What Is Selection Rate?

How Selection Rate Works

A generative engine builds an answer in stages. It interprets the prompt, runs a query fan-out that issues several background searches, retrieves a set of candidate documents, and then synthesizes a reply that stitches together passages from a handful of those candidates — usually with citations. Selection rate measures what happens at that last stage, and only that stage.

The key idea is that retrieval and selection are two different events. Retrieval decides whether your page is even eligible to be quoted — whether it clears crawlability, indexing, topical relevance, and grounding into the engine’s live index. Selection decides whether, out of everything the engine pulled, it actually reaches for your passage when it writes the answer. Most SEO instinct is aimed at the first event. Selection rate is aimed squarely at the second.

This matters because the two failure modes need opposite fixes. If you are never retrieved, you have a visibility problem: the page is not crawlable, not indexed, or not on-topic, and the fix is technical and topical. If you are retrieved constantly but rarely selected, you have a synthesis problem: the page is eligible but not quotable, and the fix is extractability — writing self-contained, attributable passages the model can lift without inheriting risk. Selection rate is the metric that tells those two situations apart. A high average position or a healthy retrieval count can hide a selection rate near zero, which is exactly the case where teams burn months building more links that never move a citation.

Optimizing this number deliberately is what some practitioners call Selection Rate Optimization (SRO): once a page is reliably in the candidate set, make it the single easiest, best-sourced, most self-contained option the engine can quote for that prompt.

Formula

Selection rate is a conditional rate. The denominator is not every prompt in the world — it is only the prompts where your page was actually retrieved as a candidate:

selection rate = times cited ÷ times retrieved (eligible)

So a page retrieved for 200 prompts and cited in 40 answers has a selection rate of 40 ÷ 200 = 20%. The conditioning is the whole point. It strips out the retrieval question entirely and asks a single thing: given that the engine had you in hand, how often did it choose you? That is why a page can be retrieved far more often than a competitor and still lose — it is winning the retrieval race and losing the selection race.

Two anchors from adjacent research show this stage is real and lossy. In classic RAG evaluation, “chunk utilization” or attribution tracks how many retrieved passages are actually used in the generated answer, and it is common for only a fraction of retrieved chunks to be cited — often on the order of two out of five (Braintrust RAG metrics). And a foundational study of generative search engines, Evaluating Verifiability in Generative Search Engines (Liu, Zhang & Liang, 2023), measured citation precision and recall across engines and found large gaps between what is relevant and what is actually cited — the same retrieved-but-not-selected gap, measured at the engine level.

Example of Selection Rate

Selection rate is a newer, proprietary-leaning metric, so the following is a deliberately illustrative example with round numbers, not a measured case study — it exists to show the arithmetic, not to report a real result.

Suppose you track one page across a fixed set of 100 buyer-intent prompts in an AI answer engine. Over a month, the engine retrieves that page as a candidate for 100 of those prompts — it is eligible every time, because it is crawlable, on-topic, and well grounded. But it is actually cited in the written answer for only 22 of them.

Its selection rate is 22 ÷ 100 = 22%. Retrieval is a solved problem here; selection is not. Roughly four out of five times the engine had this page available and chose something else.

Now suppose a competitor is retrieved for only 60 of the same prompts — worse eligibility — but gets cited in 30 of those answers. Their selection rate is 30 ÷ 60 = 50%. Despite weaker retrieval, they win the synthesis stage more than twice as often, because when they are in the set they are the cleaner, more quotable option. The raw citation counts (22 vs 30) understate the gap; the selection rates (22% vs 50%) expose it. That is the diagnostic value of the metric — it separates “can’t be found” from “won’t be quoted,” and here the fix is clearly the second.

Selection Rate vs Click-Through Rate

The two metrics look similar — both are ratios about whether something got chosen — but they measure different actors at different stages of completely different pipelines.

Click-through rate is a post-ranking, human-attention metric: a link has already won a position, been displayed, and CTR asks how often people click it. Selection rate is a pre-display, machine-decision metric: a page has already been retrieved, and selection rate asks how often the model quotes it. Confusing the two leads teams to optimize titles and snippets for a stage that, in an AI answer, may never involve a human clicking anything at all — the zero-click reality of generative search. In classic SEO, CTR is where you compete for the click. In generative engine optimization, selection rate is where you compete for the citation, and it sits alongside citation share as the pair of numbers that actually describe AI visibility.

Frequently Asked Questions

The Bottom Line

Selection rate splits AI visibility into two questions that used to be blurred together: can the engine find you, and will it actually quote you? Retrieval answers the first; selection rate answers the second, and it is where most GEO effort should go once a page is reliably eligible. Optimize the passage the model has to choose between, not just the page it has to find.