A bare term is a gesture toward a category. The answer it receives is the average of everyone who ever gestured that way.

A query that names a category and leaves everything else open. Holiday, insurance, clinic. Nothing has been asked in the ordinary sense. A direction has been indicated, and the system answers it with the average of everyone who has ever indicated that direction.

The linguistic name for what a bare term does. An expression omits values that a listener then supplies from general principles, from habit or from the situation. It is the same family as deixis, which leaves place and time to the situation, and presupposition, which leaves an assumption unstated.

“underspecification is an analytic strategy in which a linguistic representation omits the value of one or more features”

Reference: Wikipedia, Underspecification. https://en.wikipedia.org/wiki/Underspecification

Each word takes one decision away from the situation and puts it into the query. Adding a place removes the question of where. Adding a month removes the question of when. Adding a purpose removes the question of what for. At the top of the ladder stands an utterance that works without its situation, which is exactly why it can be answered, and exactly why almost nobody has asked it.

One ladder, five rungs, German market, measured on 17 September 2026. From holiday to holiday italy study trip venice october 2026. The result sets changed sharply: the domains shared with the bare term fell from all of them to 31 per cent at the first rung, and to around 6 per cent at the top. The engine does read the added words.

The last one. Adding the month left six of eighteen domains in place, and those six were already there for the rung below. The most precise element of the question, the one that decides whether a trip is even possible, moved the page least.

The place, and little else. On the top rung the results were checked for the words of the question. Venice appeared in fourteen of nineteen addresses, Italy in eight, October in two, the year in one. The word that named the purpose of the trip appeared in none. Alongside the travel pages the page carried a broadcaster's video about a dolphin, a news report about a boating accident and a magazine piece about a town said to be nicer than Venice.

Because a place is an entity and a purpose is a qualifier. Retrieval systems hold entities, link them and can match them across documents. A word that states what something is for has no such anchor, so it is the first thing dropped when the reading is widened. The consequence is the one you can see on the ladder: a question about a form of travel comes back as a subject area.

The number of results. Every rung returned a full page, seventeen to nineteen organic entries, including the rung that almost nobody has ever typed. That is the finding that matters: a result page has a fixed size, so it cannot report scarcity. It cannot say that four people have written on this and that three of them are shops.

Because it is built to be filled. Ten slots exist, and ten slots are served, so the system broadens the reading of the query until enough documents qualify. The broadening is invisible to the reader, and its result looks exactly like abundance. A reader who could see how thin the ground is would ask a different question or accept a different answer.

It is the process of reformulating a query to improve retrieval, by adding synonyms, related terms and rewritten forms. It was introduced to solve a real problem: a person writes heart attack where a document says myocardial infarction. It also carries the specific back toward the general, which is what fills the page when the specific alone would leave it empty.

“the process of reformulating a given query to improve retrieval performance”

Reference: Wikipedia, Query expansion. https://en.wikipedia.org/wiki/Query_expansion

Asked and answered, which is the consensus visible on the result page. Asked and thinly answered, which is what the people-also-ask boxes collect. And never asked, which appears in neither, because both are built out of questions that were put. The third layer has to be brought in from outside.

Arithmetic. A system assembled from questions that people have entered can contain no question that nobody has entered. This is no shortcoming of the feature. It is its construction, and it is the reason the third layer is worth paying for.

Near the top, and usually silent. The person with a decision to make carries the place, the date, the budget and the condition in their head whether or not they type them. A page written for the bare term meets them with the average of the category. A page that names the rung meets them with their own case.

The rare words are where the specificity lives. In a clinical study of five hundred queries, retrieval by meaning alone failed for 22.2 per cent of them, and keyword retrieval surfaced hundreds of documents that the semantic path never reached. The exact wording of a rare question is the part that generalisation removes first.

“dense-only retrieval fails for 22.2% of clinical queries even at the most permissive threshold”

Study: BM25 and Dense Retrieval Are Complementary for Portuguese Clinical Text, preprint, Research Square, 2026. https://www.researchsquare.com/article/rs-10757954/v1

Write for a rung, and say which one. Name the place, the period and the case inside the sentences, so the page answers a specified question rather than a category. This is the same instruction that follows from Deixis, arrived at from the other side.