anything you buried. Lead your key pages with the part you want quoted, the comparison table, the feature list, the positioning line. The page that opens with 2,000 words of throat-clearing loses to the one that opens with the answer.
As more of these answers come from agentic fetches rather than open search, tracking how AI agents move through your site becomes its own signal worth watching.
ChatGPT quietly dropped structured citations, and it's breaking dashboards this week
Zero anchors on all 50 Astra answers. Sol shipped 399.
content_references is the structured array that maps each citation to the sentence it supports and the URLs behind it. It powers ChatGPT's inline citation bubbles, and it's what most AEO dashboards parse to count citations. On Sol it was full. On Astra it's empty, every time.
The URLs didn't disappear. Astra moves them into the answer body as ordinary markdown links, [HubSpot pricing](https://www.hubspot.com/pricing/sales). We counted 273 across the 50 answers. The citation rate held; only the storage format changed.
What to do: Check which field your citation tracker reads. If your ChatGPT share fell off in the last two weeks, this is almost certainly why, not a real drop in visibility. Point your extractor at the inline links in the answer text and your Astra numbers come back.
If you'd rather not rebuild your parser, Writesonic's AI Visibility Tracker already reads Astra's inline citations, so your ChatGPT share stays accurate through the format change.
The full metric snapshot
Note for CMS: keep the nine core figures (first-party %, site: %, response length, content_references count) written out in the paragraph below too, so the numbers stay in the page's HTML and not only inside the image.
First-party citations up 13 points to 71.1%, site: usage to 75.1%, answer length down 54% to 305 words, content_references from 399 to zero, and 273 inline markdown links where Sol had none. Search still fires on 98% of conversations. Astra didn't stop searching; it changed what it searches for and how it reports what it finds.
Google alignment barely moved
85% of the domains Astra cites are absent from both Google and Bing's top 10 for the same query, against 89% on Sol. Barely moved, both overwhelming.
It's the direct result of the site: strategy. When Astra fetches site.hubspot.com pricing, it lands on a page that will never rank in Google's top 10 for "best CRM," and doesn't need to. Only 13% of Astra's cited domains appear in Google's top 10 for the matching query. Your rank tracker and your AI citations now measure two different worlds.
Astra's default-answer brands, regardless of where they rank: hubspot.com, salesforce.com, xero.com, freshbooks.com, zoho.com, quickbooks.intuit.com.
Dig deeper: which sources AI models actually cite, across every major engine
Where the brand gains actually landed
The 71% average hides a wide spread. Travel jumped from 68% to 94%, Finance and Services hit 100%, Education and Legal climbed into the 90s. If you sell in those categories, Astra's default answer now points at your own domain far more often than Sol's did.
Two categories moved the other way. Fitness fell from 53% to 47% and Home from 57% to 47%, where Astra leaned harder on aggregators and review sites than Sol did. Single-run category cells are small, so treat those two as a signal to watch, not a verdict. We're rerunning them.
The takeaway isn't the average. A single release can hand your vertical 26 points or take 10 away, and the blended number tells you nothing about which happened to you.
What this means for your next quarter
Make your pricing and product pages quotable. Astra's default move is a domain-scoped query against your own site, and it quotes what comes back. A clean comparison table, clear value props, and liftable feature bullets do more for your AI visibility right now than another top-of-funnel post.
Write for a 305-word answer. The summarization step got sharper. Front-load the pages you want cited with the structure a model extracts first, not narrative it has to dig through.
Fix your citation parsing this week. If you track your ChatGPT citation share, confirm your extractor reads inline markdown links and not only content_references. Anyone still on the old field is looking at a false zero for the newest model in the market.
Judge the change per category, not on the average. Astra handed 20-plus points to Travel, Education, and Finance while pulling back in a couple of others. The blended number hides both moves.
Questions we're still chasing
- Is the empty
content_referencesarray permanent, or a Work-tier beta artifact? If OpenAI restores it when Astra reaches the standard picker, dashboards recover on their own. - What's Astra's citation freshness? Its inline links don't carry the publication dates Sol's anchors did, so measuring it means matching URLs back to the search payload.
- Do the category pullbacks hold across repeat runs, or are they single-run noise?
- When does Astra reach the classic Chat picker on Plus, and does its behavior shift on that surface?
- Does a 305-word answer change how often anyone clicks a citation at all?
FAQ
How we ran the study
Same 50 prompts we've run against every model since March, 16 categories from SaaS to travel, legal, and food. Astra isn't in the standard Plus picker yet, so we captured it in the Work workspace, the only surface exposing it. Sol figures come from our previous study.
The workspaces differ, so we controlled for it. Every conversation was verified on model_slug: gpt-6-astra-wm and thinking_effort: standard before we kept it. Standard is the effort Sol Medium ran at, so every number below is a version change, Astra against Sol, not an effort or workspace artifact. Claude Haiku 4.5 classified each cited URL as first-party or third-party using the same prompt as every prior study.
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.
