Autocomplete That Helps: Designing Search Suggestions People Trust
Autocomplete is the search box making promises. Every suggestion tells the shopper this query will work here, and a suggestion that leads to zero results breaks trust faster than no suggestion at all. Good autocomplete blends query popularity, catalog reality, and the individual context into a short list that feels obvious in hindsight.
Where suggestions should come from
The strongest suggestion lists merge three sources. Popular queries from your own search logs carry real demand in your shoppers language. Catalog matched completions guarantee the suggestion resolves to products, because they are generated from terms the index actually contains. Personal context, recent views and past searches, tailors the ranking for the individual. A list built from only one source underperforms: log only lists repeat other people typos, catalog only lists sound like a warehouse, and personal only lists trap returning shoppers in their last session.
Filter ruthlessly by outcome. Any suggestion that resolves to zero or near zero results should be suppressed automatically, and that check should run against the live index, not a weekly export. Nothing in autocomplete damages credibility like the store suggesting a search and then apologizing for having nothing. Treat suggestion quality as a promise keeping metric, and audit it like one.
Ranking and the shape of the list
Rank by expected usefulness, which usually means a blend of query popularity, result count health, and conversion history, with an exact prefix match boosted so typing feels responsive to the shopper own words. Cap the list at a length a person can scan while typing, typically six to ten suggestions, and group when you mix types: terms first, then categories, then specific products with thumbnails where the design supports them.
Product suggestions inside autocomplete are powerful and risky. A thumbnail with a price answers some queries instantly, a specific item can also hijack a general intent. Reserve product rows for queries that look navigational, near complete product names or model numbers, and keep category and term suggestions dominant for exploratory typing. The shopper typing running is browsing, the shopper typing a full model code is hunting, and the list should behave differently for each.
Keyboard, touch, and speed details
Autocomplete is an interaction before it is an algorithm. Arrow keys and enter must work flawlessly, escape must dismiss, and the highlighted state must be visible at a glance. On touch devices, targets need real size, the list must not be hidden by the keyboard, and tapping a suggestion must fire reliably on the first touch. These details sound basic, and they are where most implementations quietly bleed usage to the plain search button.
Speed is a feature of the list itself. Suggestions that arrive after the shopper has typed three more characters train them to ignore the box. Debounce lightly, cache popular prefixes aggressively, and render partial updates without flicker. Measure suggestion latency as a user facing metric with a budget, the same way you budget page load, because a slow autocomplete is functionally a missing one.
Learning from autocomplete analytics
Log impressions as well as clicks. A suggestion shown a thousand times and chosen ten times is actively hurting, either by ranking too high or by promising a weak result. Track selection rate per suggestion, the zero results rate of autocomplete driven searches separately from typed searches, and the queries shoppers complete by hand after ignoring suggestions, since those hand finished queries reveal gaps in the suggestion sources.
Refresh suggestion sources on a cadence that matches your catalog churn. Fast changing catalogs need daily regeneration of catalog matched completions and outcome filters, slower catalogs can run weekly. Seasonal language should enter the popularity pool quickly at season start, because shoppers start typing holiday and weather terms weeks before they buy, and autocomplete that has not noticed the season feels deaf.
Keep reading
Related guides on this site: Search Synonyms and Tuning: Teaching Your Store Search Your Customers Words Faceted Navigation and Search Relevance: Filters That Help Instead of Hide Practical answers about personalization.
Frequently asked questions
How many suggestions should autocomplete show?
Six to ten is the workable range for most stores. Fewer starves choice, more overwhelms scanning while typing. Group by type when you mix terms, categories, and products.
Should autocomplete include products or just search terms?
Both, weighted by intent. Term and category suggestions serve exploratory typing, while product rows with thumbnails serve near complete, navigational queries. Product rows for every keystroke usually backfire.
Why do bad suggestions keep appearing?
Because suggestion lists are often generated once from raw query logs without outcome filtering. Suppress anything that resolves to weak or zero results, and recheck against the live index regularly.
Does autocomplete actually increase conversion?
Well tuned autocomplete raises search usage and search conversion because shoppers reach working queries faster. The effect disappears when suggestions are slow or frequently lead to dead ends.