Natural-language interfaces have nothing to scan, so discoverability is worse than in graphical ones

What it is

A graphical interface lays executable acts on the screen: menus, buttons, icons can be swept by the eye. A natural-language interface is often just an empty field. The user cannot build a list of “what verbs live here” by scanning, and has to recall or guess the first sentence. That gap is discoverability without scannable controls. It is not “the prompt was poorly written.” The interface has no action set available for recognition.

Help pages, tutorial videos, and a handful of starter lines on an empty state do not change the structure: once the session starts, there is still nothing to scan.

Why it happens

Recognition is cheaper than recall. Menus turn acts into recognition; command languages turn them into recall. Natural language is harsher than a classical CLI: a CLI at least has a finite verb list and --help. A dialogue system’s legal phrasings are an open set. The verb table lives in the model, not on the glass.

Scanning also supplies negative evidence: “I scanned the toolbar, there is no Print, so printing probably is not here.” An empty field supplies none. People neither know what is possible nor what is impossible, and they read failure as “I did not say it right,” not “the function is absent.”

Studying it

Use an action-enumeration task: implement the same functions as a toolbar UI and as a chat-only UI, and ask novices, on a clock, to list “everything you think you can do here.” Independent variables: control visibility (fully shown, grouped-and-collapsed, hidden into language), task familiarity. Dependent variables: coverage against the designer’s capability list, count of invented functions, time to first legal request.

Do not stop at eventual task success. A lucky first guess can still leave the person with no map of the capability space. In the lab, “please try as much as you can” inflates exploration; the distribution of first utterances in product logs is closer to real missing scan.

Where it stops holding

Once frequent acts are chips or slash commands, those acts re-enter the recognition channel and this claim stops holding for that slice; the long tail that never got a visible entry still holds. On audio-only devices there is no scan path at all; the problem becomes auditory traversability — same mechanism, different measure (spoken recall). For experts who have internalised the verb table, discoverability recedes and efficiency plus ambiguity take over.

Applying it

• Keep a scannable set of action entries next to the field (chips, a slash menu, direct manipulation on a canvas) covering the verbs the product actually wants used — not a one-frame flash on the welcome screen.
• Every visible entry must be clickable, not ornamental copy. A click either fills the field or runs the act, so recognition connects to action.
• If a class of things is firmly out of scope, put one negative sentence near the field (“cannot read your calendar”) so scanning has negative evidence.
• Check: with someone who has never used the product, hide every explanation, leave the main surface for 30 seconds, and have them speak the acts they believe exist. Map that against the capability list; high-frequency verbs they miss are the missing controls. Then check first-week logs: a capability that is on the list but almost never appears in first utterances does not exist on the interface.

Related

• Same group: L2.09.2 Examples draw the range of what is possible, so diversity matters more than count · L2.09.3 People treat seen examples as a ceiling; a narrow set shrinks real use · L2.09.4 Suggestions that appear only after failure arrive past the point of giving up · L2.09.5 Capability hints must move with the dialogue; a one-shot start-screen hint does not cover later turns
• Nearby: L2.02 Discoverability of What Can Be Said · L2.01 Openness of Natural-Language Commands and Its Cost · L2.04 Complementarity of Parameter Controls and Natural Language
• Search terms: discoverability without scannable controls · recognition versus recall · conversational UI affordance