AI discoverability: helping users learn what they can ask

AI discoverability: when users don't know they can ask for it. What NN/g found about AI features nobody notices and which patterns make them visible.

info Quick Definition
Discoverability is the ability of a user to notice, recognize and understand a feature they didn’t know existed. In AI, it means the person learns what the system can do without having to guess or read a manual.

How does discoverability differ from findability?

They are two different things that often get confused. Nielsen Norman Group defines them like this: a feature has good findability if people can locate it when they look for it, and good discoverability if people notice, recognize and understand it even when they were not aware it existed. Searching for the export button because you know it exports is findability. Learning that a tool can summarize meetings is discoverability: nobody searches for what they don’t know exists.

In traditional interfaces, discoverability comes from visible affordances: menus, toolbars and labelled buttons. A chat box offers none of those cues. It presents an empty text surface, and the real capability sits inside the model.

What did Nielsen Norman Group find about AI features?

In March 2025, Kate Moran and Feifei Liu published an analysis based on qualitative usability testing, with Amazon as the main case. The article does not say how many participants took part, so read it as a qualitative signal rather than a statistic. Its central conclusion: most of the AI features tested lacked both discoverability and findability. Participants did not expect AI features, did not notice them while using the product and struggled to find them when asked. They often found the features useful once a researcher pointed them out.

Three findings from the Amazon case work as concrete examples:

  • Rufus, the chat assistant. No participant noticed it until the researcher showed it. On desktop, its button sat in an unusual spot (upper left), had an unfamiliar name and an ambiguous sparkles icon.
  • AI prompt suggestions in search. No participant tapped them unprompted: the search bar looked unchanged, the suggestions blended in with keyword suggestions, and people assumed the bar was only for finding products.
  • AI Q&A on the product page. It sat at the bottom of dense pages, and most participants scrolled past it.

NN/g’s recommendation runs against creativity: use standard patterns (for example, the chat button in the lower-right corner), plain language and familiar icons, and avoid “unnecessary labels, cute names, or inscrutable icons”.

Why doesn’t a chat box communicate what it can do?

Because the first contact fixes an idea. Leon Furze, in a 2026 essay aimed at education, describes the typical chatbot interface as a text box with a blinking cursor that gives no signal about what the system can do, and argues that this first interaction becomes a ceiling: people do not explore beyond what they imagined at the start. It is a reasoned opinion, not research data. Furze also warns that the problem will not be solved by better interface design alone; it also depends on how people think about these systems.

For a designer the practical consequence is clear: a feature nobody discovers does not exist for the user, however well it works. And if the first experience is narrow, people don’t come back to try new capabilities.

Which patterns help?

Five patterns drawn from NN/g’s guidance and from practice:

  1. Examples on the empty screen. A short set of concrete tasks, such as “Summarize this meeting” or “Compare these two suppliers”. NN/g calls them use-case suggestions and describes them as useful mainly for newcomers. More in prompt suggestions and the chat empty state.
  2. Follow-up suggestions. Below each answer, questions based on the conversation. NN/g notes these tend to be the most relevant and useful because they are tailored to a specific need.
  3. Prompt autocomplete. A suggestion below the input that continues what the person is typing. Useful as long as it does not interrupt. See AI autocomplete.
  4. Familiar names and icons. “Ask AI” or “Summarize” say what they do. An internal codename or an unlabelled sparkles icon tells nobody anything.
  5. AI where the work happens. A button inside the current flow is noticed more often than a separate page of AI features. See inline AI and AI feature onboarding.

One detail: these aids should be easy to see without being intrusive. A pulsing bubble that demands attention every time someone opens the app becomes noise.

When do suggestions NOT help?

  • When they appear in the wrong place. A suggestion that covers content or moves while the person types gets ignored or resented. NN/g observes that people often ignore suggestions that are not in the right place or don’t feel useful.
  • When they are not relevant. Generic suggestions unrelated to the person’s task are ignored.
  • When the user is already an expert. Basic examples become tiresome for someone who masters the tool. Offer a way to collapse or turn them off.
  • When they look like something else. In the Amazon case, AI suggestions blended in with keyword suggestions and read as more of the same. If they look identical to what’s familiar, they are not discovered.
  • When the real problem is something else. If the feature doesn’t solve anything, showing it better won’t make it useful. Before redesigning how it is found, check that the feature earns its place.

Example

Illustrative example: a customer management app adds an assistant for drafting follow-up emails. At first the assistant sits in a three-dot menu and almost nobody uses it. The team moves it onto each customer’s record, with a “Draft follow-up” button next to the call history. The assistant’s empty screen shows three examples: “Remind them about Tuesday’s proposal”, “Ask for the renewal date”, “Summarize the last meeting”. After the first answer, a follow-up appears: “Send it by email?”. The feature did not change; where it appears and what it shows about itself did. To check it worked, the team compares how many people use it for the first time and how many come back the following week, not just clicks on the button.

Common mistakes

  • Hiding the feature behind an unlabelled icon.
  • Using internal or marketing names instead of the action the feature performs.
  • Showing generic suggestions unrelated to the task.
  • Measuring discoverability only by clicks on a banner, without checking whether the feature was used afterwards. A better measure combines activation rate, time to first use and repeat use.
  • Assuming the chat box communicates the capability. Compare chat vs. structured UI before choosing the surface.
  • Testing only with people who already know the feature. Discoverability shows up in usability tests with people who don’t know it.

Sources

Frequently asked questions

What is discoverability of an AI feature?
It is the ability of a user to find features they didn’t know existed. It differs from findability, which is locating something you already know about.
Why don't users discover AI features?
In Nielsen Norman Group’s qualitative usability tests (2025), participants did not expect AI features, did not notice them during use and struggled to find them when asked. Many also live inside a text box with no buttons or menus to announce them.
What helps show what AI can do?
Concrete examples on the empty screen, follow-up suggestions below each answer, familiar names and icons instead of clever ones, and standard placement inside the flow where the person already works.
Do prompt suggestions always help?
No. According to NN/g, people often ignore them when they are not in the right place or don’t feel useful. They should be relevant to the task and to the user’s experience level.