Skip to content
Article Nielsen Norman Group Jul 2026

NN/G: the 5 qualities of site-specific AI chatbots

A site-specific AI chatbot has a narrower job than a general assistant: it exists to help users with a defined set of tasks on a particular website, and users evaluate it against a human support agent they could have reached instead. Nielsen Norman Group published a framework in July 2026 that identifies five qualities separating chatbots users find trustworthy from ones they abandon.

The framework is grounded in evaluations of real chatbots — from retailers such as Home Depot, Williams Sonoma, and Arc’teryx — combined with participant feedback from user testing. The authors are explicit that each quality requires calibration through actual user testing; the framework names what to optimize, not the exact settings.

The five qualities

Handoff willingness. Users want a clear path to a human when the bot cannot help or when they feel frustrated. Offering escalation too early raises support costs, but refusing to escalate is worse. The article recommends surfacing a handoff option immediately when a user asks, and proactively offering one when someone rephrases the same question without getting a useful response.

Flexibility. Users make mistakes, change direction, and ask adjacent questions. A chatbot should handle these gracefully within defined guardrails — signaling uncertainty when it misunderstands, offering multiple-choice clarifications when it cannot repair a conversation, and handing off if repeated repairs fail. The right scope is “appropriately scoped,” not maximally narrow.

Proactivity. Two kinds matter. Clarification proactivity means asking follow-up questions only when the answers will change the response. Directional proactivity means offering next steps as scannable options — buttons linked to products or information — that are tied to the user’s current goal, not to upselling before that goal is met.

Emotional responsiveness. Sensitive domains, such as insurance claims or billing disputes, require a higher emotional baseline. The article recommends acknowledging a user’s situation without claiming feelings or assigning emotions to them. Acknowledgment should never substitute for actually resolving the issue.

Transparency. Four areas: identity (persistent indicator that the user is talking to AI), capability (surfacing features contextually and being specific about what the bot cannot do), rationale (explaining judgment calls such as product recommendations or declined requests), and privacy (explaining why sensitive information is needed, in the conversation, not in a separate policy).

Who this is relevant for

The framework is written for UX designers, content or conversation designers, and product managers responsible for customer-facing AI on websites and apps. The examples cover retail and service contexts, but the five qualities apply broadly. Teams shipping AI chatbots in regulated industries, where emotional responsiveness and transparency requirements are stricter, will find the framework particularly applicable.

The companion article, 10 Guidelines for Designing Your Site’s AI Chatbots (April 2026, same authors), offers more prescriptive recommendations for teams ready to move from framework to implementation.