How does ChatGPT choose its citations?
I’m Eddie – a certified SEO professional. If you’ve read my piece on AI search optimisation, you’ll know the headline problem. Someone asks ChatGPT for a recommendation, and it either names your business or it doesn’t. What I didn’t cover in that blog is how it decides which sources make the cut, anddd which don’t. It’s worth knowing though, since it will affect your approach to “optimising for AI search”, and what it means for a small business.
The formula behind the citations
Earlier this year, SEO analyst Metehan Yeşilyurt inspected the network requests ChatGPT’s browser makes via Chrome’s DevTools. He found a specific method at work: Reciprocal Rank Fusion, or RRF. It isn’t new – as is the case with much of AI. RRF was first described back in 2009 as a way of merging results from different search methods into one ranking. Don’t get me wrong though – seeing it inside a live consumer AI product is notable.
Here’s the plain-English version. When you ask ChatGPT something, it doesn’t run one search. It typically breaks your question into several related queries and runs each one separately, a process known as ‘query fan-out.’ Each of those searches comes back with its own ranked list of results. RRF merges those separate lists into a single, final ranking. Every source gets a score based on how high it appeared across the different lists. A source that shows up near-ish the top of multiple lists scores higher than one that only ranked well once. Think of it this way – getting a third of a pizza four out of five times (0.33*4=1.32) beats getting half a pizza two out of 5 times (0.5*2=1) – the first scenario gets you 33% more pizza! (And that’s kind of how RRF works – being there multiple times matters more than winning once or twice).
Showing up well for one exact phrase isn’t enough. RRF rewards a source that keeps reappearing near the top across several related searches on the same topic. That’s really just a way of looking at topical authority – and it’s harder to game.
Why “being retrieved” matters more than “ranking first”
This is the part most small businesses miss. RRF can only score the sources that get retrieved into the results in the first place. Independent testing referenced in 201 Creative’s breakdown of the mechanism suggests ChatGPT typically pulls somewhere between 40 and 65 sources per query, before fusing and ranking them. If your page isn’t in that initial pool at all, it scores nothing, no matter how good the content is.
That sets out goalposts. Chasing the #1 spot on one single keyword is far less valuable here. Showing up somewhere in the pool across several related searches has that value. A business ranking a solid but unspectacular 15th–30th on three or four related queries will could well beat a business ranking #1 on just one. And that’s a very different game to SEO – you’ll know well as someone who uses Google and search tools that very rarely will you ever stray past the first couple results.
A test you can run on your own business today
Pick one core service you offer. Then think of three or four different ways a customer might phrase a search for it. Not just your main keyword, but the variations, questions, and comparisons people actually type. Search each one on ChatGPT (while signed out!) and note whether your business appears anywhere in the answer, not just whether it’s named first. If you show up in two or three of those variations but not all four, that’s a good start. It’s a sign you’re already inside the retrieval pool – which means all you now need is just a broader coverage of the topic.
What this means for your content strategy
Practically, this points towards building out topical breadth, rather than a single, heavily optimised page. Several pages or sections covering different angles of the same core service give search more chances to find you. Think the “what,” the “how much,” the “how long,” the “compared to X.” It’s the same principle behind the FAQ-style sections I build into every client site. Each question is, in effect, its own small chance of getting pulled into an answer.
It’s worth being clear-eyed about the limits here too. Yeşilyurt himself has been upfront that this isn’t a “dominating formula” to reverse-engineer. OpenAI hasn’t published the exact ranking code. The specific thresholds researchers have floated are their own experimental estimates, not confirmed constants. What’s well-established is the general pattern: fan-out, retrieval, then fusion. We don’t know the exact tuning used – but frankly, we don’t need to. We know the method – and better is always better.
Frequently Asked Questions — How ChatGPT Chooses Citations
Does this replace the need for traditional SEO? No – if anything it depends on it. RRF can only fuse and rank sources already inside that retrieval pool. Strong traditional SEO fundamentals — structure, relevance, authority — are still what gets a page into that pool in the first place.
Is this the same as the AI search optimisation (AEO) you’ve written about before? It’s the mechanism behind it. My AI search optimisation post covers the business case and the practical steps. This one explains the specific process that makes those steps work.
Can a small business realistically compete with bigger sites on this? Yes, more easily than with traditional link-building. RRF rewards consistent topical coverage over a single dominant page. A small business willing to build breadth on its own specialist area has a real shot at outscoring a much bigger competitor who’s thin on that exact topic. If you’re interested in getting help here, check out my blog post on my copywriting service.
How do you apply this for client sites? It’s built into the same SEO and AI search groundwork I already do for every project. That means structured FAQ content and consistent business details. It also means covering the different ways a customer might phrase the same question, rather than treating it as a separate, bolted-on service.
Curious Where You Stand? Get in Touch with Eddie
If you want to know whether your business is showing up in ChatGPT’s results before you worry about ranking first within them, get in touch for a free, no-obligation discovery call. We can chat through your goals and how I can help, together.