Does GPT-6 Astra change how ChatGPT decides which sources to cite?

Yes. GPT-6 Astra searches less but cites almost as often, and it has switched from search-tag citations to quoting the actual source text behind an answer. It also sends most of its searches straight to a single brand's own domain rather than scanning the open web, which changes what "getting cited" even means.

What GPT-6 Astra actually changed

GPT-6 Astra went live to the general public on 4 September 2026, with GPT-6 Sol and GPT-6 Luna following later that month on 22 September 2026 (Wikipedia). OpenAI called it a "generational leap," and company president Greg Brockman went as far as suggesting it could eventually be seen as the arrival of artificial general intelligence (Wikipedia).

On the technical side, OpenAI's model documentation lists a 1,050,000 context window, 128,000 max output tokens, and an April 30, 2026 knowledge cutoff (OpenAI community post). OpenAI's developer docs also note that Astra gets stronger results using substantially fewer output tokens, giving it a lower estimated API cost per task despite higher per-token pricing than earlier models (OpenAI developer docs).

None of that tells you directly how citations work. For that, you need to look at how the model actually searches.

How Astra searches and cites sources differently

A GPT-6 Astra citation study found the model issues 6.1 search queries per answer, compared with 10.3 from the model it replaced, GPT-5.6 Sol, run in the same environment (Writesonic's citation study). That is roughly a 41% drop in how often the model reaches for the web (Writesonic's citation study). It also reads less once it does search: 50 web results per answer against Sol's 82 (Writesonic's citation study).

This fits the broader toolset OpenAI documents for Astra, which includes web search alongside file search, code interpreter, hosted shell, computer use and MCP (Yottalabs). Web search is one tool among several, and Astra appears to reach for it more sparingly than its predecessor did. If you've followed the pattern of models skipping search entirely for a large share of prompts, this will feel familiar; see our earlier coverage of ChatGPT skipping the web for six in seven prompts.

The 40%-fewer-searches, direct-to-brand-pages pattern

The headline number from the Writesonic study is that Astra runs 40% fewer searches than the model before it, yet sends 83% of them straight to brand-owned pages (Writesonic's citation study). 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 (Writesonic's citation study). Another 8% skip search entirely and open a brand URL directly, such as hubspot.com/pricing/sales, because the model already knows the address (Writesonic's citation study).

Despite the smaller footprint, the study notes the answer a user gets is about as citation-dense and points to brands about as often as before, meaning Astra is doing the same job with far less searching (Writesonic's citation study). That changes the practical question for brands from "will I get crawled broadly" to "does my own domain carry the pages the model wants." It is also consistent with the separate branded-vs-non-branded study from the same research team, which found every platform increases product page citations in branded queries, with lifts of 20-60% (Writesonic's branded-vs-non-branded study). The same study found pricing pages drop 25-96% in branded contexts on other platforms, with Claude nearly abandoning them, while user-generated content climbs 46% on platforms like Copilot and Grok (Writesonic's branded-vs-non-branded study). Astra's own behaviour, heavily weighted toward site:-scoped brand searches, sits at one end of that spectrum.

If you want to see how this plays out for a specific product category, our piece on AI Overviews covering 88% of buying-intent searches covers similar territory from the Google side.

The new "quoted source" citation format

The mechanics of citation itself have also changed. OpenAI's own changelog-tracking site reports that with GPT-6 Astra, a source reference is attributed by the quoted text behind the answer rather than by a citation tag, so checking where a response came from is a matter of reading it rather than clicking a link marker (V7Labs' changelog). The same changelog notes that run transcripts changed too: every user turn in a run is kept, so an audit or a playback shows the whole conversation rather than a fragment of it (V7Labs' changelog).

This shifts the unit of citation from a URL stamped on an answer to a block of quoted text. If your page doesn't contain a clean, quotable sentence that answers the question, the model has less to lift directly, even if it visited your page. This echoes the shift we covered in ChatGPT's Intelligent UI shaking up what counts as a citation, where the definition of "being cited" kept moving under marketers' feet.

