ChatGPT’s premium model reads your actual website before citing you; its free model doesn’t even visit, and the two share only 7% of sources.
Writesonic ran 50 prompts across ChatGPT’s two current models and extracted every fan-out query and citation. The headline number: GPT-5.4 Thinking (the premium model) sends 56% of its citations to brand websites. GPT-5.3 Instant (the free default) sends 8%.
That’s not a rounding error. That’s two completely different information architectures wearing the same interface.
What’s actually happening under the hood
The screenshot making the rounds on LinkedIn shows GPT-5.4’s fan-out queries for a recruiting software comparison. Thirteen sub-queries total. The first few are broad discovery searches, the kind you’d expect: “best applicant tracking system AI features,” “G2 ATS 2026 Greenhouse Lever Ashby.” Normal stuff.
Then something shifts. Queries 4 through 12 are all site: searches. site:greenhouse.com AI features. site:ashbyhq.com AI features. site:lever.co AI recruiting. One by one, the model walks through each brand’s actual website looking for specific product information.
GPT-5.3 doesn’t do this. It fires one or two broad queries, pulls from whatever third-party listicles rank well, and builds its answer from there. The premium model identifies the brands through its initial discovery queries, then goes to each brand’s site to verify claims and pull details directly.
Writesonic found that GPT-5.4 used site: operators in 156 out of 423 total queries. GPT-5.3 used zero.
Two models, two visibility games
This creates a genuinely strange situation. The same product, asked about by two different users, gets described using completely different sources depending on which ChatGPT tier they’re paying for.
If someone on the free plan asks “what’s the best CRM for a 50-person sales team,” ChatGPT pulls from TechRadar, Forbes, Tom’s Guide. The brand’s own website barely factors in. Your visibility depends on third-party coverage.
If someone on the paid plan asks the same question, ChatGPT goes to each CRM’s actual website. It reads pricing pages, feature pages, comparison pages. Your visibility depends on what’s on your own site.
The two models shared only 7% of their cited sources across the same prompts. Practically zero overlap.
The pricing page problem
Here’s where it gets concrete. GPT-5.3 cited 4 pricing pages across 49 conversations. GPT-5.4 cited 138.
If your pricing page says “contact sales,” GPT-5.4 has nothing to work with when a user asks it to compare your pricing against competitors. It won’t guess. It won’t fill in the gap with a generous assumption. It moves on to the next brand that actually publishes its numbers.
Every B2B SaaS company that gates pricing behind a form is now invisible to the model that actually reads their website.
What this means for the “GEO” conversation
The emerging discipline of optimizing for AI search just got more complicated. There isn’t one game to play. There are at least two, and they reward different things.
For GPT-5.3 (the model 95% of users interact with), the playbook looks a lot like traditional digital PR. Get mentioned on review sites. Earn spots in roundup articles. Build a presence on G2 and Capterra. The model trusts third parties more than it trusts you.
For GPT-5.4, the playbook is your own website. Clear product pages. Published pricing. Feature comparisons written in plain language that a model can parse. Structured content that answers the exact questions buyers are prompting.
Philipp Götza made a sharp observation in the replies to Chris Long’s original post: this behavior is specific to GPT-5.4 Thinking. The reasoning layer seems to drive the site: search behavior. If and when that reasoning capability trickles down to the free tier, the rules change for everyone. But right now, the free model and the premium model are playing entirely different games.
The uncomfortable implication
There’s a class dynamic buried in this that nobody’s really talking about. The users most likely to be paying for ChatGPT Plus are also the users most likely to be making purchase decisions at companies. They’re the ones comparing CRM platforms, evaluating recruiting software, researching marketing tools.
Those users are getting answers sourced directly from brand websites. Everyone else is getting answers sourced from whatever third-party content ranks well on Bing.
If you’re a B2B brand, the model that matters most for your pipeline is the one that’s actually reading your site. And right now, that model is judging you on content most companies treat as an afterthought: the pricing page, the feature comparison, the product overview that hasn’t been updated since last quarter.
The brands that figure this out first won’t just show up in AI search. They’ll show up in the version of AI search that paying customers use.