Query fan-out is the process Google and ChatGPT use to turn one search into a set of related sub-queries, so the AI can pull a wider range of results before it answers. In September 2026, Google Research published new query fan-out research showing a faster method built on a diffusion model. In Google’s tests, the diffusion model ran 12 to 20 times faster and returned more diverse results than the current method.
For SEO, Google’s research means search engines are about to get much harder to trick. At TJ Digital, we run AI search campaigns for 40+ clients, and in just the last six months, I’ve watched a lot of SEO tricks become less effective.
I’ve been doing SEO for 17 years, and the approach we use for clients is simple. We build pages that help people with what they’re actually searching for.
What is a query fan-out?
Query fan-out is when an AI search tool takes your prompt, generates a series of similar queries, and runs each one to find a diverse set of search results. Google’s own example is a search for camping gear. A useful answer includes a tent, a sleeping bag, a stove, and a headlamp, and ten slightly different four-person tents would be a bad answer.
Google’s AI Mode and ChatGPT both run query fan-out before they recommend anything. I cover the bigger picture in my breakdown of how AI search works. The short version is that every sub-query is another chance for your page to get pulled into the answer.
Does ranking in Google still matter for AI search?
Yes. Google is far ahead of everyone else in search and information retrieval, and if you want to get recommended by AI, you need to show up in search. Anyone else building a search engine is mostly copying Google, which makes Google’s research the best preview we have of where search is headed.
Other labs are ahead on raw model intelligence right now, so I understand why some people think Google is on its way out. Google already had a huge head start in search, and from what I can see, it’s now investing more resources in this technology than any other lab. Its researchers also invented the Transformer, the architecture behind ChatGPT and most modern language models.
Why is AI search slow and repetitive right now?
AI search is slow and repetitive because current models write near-duplicate sub-queries unless they think for a long time, and they judge each search result on its own. Google’s paper opens by laying out both problems.
Why are AI fan-out queries so similar?
AI fan-out queries are similar because a model that doesn’t think for long tends to write near-duplicates, a problem Google calls paraphrastic collapse. In Google’s example, a standard model given “bohemian festival style” might come back with “bohemian festival fashion” and “bohemian festival clothes,” which would pull up almost the same results.
Ideally, the model finds queries that stay closely related to your search while being as diverse as possible, so it uncovers information you didn’t think to ask about. Doing that well with current models takes 10 or 20 seconds of thinking, which is far too slow and expensive for Google search. Google’s own tests found the standard approach took nearly 50 seconds at large scale, and a search bar needs to respond in under a second.
Why do AI search results repeat the same information?
AI search results repeat the same information because the AI judges each search result on its own. You might get 10 recommendations, and five of them say essentially the same thing.
Google’s current algorithm handles this to some degree by re-ranking pages after it finds the initial set. A better approach is to judge each possible set of results as a whole. Diversity only exists across a group, so you can’t measure it one page at a time.
How does Google’s new diffusion model work?
Google’s new diffusion model generates a full set of search results in one pass. It starts with a complete output that’s mostly random, then keeps refining that output until it reaches the final result. Google calls the framework Retrieve-for-Train, or R4T.
The original AI image models, like Midjourney and Stable Diffusion, work the same way. Large language models like ChatGPT and Gemini generate text one token at a time.
Why use a diffusion model for search?
Diffusion models tend to be faster and more efficient than language models. They’re a poor fit for language or code, where you want the model to work through the problem line by line. For generating an entire collection at once, like a grid of pixels or a set of search results, they might be the better choice.
How did Google train its new fan-out model?
Google built its system in three steps. First, it used reinforcement learning to teach a language model what a good set of sub-queries looks like, scored on three things:
- Groundedness keeps every sub-query tied to something real in the database.
- Diversity is measured across the whole set, which pushes the model to cover different angles.
- Alignment keeps every sub-query connected to the original search.
Next, that trained model generated its own training examples, with no human labeling required. Finally, Google compressed the behavior into a small diffusion model with 53.9 million parameters that produces the whole set in a single pass. In testing, it ran 12 to 20 times faster than the step-by-step approach and beat standard search, basic query expansion, and a heavily optimized baseline on result quality.
