A few weeks ago, Google Research published a paper that tells you almost everything you need to know about where search is headed, and it has nothing to do with Google losing to ChatGPT. The paper is about how Google is rebuilding the engine behind AI search itself, and it points to an algorithm that is about to get a lot harder to trick and a lot better at finding content that is genuinely useful. If your marketing still leans on tricks instead of substance, consider this your warning shot.
Why Would I Care About Google? Because AI Search Runs on It
I hear some version of this question all the time: “Nathan, why would I care about Google ranking? Everyone’s searching in ChatGPT now.” Fair question. It is true that Google does not have the single smartest AI model on the market, even with Gemini 4 closing the gap fast. But “smartest model” and “best search engine” are not the same claim, and mixing them up is costing people real visibility.
Google has a 25 year head start on retrieving and ranking information at scale, and it is pouring more resources into that specific problem than any other lab on earth. Every other company building a search or AI-search product, OpenAI included, is working from a playbook Google wrote first. If you want ChatGPT, Perplexity, or Gemini to recommend your practice, you need to show up in search, because that is largely where these systems go to find an answer. Skip Google and you are not skipping a step. You are skipping the step everything else depends on.
That is why this paper matters, even if you have never heard the term “query fan-out” before today.
What Query Fan-Out Actually Is
When you search Google, or ask ChatGPT a question, the system does not just match your words to a page. It runs something called query fan-out first. The model takes your original question and generates a batch of related sub-questions, so it can pull a wider, more diverse set of results instead of 10 pages that all say the exact same thing.
That part works fine in theory. In practice, Google’s own research flags 2 real problems with it.
Problem 1: fast and shallow, or slow and good, pick 1
Unless a model is allowed to “think” for 10 or 20 seconds per query, its fan-out questions tend to cluster too close together. The model struggles to cover the real semantic range of a topic quickly. Give it more thinking time and the results get better, but 10 to 20 seconds of compute per search is far too slow and far too expensive to run at Google’s scale. Researchers call the lazy version of this failure “paraphrastic collapse”: the model just spits out synonyms of your question instead of genuinely different angles on it.
Problem 2: 10 results, 2 actual ideas
The second issue is that older systems score each result one at a time instead of judging the whole set together. That is how you end up with 10 search results where 5 of them are saying the identical thing in slightly different words. Google’s current algorithm already patches this with re-ranking after the first pass, but the new paper lays out a better way to solve it from the start: judge the group, not just the individual.
Google's Fix: Swap the Thinker for A Sprinter
Here is the part of the paper that is actually new. Google built a diffusion model to handle this instead of a traditional large language model.
A quick, non-technical definition: a diffusion model starts with something close to random noise and refines it, pass by pass, into a finished output. That is how the original AI image generators worked, Midjourney and Stable Diffusion both run on this idea. Diffusion models are fast and efficient at generating a whole collection at once. They are a poor fit for language or code, where you usually want a model reasoning sequentially, line by line. But for producing a full set, like a set of search results, a diffusion model turns out to be the better tool for the job.
Google’s actual framework is called Retrieve-for-Train, or R4T, and it runs in 3 stages. First, a fan-out language model, built on Gemma3-4B or Qwen3-4B, gets trained with reinforcement learning to generate genuinely diverse, property-aligned sub-queries. Second, that trained model’s behavior gets converted into training data. Third, a much smaller diffusion model, just 53.9 million parameters, learns to copy that behavior directly, skipping the slow step-by-step generation entirely.
The payoff is real. Google reports the diffusion approach runs 12 to 20 times faster than the old autoregressive method. The old approach stretched out toward 50 seconds of fan-out latency at a large batch size. The diffusion version holds that same job to roughly 4 seconds. That is the difference between “technically possible” and “something you can actually run on every search, every time.”
You can read Google’s own writeup here: Bypassing Inference Bottlenecks: Accelerating Complex AI Search with Retrieve-for-Train, and the full research paper here: Efficient, Property-Aligned Fan-Out Retrieval via RL-Compiled Diffusion.
