Source: Production search UX methodology; Volume 2 query understanding outputs; major web search UX (Google, Bing, DuckDuckGo)
Classification — Patterns for presenting spell correction, query expansion, and reformulation suggestions to users with appropriate confidence-based behavior.
Surface query understanding outputs as user-facing affordances that improve search outcomes without removing user agency over their query intent.
Volume 2's spell correction can produce wrong corrections (the user actually meant the unusual spelling); auto-correcting confidently incorrect spellings produces frustrating failures. Showing too many alternatives clutters the UX; showing too few misses opportunities to recover from common errors. The patterns documented here address the confidence-based handling that production systems converge on.
Confidence-based behavior. The spell correction backend produces a corrected query plus a confidence score. The UX response depends on the confidence: very high confidence (the user almost certainly meant the corrected spelling): auto-apply the correction, show "Showing results for X. Search instead for Y" — the user sees the correction was applied with a one-click escape. Medium-high confidence: show original-query results but with prominent "Did you mean X?" link — the user can keep current results or switch. Lower confidence: show suggestions in less prominent places (related searches, sidebar) without claiming they're corrections.
Preserving user intent. Even when auto-applying corrections, the user's original query should be preserved as an affordance. Pattern: "Showing results for 'running shoes'. Search instead for 'running shoses'." The 'search instead for' link is essential — if the user really did mean the original (an unusual product name, a person's name, a technical term), they need a one-click path to get those results. Without the escape hatch, the team is paternalistically overriding user intent.
Confidence thresholds. The thresholds are workload-specific. Production patterns: auto-apply at very high confidence (typically 0.9+ in calibrated terms, when zero-result avoidance is the primary concern); show "did you mean" at medium-high confidence (0.7–0.9); show in less prominent places at lower confidence (0.5–0.7); don't suggest below 0.5. The thresholds should be tuned based on observed outcomes — a high auto-apply rate that produces user complaints about "it changed my query" suggests the threshold is too low.
Suggestion presentation. The "Did you mean" affordance is typically placed prominently at the top of results: a single line with the suggested correction as a link, plus the original query rendered with the suspect terms underlined or italicized. Visual distinction: italic for the original ("you typed"), regular weight for the suggestion ("we think you meant"). The link's click target should be the suggestion text plus enough padding to be tap-friendly on mobile.
Multiple suggestions. Sometimes multiple corrections are plausible. The UX choice: show one (the most confident) and risk being wrong, vs show multiple ("Did you mean X, Y, or Z?") and risk overwhelming. Production patterns favor showing one in most cases — reduces cognitive load and works for the vast majority of corrections. If the team chooses multiple, present them as a small ordered list rather than a long string.
Related searches. Beyond spell correction, query expansion produces related queries (synonyms, broader/narrower terms, popular reformulations). These should be surfaced as "Related searches" or "You might also try" — typically in a sidebar or at the bottom of results, not at the top (where they'd compete with did-you-mean). The presentation is browseable rather than prescriptive: "here are some other paths you might try."
Inline reformulation in empty states. When results are zero, the suggestion becomes the primary affordance. Empty-state pattern: "No results for X. Did you mean Y?" with the suggestion as a prominent link. The empty state is the place where reformulation suggestions can be most assertive because there's nothing competing for attention. Even low-confidence suggestions can appear in empty states as exploratory options.
Logging and learning. Did-you-mean clicks are valuable signals. Tracking which suggestions users accept vs reject improves the underlying correction model and informs the confidence thresholds. Production patterns: log every suggestion shown and whether it was clicked; aggregate to compute acceptance rates per correction pair; feed back into the spell correction model's training or tuning; revisit confidence thresholds periodically based on the data.
Any search system where users sometimes misspell or use vocabulary mismatched with the index. The patterns are foundational; the implementation can be simple (a few rules) or sophisticated (calibrated probabilities, contextual thresholds), but some form of did-you-mean handling is appropriate for almost all search.
Alternatives — silent auto-correction for very controlled workloads where the corrections are known reliable. Strict no-correction for specialized workloads (legal search, scientific search) where the user's exact terms matter. The middle ground (confidence-based suggestions) is the production default.
- Volume 2 of this series for the spell correction backend
- Production methodology writings on search UX
- Major search engine UX as observed (Google, Bing, DuckDuckGo)