Ranked first in the pool, cited zero times

The two events people keep confusing

A retrieval engine does two jobs in sequence. First it assembles a pool: it searches, and it puts a list of candidate pages in front of itself. Then it writes an answer, and while writing it decides which of those candidates, if any, to name. Position in the pool belongs to the first job. Citation belongs to the second.

Almost every tool in this category reports the first and lets you assume it is the second. That is not a lie exactly — the first is real, measurable, and it does move — but the two come apart often enough that a page can win the search and lose the answer every single time.

What the numbers say

• 10 pages entered their pool at position one. 5 of them were cited zero times.
• 9 of the 22 cited pages — 40.9% — did not rank in the top three on entry.
• The most-cited page in the whole dataset came in at rank 3 with 2,038 words, and was named in 6/6 runs.
• The longest page in the dataset — 11,582 words, rank 1 — was named in 0/6 runs, and blocks retrieval crawlers at the root.

Being first in the pool is a search result. Being cited is a different event, decided later, by a different process. Treating the first as a proxy for the second is the most expensive mistake in this category.

The extreme case, and why it is instructive

That 11,582-word page is the single most useful row in the dataset, because it kills three intuitions at once. Length does not rescue you. Ranking first does not rescue you. Having structured data does not rescue you either — it has JSON-LD.

What it does have is a closed door: retrieval crawlers are refused at the root. Whatever else is wrong with it, the engine that would have to quote it cannot read it in the first place. That is the first of the two failures, and it is the one you can fix today.

So what decides the answer

In our runs the answer was decided at the summary layer — the sentence the engine writes about your page, not your page itself. When the summary matched the shape of the question word for word, the page got named. When it did not, the page was skipped even when it was sitting at position one with a better page behind it.

That observation is from the second round of runs, where we read the model’s stated reason for each citation. It is a small sample and we treat it as a hypothesis, not a result.

What this does not tell you

Rank still matters — it is just not sufficient. A page at position 12 with a perfect summary may still lose to a page at position 1 with a mediocre one. What the data rules out is the reverse inference: from “I rank first” to “I will be cited”.