New Research: ChatGPT Tripled It's Fan-Out Queries + Looks For Authoritative Sources
OK SEOs so really excited to share with you all some of the new research that we've been up to at Nectiv. Last year we did two separate analyses that looked at both ChatGPT and Google Gemini fan-out queries to understand the inner workings of how both AI systems search. However, a lot has changed since then. There have been multiple model updates to both LLMs (GPT 5.6 Sol, Gemini Flash 3.5), so we figured it was time to give that research a brand new refresh.
After reviewing the results, it's been VERY interesting to see how OpenAI has made changes to the fan-out query systems power of their AI search models. Without further ado, let's get into it.
A Recap of Last Year's Data
So OK as I mentioned, last year, we performed a study around how ChatGPT searches. With that methodology, we took a very large list of prompts and ran that exact same set of prompts through it's newest model (GPT 5.6 Sol) to and extracted all the fan-out queries associated with them. We also segmented these prompts into popular vertical areas such as software e-commerce travel, so we could understand how fanout queries differ across multiple industries.
As a reminder, here is the data from the study:
ChatGPT Search:
Average Number Of Fan Outs: 2.17
Average Words Per Query: 5.48
Maximum Searches: 4
Common N-Grams: Reviews, Year (2025), Free, Features, Comparison
We wanted to get to the bottom of what exactly has changed since 2025.
Methodology
For this analysis, we replicated the study fairly closely. We took a subset (about 4K) of the exact same prompts that we ran in our study last year and insured the equal representation across all of the different verticals. We then extracted the fan outs queries for those prompts using 5.6 Sol. Fortunately, OpenAI has an API that makes extracting fan-queries and scale very very easy and surprisingly fast.
1. ChatGPT Has Tripled It's Use Of Fan-Out Queries
From using ChatGPT a lot during the past year, one of the things that stood out to us was that it seems to be searching a lot more than it used to. However, we wanted to get data to actually verify this. As a result one of the first things we looked at is the frequency of the number of unique fan out queries from ChatGPT. Well the data certainly backed this up.
You can see the distribution below:
Some key data points include:
Average Fan-Out Number Of Queries: 7.61 (+250%)
Maximum Fan-Out Query: 29 (+625%)
Average Words Per Query: 6.82 (+24%)
The stark finding here is that ChatGPT has more than tripled the number of fan out searches that it's performing. When looking at the distribution 55% of the fan-out queries fell between 6-12 searches. As well, 8 searches was actually the most popular bucket with 19% of all prompts.
So if you've been thinking ChatGPT is searching more than it ever has...the answer if unequivocally "yes".
2. The Maximum Number Of Fan-Out Queries Is Much Longer
As well the fan-out query chains have the potential to be much longer than they've ever been before. Last year the maximum number of fan outs we saw in the data set was just four. This year the maximum number was all the way up to 29 queries in a single chain!!! More regularly, we could see fan-out query chains that exceeded 15 searches in length.
In case you're wondering that prompt was what are the best high knee boots for women wear chat? ChatGPT was searching all kinds of different sites such as UGG, DSW Franco Sarto, and many more retailers.
3. Software, Real Estate And Fashion Have The Most Fan-Outs
We also broke down the searches ChatGPT was doing by different industry. Similar to our previous study we wanted to see if different industries see more aggressive, or less aggressive, searching behavior.
In this data study, "Software" was the most searched industry by a decent margin followed by Real Estate (8.7) and Fashion (8.5). Travel and Credit Card were among the two least search industries averaging less than six searches for per individual prompt. However, this is still a lot more than the average ChatGPT search from last year.
Software queries had an average of 10.7 fan out searches per prompt. I'm guessing this is because these queries take a lot more research for LLMs to confirm things such as features pricey capabilities and a lot more variables for very complex buying cycles.
4. The Unigrams Are Dramatically Different Than Last Year
Finally, we looked at the Ngrams for the data set as a whole. The Ngrams represent unique search terms that ChatGPT used that we're not part of the original prompt. The goal here is to identify the common nomenclature that ChatGPT is using to search. That way marketers can better understand the terms they might need to optimize their sites for.
This is probably where the most surprising insights came from.
Let's directly compare this to the list of extracted Unigrams that we found last year. When you look at the list and compare the two, there is very little overlap between the most common search terms. Also I need to call out how dramatically better our charts have gotten in the past year.
5. ChatGPT Uses "site:" Search In 64% Of Fan-Outs
So first off the clearest difference from 2025 to 2026 is just how clearly ChatGPT is trying to narrow in on truly authoritative and official sources. The terms "site:", official, gov all appear within the top 5 Unigrams. This is clearly one of the ways that ChatGPT is trying to eliminate spam and find sources that it feels are credible, and that it can trust.
ChatGPT is performing a freaking TON of "site:" searches. ChatGPT performs a "site:" search in 64% of all fan-out queries. Often times what we see is it's is narrowing down to a specific set of companies or vendors, and then performing site searches to look for information directly from the source.
Here you can see an example for "what is the best corporate management software", ChatGPT starts with some non-site searches for the first queries but then eventually shifts into looking at key vendors directly on their site.
However, site search doesn't just take place across brand sites. ChatGPT can also perform site search against resources such as official government sites are even the social ecosystem. For example, in this search for "what is the best home healthcare software reviews" ChatGPT performs site searches on both the Medicaid and the CMS government sites directly.
This is critical as ChatGPT is clearly looking for trusted entities and data sources that it can reliably extract context from. The insight for a lot of brands is that by understanding the common authoritative sources that ChatGPT searches against is important for understanding your overall strategy and the aspects of the product that you should be focusing on for your own site.
6. "Official" Is The Second Most Searched Unigram
Another way that ChatGPT is trying to narrow in on authoritative sources is literally just using the term "official" in the actual search itself. This seems to be basically a substitute and work around for a site search or even used in conjunction with it. Here you can see for the prompt "what is the best martial art software" that it tacts on "official" right after a "site:" search for multiple vendors.
It might make sense for companies to start optimizing for "official" terms on pages where it makes sense. This would help them better optimize for the nomenclature that ChatGPT is fan-out query systems are using.
7. Freshness Is Still A Factor + ChatGPT Performs Multi-Year Search
There are also some surprising insights when it comes to freshness/year signals:
They used to be the most popular Unigram, now it's dropped outside of the top 5
ChatGPT is executing multi-year search. It will often search for both "2026" and "2025" queries in an effort to get both the previous and current year.
The multi year search finding is something that we've been seeing Gemini do for a while now. However, last year when we looked at ChatGPT searches, it would only search against the current year. Now it seems like it's adopting a similar technique as Google to capture both the current year and the previous year content as context. So freshness still matters for ChatGPT - maybe just not as much as it once did.
Conclusion
The biggest takeaway from this year's research is that ChatGPT's search behavior has become dramatically more sophisticated. It's not just performing more fan-out queries. It's conducting deeper research, using longer search chains, narrowing in on specific brands and sources, and increasingly looking for information directly from authoritative and official websites. For SEOs, that means optimizing for AI search is becoming less about simply appearing for a broad set of keywords and more about understanding the research journey an LLM takes to answer a question.