Suggestions cut typing cost and demonstrate what can be searched
Aliases: query suggestion · type-ahead · search suggestions
What it is
The dropdown under a search box does two different jobs as query suggestion / autocomplete. One is completion: the prefix “travel” is already typed, and the rest of the phrase is filled so fewer keys are hit. The other is demonstration: a handful of complete, actually searched or actually existing phrasings tell the user what lives here and what those things are called. Hearst treats autocomplete as both fewer keystrokes and an instant sample of query language. The two jobs share a list but not a success criterion—keystroke saving is “picking a row was faster than finishing the string”; demonstration is “the user learned a word they could not have generated.”
Suggestions are not results. Choosing one should still land on a result page, not open a single document; otherwise demonstration is silently turned into a navigation shortcut and the list’s meaning is swapped.
Why it happens
Query formulation is expensive because words must be generated from memory. Suggestions turn generation into recognition: seeing “travel reimbursement policy” is cheaper than recalling “policy” from a blank. Recognition also sidesteps the vocabulary problem—someone who would only type “reimburse” sees “travel reimbursement,” so index language is shown before submit, and a short query is lengthened to a distinctive one.
Demonstration only holds if the list is an honest sample of the collection. All-popular terms never demonstrate rare-but-real types, and people infer those types cannot be searched. All-history terms demonstrate “what I searched before,” not “what exists here.” Prefix matching solves keystrokes; covering diverse types solves scent—each row is evidence that searching in that direction will yield something.
Studying it
Split “how many keys were saved” from “did they learn a searchable phrasing” as dependents; do not let click-through stand for suggestion success.
- Paradigms: keystroke savings (pick vs finish typing) versus vocabulary uptake (can the person later produce the full phrasing unaided); query logs of length change after a suggestion click; Hearst on autocomplete as query support.
- Independent variables: suggestion source (prefix completion / popular queries / content titles), whether the list covers multiple content types, whether an empty prefix still shows a demonstration.
- Dependent variables: keystrokes to a finished query, suggestion click rate, whether a click hits a target the user could not name, whether the user can say what else is searchable here.
- Methodological note: high click rate may only mean popular rows are clicked often; it does not prove coverage was demonstrated. If the lab sheet already prints the full name, suggestions are tested only as keystroke savers and demonstration is switched off. “Guesses” shown on an empty prefix are recommendations, not completion; report them separately.
Where it stops holding
For known items whose clue is a literal that must be exact (IDs, error codes), popular demonstrations steal attention; complete on the literal prefix and do not inject “trending.” Privacy-sensitive queries (names, records, unreleased projects) used as popular demonstrations leak; filter the source. In a tiny collection whose titles are already all visible, demonstration value is near zero and only completion remains. IME candidate bars stacked on search suggestions will hide the demonstration, especially on mobile.
Applying it
- Serve both successes in one list: prefix completions first, then complete phrasings of different content types—avoid ten variants of the same hot term.
- Demonstrate with real titles or real, legitimate frequent queries, not internal field names; each row should let someone predict what class of object a search would return.
- A click fills or submits a query; it does not jump into a detail record, so the list stays “this is a query, not a result.”
- Verify with someone who cannot produce the canonical title: can they pick from suggestions a phrasing that later proves to exist. Saving keystrokes without teaching a new word means demonstration never happened. A list of only the hot terms they already know is accelerating known-item search, not showing the collection.