How does diffusion fan-out compare to current AI search?
Diffusion fan-out is faster, produces more varied sub-queries, and judges results as a full set. Here’s how the two approaches compare:
| Factor | Current AI fan-out | Google’s diffusion approach |
| How sub-queries get created | A language model writes them token by token | One model generates the full set in a single pass |
| Speed at large scale | Up to nearly 50 seconds | Under a second to a few seconds |
| Variety of sub-queries | Often near-duplicates without long thinking time | Trained to cover distinct angles |
| How results get judged | One result at a time | The full set together |
| What searchers likely see | Repetitive recommendations | More varied recommendations |
Is Google using this in search results yet?
We don’t know if Google has deployed this technology in its search results yet, or if it ever will. The published tests used a fashion dataset and a music playlist dataset, and Google’s write-up says nothing about Google Search specifically.
Still, I think a rollout is likely. Google’s algorithm is improving fast, and this research targets the exact problems that make AI search slow and repetitive today.
Is Google getting harder to trick?
Yes. Since search engines first appeared, it has gotten steadily harder to trick the algorithm, and that trend is speeding up.
The trend moved slowly over the last 20 years. It’s moving much faster now, and I expect the pace to keep accelerating. In one or two years, we might be looking at a search algorithm that’s much harder to trick than today’s and much better at surfacing diverse, high-quality information.
Do SEO tricks still work in 2026?
Plenty of SEO tricks still work in 2026, but a lot of them have become less effective in just the last six months. I’d count on that trend continuing.
My read is that set-level evaluation will make repetitive content even less useful. If Google judges results as a group, a page that repeats what five other ranking pages already say has little chance of making the set.
How should you optimize for query fan-out?
To optimize for query fan-out, create content that’s actually helpful for your users and matches what they’re actually searching for. That lines up with Google’s own guidance on people-first content, and it’s where this research points.
Here’s what that looks like in practice:
- Start with what your customers search for. Build pages around the questions they type into Google and ChatGPT.
- Answer the follow-up questions. Fan-out means AI runs several related searches for every prompt, so pages that answer specific sub-questions get more chances to be cited. After months of testing, our approach is to build pages that answer sub-queries better than what’s currently ranking.
- Add something the other results don’t have. Your own data, examples, and experience give a set-based algorithm a reason to include your page.
- Stop trying to trick Google. Keyword stuffing and near-duplicate pages for every keyword variation are the kinds of tricks that keep losing ground.
How can you keep up with Google search changes?
The best way to keep up with Google search changes is to follow the research Google publishes and the people who break it down. Papers like this one are the earliest signal of where search is headed.
I post a video every day on YouTube and TikTok sharing what I’m learning about AI search and SEO. For SEO and AI news like this, I also recommend joining Marie Haynes’ Search Bar community.
What else should you know about Google’s fan-out research?
What is paraphrastic collapse?
Paraphrastic collapse is when an AI model writes sub-queries that are just rewordings of the original search. The result is a set of search results that all say roughly the same thing.
What is Retrieve-for-Train (R4T)?
Retrieve-for-Train is Google’s framework for teaching a model good fan-out behavior offline, then compressing that behavior into a fast diffusion model. The heavy reasoning happens during training, so live searches don’t have to wait on it.
How much faster is Google’s diffusion retriever?
Google reports a 12 to 20 times speedup over standard step-by-step fan-out. At its largest test scale, the older approach took nearly 50 seconds, while the diffusion model finished in under a second to a few seconds.
Does ChatGPT use fan-out searches too?
Yes. ChatGPT and Google’s AI Mode both break a prompt into multiple searches before answering, so the same content principles apply to both.
How do you get your business recommended by AI?
You get your business recommended by AI by showing up in the searches AI runs when your customers ask for what you sell. At TJ Digital, every AI search campaign is built on data from AI visibility tracking, search performance, and competitor research. We track where AI recommends you, then build the pages and third-party mentions that fill the gaps.
For one client, shifting to topics we found through AI search query analysis doubled organic traffic in one month. Contact TJ Digital for a free digital marketing audit, and we’ll show you where your site stands in AI search and what will get you more leads.