Mixing history with suggestions blurs where a query came from

What it is

One row in the dropdown is what you typed last week; the next is a site-wide popular query or a model completion. If they share one visual style, people cannot tell “this is my record” from “this is the system selling a query.” Source confusion happens when history and suggestions are interleaved without a learnable mark. The costs of a wrong click are asymmetric: treating a trending query as history issues a search never intended; treating history as a suggestion makes a sensitive string look like a public example.

Suggestions cut the cost of constructing a new query; history cuts the cost of repeating an old one. Folded into one undifferentiated list, the two mechanisms impersonate each other.

Why it happens

Autocomplete is a recognition task. The eye scans candidates; working memory keeps “which line looks like what I want.” Source is second-order information: it needs a label, a group, or an icon to enter working memory, and by default it is not encoded. Choice is then driven by surface similarity—the shorter, bolder, closer-to-the-prefix line wins, whether it came from a personal log or a global language model.

Confusion also contaminates attribution. People treat the issued query as a record of their own intent. If the click was actually a trending item, the later “what I searched” file is polluted, re-finding fails, and the privacy list contains sentences they never wrote. The reverse holds too: a personal medical query that looks like a suggestion will be read by bystanders as something the site is promoting, and by the searcher as a public term the system has already adopted.

Studying it

Measure source judgments, not whether completion speeds typing.

• Paradigms: under one prefix, offer both history and suggestion rows; manipulate grouping and “Recent / Popular” labels; ask before each click “yours or the system’s,” or load a dual task and watch source judgments collapse. Log analysis of mixed autocomplete can report click share of personal versus public rows, but a click does not tell you whether the person knew the source.
• Independent variables: visual separation (grouped vs mixed), presence of source labels, textual similarity between a history row and a suggestion row.
• Dependent variables: accuracy of source judgments, trending queries issued as if they were history, later recognition of strings in “my history” that the person never wrote.
• Methodological note: if history and suggestion wording is far apart, mixed lists rarely misfire and harm is underestimated. Use stimuli where both lines are plausible after the same prefix. Measure over-the-shoulder conditions separately; that is embarrassment, not efficiency.

Where it stops holding

A product with history and no suggestions cannot mix them. Neither can a product with suggestions and no stored history. On a single-user machine with a very short history, people may recognize their own sentences by content, and labels add less. A controlled-vocabulary dropdown (pick a field value from a whitelist) is neither suggestion nor history; hanging source logic on it mislabels the control. Screen-reader users hear a linear list; group headings matter more than color or icons.

Applying it

• Split history and suggestions into two blocks with headings (“Recent searches,” “Suggestions”). Do not ship one unlabeled list.
• Mark history with a clock or “you searched this,” and suggestions with a different mark. Color must not be the only distinction.
• Sensitive history should not appear in a suggestion slot that can be overseen; at minimum, source must be readable at a glance.
• Verify by placing one of the person’s own queries and a similarly worded popular suggestion under the same prefix, and asking someone who has not been briefed “which line did you write.” If they cannot tell, or they click the popular row and call it “what I searched last time,” mixing has already confused source.

Related

• Within the group: G3.11.1 History lowers the cost of typing the same query again · G3.11.2 Search history is privacy-sensitive and must be deletable
• Adjacent: G3.03 Search suggestions · G3.02 Query formulation · G3.04 Spelling correction
• Search terms: source confusion · autocomplete · search history