TL;DR
-
We ran a direct comparison of 346 matched prompts to see whether ChatGPT-5.6 Luna behaved differently from GPT-5.5
- GPT-5.6 Luna used web search more often but ran fewer queries, resulting in little movement on the number of fan-out queries compared to GPT-5.5
- Domain-target searching increased drastically (more than 3x), while consulted sources and inline citations dropped
- Our take: Luna is getting better at finding information and needs fewer resources, so brands should pay more attention to the full retrieval process to measure AI search visibility
How Do ChatGPT-5.5 and GPT-5.6 Differ?
We’ve researched ChatGPT-5.5 extensively, including a deep dive into fan-out patterns. Now that GPT-5.6 Luna is the default model as of July 9, 2025, it's the model people get when interacting with ChatGPT and the model brands will be reporting on through AI visibility trackers.
With that in mind, we wanted to know how the two models compared. Specifically:
- How did the retrieval process change across domains and citations?
- Does GPT-5.6 Luna search similarly, or were there notable differences?
- Did fan-out query behavior shift or stay the same?
We compared GPT-5.5 and GPT-5.6 Luna across 346 matched prompts from six industries.
Spoiler alert: At first glance, the two models looked very similar. Across the 346 matched prompts, they ran almost exactly the same number of fan-out queries. The difference is how targeted Luna's searches are. It used the site: operator in 76% of its fan-out queries, up from 22% with GPT-5.5.
That leaves far fewer open searches for everyone else. So if Luna isn't already searching your site directly, your barrier to showing up is higher than it used to be. Check out our top findings below.
1. GPT-5.6 Luna Uses Web Search More Often, But Runs Fewer Queries Per Search
GPT-5.6 Luna turned to web search more often than GPT-5.5. Across the six industries, Luna searched on 85.2% of calls, up from 80.8%. But each time it searched, it ran fewer fan-out queries: 3.42 per search, down from 3.63.
Those two changes almost exactly cancel each other out. Luna searched more often but ran fewer queries each time, so the total number of fan-out queries barely moved. GPT-5.5 ran 733 and Luna ran 729, a difference of just four.
We thought there would be more fan-out queries after the model update, especially given the increase we saw when GPT-5.5 was released. Instead, the total stayed the same, and the change showed up in how Luna searched.
Our hypothesis is that Luna needs fewer queries per search because its queries are more targeted. When a fan-out query goes straight to a specific website (more on that in the next finding), one query can do the job that used to take several broader ones. So the two changes balance out in volume, but not in effect. More prompts now send ChatGPT to the web, and each search covers less ground.
2. Domain-Targeted Searching More Than Tripled, Especially on Non-Brand Prompts
GPT-5.5 had already begun using the site: operator (a search command that limits results to a single website) in fan-out queries to target specific brand sites for answers and, in our opinion, prevent low-quality responses. GPT-5.6 Luna elevated this retrieval method even further.
GPT-5.5 conducted site: searches in about 22% of its fan-out queries, while Luna did so in 76%.
On branded prompts, third-party sites (publishers, review sites, research orgs, etc.) appeared at roughly the same rate with both models: 26.6% with GPT-5.5 and 24.4% with Luna. So even when a user asks about your brand directly, ChatGPT still looks beyond your website.
That means the perception of your brand outside of your website matters just as much as the narrative you share on sources you control.
The actual jump came from prompts that didn't mention a brand. Non-brand prompts saw a rise in site: searches, from 20.6% in GPT-5.5 to 77.3% in GPT-5.6. That means Luna was the one to decide on a domain to search even when the user didn’t specify a brand or website.
Since ChatGPT is getting more targeted with its retrieval, the overall number of consulted sources (the sources ChatGPT pulls in while searching, before it decides which ones to cite) fell 43%, from 6,134 with GPT-5.5 to 3,481 with Luna.
This presents a unique opportunity for brands launching AI search programs. If GPT-5.6 Luna is already targeting your site with site: searches, make sure the information it's looking for is easy for both people and AI to find.
If your brand isn't one of the sites Luna searches directly, you may see fewer citations now that it's the default model. Open searches, the ones any site could appear in, fell from 575 to 175, so getting results will likely take more work.
Look at every stage of retrieval, from fan-out queries to consulted sources to citations, to see how and where ChatGPT finds information about your brand.
3. Inline Citations Fell 13%
We know Luna searched in a more targeted way and consulted fewer sources. In the final answers, it also gave fewer inline citations (the sources linked in the answer itself).
Across all 346 matched prompts, GPT-5.5 produced 2,054 inline citations, while Luna produced 1,790, a decrease of 12.9%. Searching more often didn't translate into more citations.
Our take is that Luna has gotten better at knowing where to find information, so it needs fewer sources to back up its answers.
This matters for any AI search program that uses citations as a KPI or as the success metric for an experiment. If your citations dropped after Luna became the default, that may have nothing to do with your strategy or the tests you have running.
Think of it like a Google algorithm update: the platform changed, and performance shifted with it. Mark the switch to Luna in your reporting so a platform-wide drop doesn't get read as a failed experiment.
So What Should You Do Now?
How can you apply these findings to your AI search strategy? Here are my recommendations:
Look further down the retrieval process when you’re measuring AI search visibility. Most AI visibility trackers only report citations, the bottom line of the retrieval process.
The bigger benefit comes from the layers behind them: the fan-out queries ChatGPT runs, the domains it targets with site: searches, and the sources it consults before deciding what to cite. That's where you'll see why your citations moved, not just that they did.
Pay closer attention to the connections the models appear to make between brands and topics. Luna often decides which sites to search before it retrieves anything, so being recognized as a relevant source for a topic matters before any page gets retrieved.
Run a base set of nonbrand prompts you’ve been monitoring to capture the fanout queries and sort the site: searches by topic. Which topics does ChatGPT search your domain for, and which does it search your competitors' domains for instead?
Compare across models or over time, and you'll see where you're gaining ground and where a competitor owns the association. The topics your competitors own are where you should focus next: both in your own content, and in your coverage on publishers, review sites, and research organizations.
Ready to level up your AI search strategy but don’t know where to start? The Seer team has extensively researched what works and what’s a short-term trend, and we can recommend where to put your resources based on your industry, budget, and team size. Chat with us to learn more.
.png)
.png)