Help-seeking indicates that the interface does not explain itself
Aliases: help-seeking rate · self-explanation · discoverability failure
What it is
Help-seeking is what people do when they will not, or can no longer, infer the next step from the interface alone: asking a moderator, opening help, searching an outside tutorial, calling someone. It measures a failure of self-explanation, not an error already executed. Error counts catch deviations after action; help-seeking catches the pause before action. High help with low errors often means people used an external explanation to avoid clicking wrong—not that the interface was already clear.
Why it happens
A learnable interface should let goals, controls, and feedback imply a next step. When that mapping breaks, people pause and check outside, because the cost of guessing wrong exceeds the cost of asking. In the lab the cheapest outside source is the moderator; in products it is a help center, a colleague, or a search engine. If the protocol allows questions at any time, help-seeking inflates; if questions are banned and in-product help is absent, difficulty reroutes into silence, trial-and-error, or abandonment—the rate falls while the problem remains. Help-seeking is therefore “how often the currently available outside source was invoked,” and must be read with whether that source was open.
Studying it
Code help separately from errors: at which step the first request occurred, what was asked, and whether success followed immediately. Split in-product help, moderator prompts, and off-site search. Keep unassisted completion as the primary contrast and treat help-seeking as a self-explanation gap; do not add it to the error rate. Hold help policy constant across design comparisons. In the field, help-page views, search queries, and self-service attempts before a ticket can proxy help-seeking, with the caveat that only traced requests are visible.
Where it stops holding
A moderate amount of help on a novice’s first pass through a complex domain can be expected learning, not a defect. Checking with assistive technology or a human assistant in accessibility settings should not be scored as interface failure. Some people will not ask because of temperament or power; zero help-seeking is not zero confusion. Making help highly visible will raise the rate by design, which is not automatically a regression—then the question is whether people can finish independently after help, not whether the rate alone is “bad.”
Applying it
- State in the script when the moderator may be asked; wait for an explicit request by default, and mark the pause location before it.
- Report help-seeking beside unassisted success; do not treat high-help success as sufficient self-explanation.
- Read help-search queries and empty results on the product side, and fold recurring questions back into wording and defaults rather than writing a longer article.
- If help-seeking falls after a change while abandonment rises, the outside source was removed and self-explanation was not added; treat the issue as open.