Seer Interactive ran 346 matched prompts through GPT-5.5 and GPT-5.6 Luna, the model ChatGPT now uses by default. Luna ran almost the same number of searches but aimed most of them at specific websites. It looked at 43% fewer sources and gave 12.9% fewer citations.
Table of contents
Key takeaways
- Luna limited 76% of its search queries to a single website with the site: operator. GPT-5.5 did so in 21.6%.
- On prompts that named no brand, that share rose from 20.6% to 77.3%. Luna picked the site itself.
- Open searches, where any site could show up, fell from 575 to 175.
- Sources consulted fell from 6,134 to 3,481, and citations in answers from 2,054 to 1,790.
- If your citations dropped after the switch, the model change alone may explain it.
What this is about
In a September 24, 2026 post, Seer Interactive compared how two ChatGPT models search the web. The newer model, Luna, sends more of its searches straight to chosen websites. For a brand that ChatGPT does not already search directly, there are fewer open searches left to appear in.
What the source says
Seer tested 346 prompts from six industries on both models. It tracked each step of the search process. First come the fan-out queries, the searches ChatGPT runs behind the scenes to answer one prompt. Then the sources it reads. Last are the citations it shows in the answer.
Seer’s reading is that Luna “is getting better at finding information and needs fewer resources”. That is Seer’s interpretation, not something the counts prove. The counts show that Luna searched in a narrower way. Whether its answers got better is a separate question the study did not test.
What the data shows
| Measure | GPT-5.5 | GPT-5.6 Luna |
|---|---|---|
| Calls that used web search | 80.8% | 85.2% |
| Fan-out queries per search | 3.63 | 3.42 |
| Total fan-out queries | 733 | 729 |
| Queries using site: | 158 (21.6%) | 554 (76%) |
| Open queries (no site:) | 575 | 175 |
| site: share, non-brand prompts | 20.6% | 77.3% |
| Third-party sites on brand prompts | 26.6% | 24.4% |
| Sources consulted | 6,134 | 3,481 |
| Citations in answers | 2,054 | 1,790 |
Two changes cancel out. Luna searched on more calls but ran fewer queries each time, so the total barely moved: 733 against 729. The real shift is in where those queries point. Three in four now go to one named site.
On prompts about a brand, the share of third-party sites stayed close, at about a quarter. Even then, ChatGPT reads beyond the brand’s own website.
One detail does not add up. At 3.63 queries per search and 733 queries, GPT-5.5 made about 202 searches. At an 80.8% search rate, that implies about 250 calls, not 346. Luna’s figures give the same result. Seer does not explain the gap, or say whether some prompts were left out.
On the seven method points, Seer states the models, the prompt count and the industries. It gives no dates for the test and no region. It does not say whether each prompt was run once or several times. That matters, because ChatGPT gives different answers to the same prompt on different runs. Seer’s post also dates Luna’s launch as the default to July 9, 2025, a year that looks like a typo.
What it does not mean
It does not mean ChatGPT stopped using the open web. Luna still ran 175 open searches and used search on more calls than before.
It does not mean third-party coverage matters less. On brand prompts, a quarter of sources still came from other sites, about the same as before.
It does not show that Luna’s answers are better. Seer measured how the model searched, not how accurate its answers were.
It is not a general law. It is one agency’s test of 346 prompts. A different prompt set could shift the numbers. For how citations work in general, see our guide to how answer engines cite.
What to do
1. Mark the model switch in your reports. If ChatGPT citations fell around the change, note the date. A drop may come from the platform, not from your work. 2. Look at fan-out queries, not only citations. If your tracking tool shows the searches ChatGPT runs, check whether it searches your domain with site:. 3. Sort those site: searches by topic. Find which topics send ChatGPT to your site and which send it to a competitor. The second list is your gap. 4. Keep working on coverage elsewhere. A quarter of sources on brand prompts are third-party. What others say about you still reaches the answer, as our piece on what AI does when it does not know your brand showed.
FAQ
What is a site: search?
It is a search limited to one website, such as “site:example.com pricing”. Luna used it in 76% of its searches, so most of its searching went to sites it had already chosen.
My ChatGPT citations fell this summer. Is it my content?
Maybe not. In Seer’s test, Luna gave 12.9% fewer citations than GPT-5.5 on the same prompts. Check whether the drop lines up with the model change before you change your strategy.
How reliable is a test of 346 prompts?
It is enough to show a large shift, like 22% to 76%. It is not enough for small ones. Seer also does not say whether it ran each prompt more than once, and ChatGPT’s answers vary between runs.
