AI search engines work in two stages: retrieval, where the model gathers many candidate pages, and recommendation, where it selects the few it actually cites in its answer. Being retrieved means you’re in the running; being recommended means you won. Studies show only about half of retrieved URLs get cited — so measuring retrieval alone hides whether your content is actually influencing answers.
- ▪AI answers are built in two stages: retrieve candidates, then recommend a few.
- ▪Retrieval means you were considered; recommendation means you were cited.
- ▪Roughly half of retrieved pages never make it into the final answer.
- ▪Tracking only “are we indexed” misses whether you’re actually chosen.
- ▪The work is closing the gap from retrieved-but-ignored to cited.
Marketers are learning to ask whether AI engines can “see” their site. It’s the wrong finish line. An LLM answering a question pulls a wide net of candidate pages into its context, then decides which handful to actually name. Getting caught in the net is easy. Getting named is the whole game — and the two are routinely confused.
This is the AI-era version of ranking on page two: technically present, practically invisible. If you only measure whether you were retrieved, you’ll congratulate yourself for losing.
Two stages, two very different bars
Retrieval is a recall problem: the model gathers everything plausibly relevant, often dozens of sources, to give itself raw material. Recommendation is a precision problem: from that pile it picks the sources that are clearest, most authoritative, and easiest to quote cleanly. A page can clear the first bar on topical relevance alone and still fail the second because it’s vague, buried, or hard to extract a confident sentence from.
| Retrieved | Recommended | |
|---|---|---|
| What it means | Considered as a candidate | Cited in the answer |
| Bar to clear | Topical relevance | Clarity + authority + quotability |
| Visible to the user | No | Yes |
| Drives referral / brand lift | No | Yes |
How wide the gap really is
The headline finding from citation research is blunt: only around half of the URLs an engine retrieves actually get cited. The rest are read, weighed, and silently dropped. So if your AI-visibility tool tells you that you were “surfaced” for a query, that’s a coin flip away from being invisible in the answer the user actually reads.
Closing the gap from retrieved to cited
Winning the recommendation stage is a content-engineering problem. Lead every section with a direct, quotable answer the model can lift verbatim. Back claims with specific numbers and sources, because models prefer citing something concrete. Structure the page with clean headings and schema so extraction is trivial. And build the external authority signals — mentions, links, consistent entity data — that make a model trust you enough to name you.
This is exactly why answer-first, well-structured content wins the AI era — it’s built to be recommended, not just retrieved.
Are you measuring the right half?
Audit how your team tracks AI visibility. If the metric is “appears as a source” without distinguishing retrieval from citation, you’re flying half-blind. Shift the target to citations in the rendered answer, and suddenly your content roadmap has a clear job: turn the coin-flips into wins.