Auto-correction must show the original query and allow reverting it

What it is

When the system turns “recieve” into “receive” before ranking, that is spelling correction acting as auto-correction. Once the rewrite happens, the interface owes two things at once: name the string that was originally submitted, and offer one-click search on that original. Hearst’s canonical pattern is “showing results for X, still search Y”—X the corrected term, Y the original. Missing either, people cannot tell whether the list answers their query or a rewritten one.

Auto-correction’s object is the query the user can see, not fuzzy matching inside the index. Internal tolerance can leave the box untouched; once the visible query is rewritten, disclosure and revert are part of that rewrite, not courtesy on the side.

Why it happens

Auto-correction swaps one query for another; every row on the result page answers the second. People check whether the system understood them by asking “does my typed word appear in the constraints.” If only the corrected word remains, the check fails: the list can look sensible and still be a different object. Blame falls on content (“what I want isn’t here”) rather than on the rewrite (“you searched something I did not type”).

Revert has to be cheap. Correction is usually right, so people settle into “Enter, then results.” A rare wrong correction that requires clearing the page, editing the box, and submitting again breaks the rhythm; they accept the wrong list or quit. Placing the original as a parallel action on the result page (“search instead for [original]”) puts the check and the undo in the same glance. Silently rewriting the box is worse: working memory and the screen disagree, and the next edit starts from the wrong current value.

Studying it

Measure “did they notice the rewrite” separately from “was the rewrite correct.”

• Paradigms: spelling-correction UX experiments—correct rewrite vs wrong rewrite vs no rewrite—recording whether the banner is noticed and whether revert is possible; in query logs, the rate of “showing results for” followed immediately by “search instead.” Hearst treats Did-you-mean as a standard search-UI part.
• Independent variables: whether correction rewrites the input box, whether the banner shows both original and new query, whether revert is one click or retyping.
• Dependent variables: detection rate of wrong corrections, time to revert, rate of attributing failure to missing content.
• Methodological note: include both successful and failed corrections, or participants learn “the banner is always right” and stop reading it. If the lab prints the misspelling on the task sheet, people expect to be corrected and detection is overestimated. Revert clicks in logs only cover people who saw the banner; silent acceptance of a bad rewrite never appears in that count.

Where it stops holding

When the “wrong” spelling is the intended one—brands, codes, deliberate nonstandard forms—the correction is linguistically right and intentionally wrong; disclosure and revert are the only safety valve. Deciding that a token is a proper noun is a neighboring problem. Command-style search (type-to-execute) has no result page for a banner; correction belongs in a pre-execution confirm, not after the fact. When there are zero results and auto-correction has not run, the issue is offering nearby queries, not “already-corrected must disclose.” Even at very high confidence, when the original would hit nothing, still disclose; do not drop the banner because “the original would have been empty anyway.”

Applying it

• After auto-correction, one sentence above the list must name both the new query and the original, with “still search [original].”
• Do not silently rewrite the characters in the box; keep the submitted original, or if you change it, make stepping back obvious.
• After a revert, do not immediately auto-correct back to the same new term, or the revert control is fake.
• Verify by submitting a word that will be corrected wrongly (a hittable proper name rewritten to a common word). If the list appears as ordinary results with no original in sight, disclosure failed. If the original is visible but restoring it requires retyping in the box, revert failed.

Related

• Within the group: G3.04.2 Proper nouns carry the highest spelling-correction risk · G3.04.3 Zero results should offer nearby queries rather than stop
• Adjacent: G3.03 Search suggestions · G3.14 Fuzzy matching and approximate search · G1.08 Terminology consistency
• Search terms: spelling correction · did you mean · query rewrite