Citations are the new ranking

Type "best running shoes for flat feet" into ChatGPT. Read what comes back.

Now ask Claude the same question. Then Perplexity. Then Gemini.

If you do this exercise across four or five different AI systems, you’ll find that the paragraph you get is broadly similar - the same five or six brands, the same caveats about consulting a podiatrist, the same vague middle ground. But the citation chips drift. One answer might cite Brooks’s own website, Runner’s World, and a Reddit thread. Another might cite Wirecutter, a sports-medicine clinic, and a different forum. The cited entities are not stable across providers, across days, or even across reruns of the same prompt.

That shifting cluster of citations is the new search results page.

The slot itself

A blue link, for two decades, had a simple economic logic. You wanted to be #1 because being #1 got you clicked. Being #4 got you maybe a tenth as many clicks. Being #11 got you almost nothing. The whole industry of SEO existed to push your URL up the ladder one place at a time.

A citation chip in an AI answer is structurally different in a way that takes a minute to absorb.

  • A citation can be visible to the user without ever being clicked. The reader sees "[1] Runner’s World" sitting next to a recommendation and walks away with the impression that Runner’s World endorsed it. That impression has commercial value even if nobody clicks.
  • There are usually three to six citations per answer, not ten. The bar is higher; the slot is narrower.
  • The citations are attached to specific sentences, not to the answer as a whole. Being cited in the "best for plantar fasciitis" sentence is different - sometimes more valuable, sometimes less - than being cited in the "general recommendation" sentence.
  • The ranking is determined by a system you can’t inspect, parameterised by signals you don’t know, run by a provider whose incentives may not align with yours.

In other words: the value of a citation is partly impression-based (your brand appears in an answer), partly intent-based (the answer is being shown to a person who already wanted to buy running shoes), and only marginally click-based.

That last point is the one I keep finding people don’t process the first time they hear it. The citation is doing work even without the click. This is closer to the economics of being mentioned in a newspaper article than the economics of being a #1 search result.

How do AI systems pick what to cite?

I don’t know - exactly. Nobody outside the providers does. But the previous piece laid out the architecture, and we can reason about which stages constrain the citation set.

A citation can only appear if:

  1. The source was in the index that was searched. If the AI provider didn’t crawl your site, or excluded it, or hasn’t re-crawled since your content moved, you’re not even in the candidate pool. Corpus inclusion is the first gate.
  2. The retrieval stage returned the source as a candidate. Lexical match on the query terms, embedding similarity in semantic space, hybrid score above some threshold - the same retrieval engineering that decides #1 vs #11 on Google decides whether your page makes it into the top-K passages here.
  3. The reranker placed it high enough to fit in the model’s context. Top-K is small. Reranker quality matters a lot at the margin.
  4. The model chose to actually cite the passage when synthesising the answer. This part is more inscrutable - it depends on how the prompt was constructed by the provider, on the model’s own calibration, on the way the system prompt requests citations.

Steps 1-3 are mostly the search-engineering you’ve been able to influence for twenty years, just with different providers. Step 4 is genuinely new and genuinely opaque.

What changes for brands

If you sell something - products, content, services, expertise - your strategic question has quietly mutated.

The question used to be: do we rank in the top three for this query?

The question now is: when a user asks an AI assistant a question we want to be in the answer to, do we appear in the citations?

These look similar from a distance but they don’t reward the same things. The old game rewarded inbound links, content depth, on-page SEO, page speed, and a hundred other crawler-visible signals. The new game rewards being in the embedding neighbourhood of useful queries, being cite-worthy in the sense that a model can pull a clean factual sentence out of you, and being included in the corpus of every provider you care about.

I don’t think most of the brands I talk to have noticed this. They’re still optimising for the previous game. The teams measuring "where do we rank for our top 100 queries" are usually not also measuring "are we mentioned when ChatGPT, Claude, Perplexity, and Gemini are asked our top 100 queries." The first is a solved tooling problem; the second is barely tooled at all.

That’s about to change. There’s a new category of measurement and visibility tooling that has to exist, because the new currency - brand mentions across AI answer surfaces - is what teams will need to know about.

The question I keep coming back to

If the visible surface is shifting from "the ten links that got the click" to "the three citations that got the mention", and if the mechanics of which citations win are mostly opaque to the brands that depend on them, what happens to the businesses that built themselves on the old surface?

I want to answer that one with numbers, not just speculation. So the next post is going to be about traffic. Specifically about where it has gone.