Why fabricated or misattributed citations are still a risk

A research-use guide to Astra warns that despite improvements, the model's ability to generate summaries does not remove the need to check that cited sources are both real and accurately described, and that fabricated citations and non-existent references remain a known risk for all LLMs, including Astra (Layer3 Labs' guide). The same guide states that while Astra is better at context selection and template matching, it can still fabricate references or misattribute claims, and that every citation and source must be verified manually before relying on its output (Layer3 Labs' guide).

This is the honest limitation here: a quoted-text citation format looks more trustworthy than a bare link, but it doesn't guarantee the quote is accurately sourced. Layer3 Labs recommends prompts that explicitly ask for direct quotations, require citation links, and ideally use a multi-step build, first listing sources, then mapping claims to sources, then writing the final summary (Layer3 Labs' guide). That's good practice for researchers using Astra, but it's a workaround for users, not a fix built into the model.

What this means for getting cited by ChatGPT

Domain authority still matters underneath all of this. A broader Writesonic analysis of the top 200 most-cited domains in AI Overviews found that sites with a domain rating of 88-100 received more than 6,000 average citations, which suggests authority still strongly predicts citation volume in the current AI-search environment Astra operates within (Writesonic's domain-authority analysis).

For "GPT-6 Astra" as a topic itself, one analysis found theverge.com and reuters.com each pulled 17 mentions among ChatGPT's cited sources, with a second tier of developer and education sites, including tembo.io and datacamp.com, at five each (GEO Toolbox's analysis). The same analysis found that across 3,121 citations on the topic, only 756 came from blogs, and that layer skewed heavily negative, 468 negative to 43 positive, with the single biggest publisher a syndicated site out of Egypt with 219 citations (GEO Toolbox's analysis).

Our view: the site:-scoped search behaviour is the biggest practical shift here. If Astra is going to open your pricing page directly 83% of the time rather than crawl the open web, the fight isn't really about general visibility anymore, it's about whether your own domain has a clean, quotable, up-to-date page for the exact question being asked. Generic blog content aimed at broad topical authority matters less for this specific pattern than a well-structured pricing or feature page on your own site. For more on why ranking well on Google doesn't automatically translate, see why topping Google means nothing to ChatGPT right now.

What to do about it

First, make sure your pricing, plans and feature pages are accurate, current and written in plain quotable sentences, since Astra is scoping three-quarters of its searches to a single brand's domain using site: operators (Writesonic's citation study). Second, don't assume a disappearing citation tag means less trust in sourcing; the quoted-text format means your content needs the exact phrasing that answers a question, not just the right topic (V7Labs' changelog). Third, keep checking what's actually being cited about you, because authority and syndication patterns still vary wildly by platform, as the skew toward one Egyptian syndication site in the GPT-6 Astra topic itself shows (GEO Toolbox's analysis).

A free scorecard like Sited's AI visibility scorecard is a reasonable starting point if you want to see where your own brand pages currently stand before Astra decides to open them directly.

Frequently asked questions

Does GPT-6 Astra cite fewer sources overall?

No. Despite running 40% fewer searches and reading fewer results per answer, the study found the answers remain about as citation-dense and point to brands about as often as before (Writesonic's citation study).

What is the "quoted source" citation format?

Instead of a citation tag pointing to a link, GPT-6 Astra attributes a source by showing the quoted text behind the answer, so you check provenance by reading the response rather than clicking a marker (V7Labs' changelog).

Can GPT-6 Astra still fabricate citations?

Yes. A research-use guide warns that fabricated citations and non-existent references remain a known risk for all LLMs including Astra, and recommends manually verifying every citation before relying on its output (Layer3 Labs' guide).

Why does Astra search a single brand's domain so often?

Three out of four of Astra's searches use Google's site: operator to pull pricing, plans and feature pages from one company's own domain at a time, and another 8% skip search and open a known brand URL directly (Writesonic's citation study).

Does domain authority still matter under GPT-6 Astra?

Yes. A broader Writesonic analysis of the top 200 most-cited domains in AI Overviews found sites with a domain rating of 88-100 received more than 6,000 average citations, showing authority still strongly predicts citation volume in the environment Astra operates within (Writesonic's domain-authority analysis).