What Is LLM SERP?
Written byFlavio Amiel Founder, Roborank
An LLM SERP is the set of external sources an AI answer engine retrieves, quotes, and cites for a given prompt — the generative counterpart of a traditional search engine results page. Instead of a ranked list of ten blue links, it is the handful of documents a system attributes beneath or inside its synthesized answer, exposed as links, footnotes, or numbered citations.
- A classic SERP shows roughly ten ranked links; an LLM SERP is usually a much smaller set — often a few cited sources — surfaced as footnotes, links, or numbered citations under a synthesized answer.
- In the Gemini API the LLM SERP is machine-readable: a
groundingChunksarray lists each cited source’suriandtitle, andgroundingSupportsmaps those chunks to specific sentences in the answer. - Appearing in an LLM SERP requires being retrieved and cited, not merely ranked — a page can sit at position one in Google and never show up among an AI answer’s sources.
- There is no stable ‘first result’ in a synthesized answer, so inclusion matters more than order; citation share replaces average position as the visibility metric.
How LLM SERP Works
A traditional search engine answers a query by ranking documents and handing the user a list — ten blue links, each a separate row to click. A generative answer engine answers differently: it interprets the prompt, retrieves a set of candidate documents, and synthesizes a single answer that stitches together passages from a few of them, usually with attribution. The LLM SERP is that attribution layer — the specific sources the engine names as the basis for what it just said. It is not a page of links you choose among; it is the evidence trail behind an answer the model already wrote.
Two things make it a genuinely different surface. First, it is short. A results page can list ten ranked URLs and a long tail below the fold; a synthesized answer typically cites only the two or three passages it actually used. Second, there is no reliable running order. Because the answer is generated rather than ranked, the “first” citation is not a rank you can defend the way you defend a position-one listing. Inclusion is the win; order is noise. That is why AI visibility is measured with citation share — your percentage of citations across answers on a topic — rather than average position.
Getting into an LLM SERP takes two things in sequence. A page must first be retrievable: reachable by an AI crawler and present in whatever index or corpus the engine draws from, which is where classic SEO still pays off. Then, at synthesis time, the model decides which retrieved passages to actually cite — and it favors text with high extractability, meaning a claim stated cleanly enough to quote without the rest of the page for context. Retrieval gets you eligible; extractability gets you cited.
What the LLM SERP Contains
Depending on the engine, an LLM SERP shows up in different forms:
- Numbered inline citations — Perplexity and similar answer engines footnote each claim with a numbered source you can open.
- Linked source cards — Google AI Overviews attach the pages the summary drew on, often as a cluster of links beside or below the generated text.
- A structured metadata object — in the Gemini API, the LLM SERP is returned as data, not just rendered UI (see the example below).
In every form the unit is a passage-backed source, not a domain. One long article can supply several distinct citations, and the same page can be cited for one prompt and absent for the next.
Example of LLM SERP
The clearest documented illustration is how Google’s Gemini API exposes an LLM SERP as machine-readable data. When Gemini answers with Grounding with Google Search enabled, the response ships with a groundingMetadata object. Inside it, groundingChunks is an array — each entry a uri and a title — that lists exactly the web sources the model used. A parallel groundingSupports array maps spans of the answer text, by start and end index, to the chunks that back them. Read together, those two arrays are the LLM SERP for that answer: the set of cited sources, plus which sentence each one supports.
The structure makes the difference from a classic results page concrete. There is no rank field and no top-ten list. There is a small array of sources and a mapping showing which claim each one justifies. A page either appears in groundingChunks for a given prompt or it does not, and appearing there depends on being retrieved and then quoted — not on where it would have ranked in ordinary search. The same passage-backed logic drives the visible citations under AI Overviews and the numbered footnotes in other answer engines, even when they don’t hand you the raw data.
This is why the discipline of generative engine optimization targets the passage rather than the page. Empirical work on how answer engines assemble their citations — such as the 2025 GEO16 citation-behavior analysis by Kumar and Palkhouski — treats “which sources make it into the answer” as the outcome worth studying, because that set, not a rank, is what a reader now sees. If your sentence is the cleanest available evidence for a claim and a crawler can reach it, you land in the LLM SERP; if it isn’t, a competitor’s does.
The trap I watch people fall into is picturing an LLM SERP as a shorter version of the ten-blue-links page — same game, fewer slots. It isn’t. On a classic results page every ranked URL gets a row of its own, and the user chooses among them. In an LLM SERP the engine has already chosen: it synthesizes one answer and cites only the two or three passages it actually leaned on. That collapses the funnel. You are no longer competing for a rank in a list a human will scan; you are competing to be one of the tiny number of sources a model is willing to quote and name. A page that ranks well but states its facts loosely loses that seat to a lower-ranked page whose sentence can be lifted cleanly. Optimize to be quotable, not just to be ranked, because the LLM SERP has room for a fraction of the winners the old page did.
Frequently Asked Questions
What is an LLM SERP?
How is an LLM SERP different from a normal SERP?
How do I appear in an LLM SERP?
Can I see the sources behind an AI answer?
groundingMetadata object listing each cited source. That structured list is the machine-readable form of the LLM SERP.The Bottom Line
If a traditional results page answers the question by handing you a ranked menu of links to open, an LLM SERP answers it directly and shows only its receipts — the small set of documents the model chose to quote and attribute. Winning there is not about placing high in a long list; it is about being one of the few sources clean and credible enough for a machine to cite by name.
Sources
Roborank shows you the LLM SERP for your topics — which sources ChatGPT, Perplexity, Gemini and Google AI Overviews cite, and which competitor gets named when you don’t.
See who gets cited →Rank & Cash — the weekly SEO breakdown
One practical teardown a week on ranking in search and getting cited by AI. No fluff.