Is this Actually Live in Google Search Right Now?
Honestly, we do not know, and neither does anyone outside Google. The research paper itself first went up on arXiv back in March 2026. Google did not publish its own blog post walking through it until September 15, 6 months later, which is a strange gap if the goal was just publicity. Search Engine Journal’s reporting on the announcement points out something worth noting: Google calls the system “production-ready” in the blog post, but the underlying paper includes caveats that did not make it into that same blog post. Google has not confirmed this is live inside Search today.
What is clear, and this is the part that actually matters for you, is that Google’s algorithm is improving fast. Since the first search engines existed, the direction has never changed: it keeps getting harder to trick the algorithm. That trend moved slowly for about 20 years. It is accelerating now, and I think that pace only speeds up from here. In 1 or 2 years, we could be looking at a search system that is extremely difficult to game and, as a direct result, extremely good at surfacing content that is genuinely diverse and genuinely helpful.
Plenty of SEO tricks still work today. But in just the last 6 months, a lot of them have stopped working, or started working a lot less well than they used to.
What to Actually Do About It
Stop trying to outsmart an algorithm that is specifically being rebuilt to catch what you are doing. Focus on content that is helpful to an actual human being searching for an actual answer. A few ways to put that into practice starting this week.
Answer the real question in the first few sentences of the page, not after 3 paragraphs of throat clearing. If someone had to scroll to find your point, a system built to reward diverse, information-dense answers is not going to favor you.
Make every section on the page stand on its own. Define your terms where you use them instead of making the reader hunt for context 2 sections back. A system designed to judge a set of results for coverage and complementarity is, by design, going to reward pages that are not redundant with everything else already ranking.
Cite real numbers and real sources instead of hedged, vague language. “Studies show” is not a source. A system trained to reward grounded, well-supported answers treats specificity as a signal, not decoration.
Stop publishing 5 near-identical pages targeting slight keyword variations of the same topic. That is precisely the kind of repetitive, low-diversity content this new system is purpose-built to filter out, and honestly, it was never a good strategy even before this paper came out.
Build out a real FAQ section with the actual follow-up questions people ask, not a generic list bolted on for the schema credit.
None of this is a trick. That is the point.
Frequently Asked Questions About Google Ranking
Query fan-out is the process search engines and AI chatbots use to turn 1 search into several related sub-questions, so they can retrieve a broader, more diverse set of results instead of a handful of near-identical pages.
R4T, short for Retrieve-for-Train, is a Google Research framework detailed in an ICML 2026 paper. It trains a small, 53.9 million parameter diffusion model to generate query fan-outs in a single fast pass, instead of the slower, more expensive method older AI systems use.
Google has not confirmed this. The company’s blog post calls the framework “production-ready,” but as of this writing, there is no public confirmation it is actively running inside Google Search or AI Overviews.
No. If anything, this research is evidence Google is investing heavily in making its own search and retrieval smarter, which means showing up in traditional search results stays relevant to showing up in AI answers too. The practices that are dying are the shortcuts, not SEO itself.
Lead with a direct answer, make each section independently useful, cite real sources and real numbers, and stop publishing near-duplicate pages aimed at slightly different keywords. Build for a reader first. The algorithm catching up to that standard is Google’s problem to solve, not yours.
Will Your Content Stand Up to Google's Next Evolution?
Google has had one job for 25 years: find the best answer and surface it fast. Every update, this one included, is just Google getting closer to actually doing that job. The practices that win from here are the ones that were always supposed to win: be the clearest, most complete, most honestly useful answer on the page. The tools for finding that answer just got a lot faster.
Want a straight read on whether your current content would hold up under something like this, or get filtered out with it? Arcane Marketing runs AI-search visibility checks for healthcare practices that actually want to know where they stand, not another vague audit PDF. And if you want the next one of these breakdowns before it turns into somebody else’s hot take, follow Arcane Marketing’s channels; links are on the site.
Is Your Website Built for the Future of AI Search?
Get a clear picture of where your practice stands, what is holding it back, and what needs to change to compete in AI-powered search.