The Ordinary

Chatbot Design for a Global Skincare Brand

Designing for conversation: what a system says, how it recovers, how it stays on-brand. A skill that only matters more as AI becomes the interface.

The Challenge

Skincare can feel overwhelming, especially for beginners.

Buyers who don't understand what they're looking at don't buy. On The Ordinary's site, that means decoding product names, cross-referencing ingredients, and figuring out a routine before adding anything to cart. We wanted to close that gap and turn curious visitors into confident ones.

The answer was a chatbot that actually knew the products. Three focused flows:

Does this sound like The Ordinary?

The Ordinary sits at the affordable end of the market with high scientific credibility. That combination has a specific voice, and the bot needed to match it on every message. I defined the personality before writing a single script: informative, knowledgeable, professional, inclusive.

Minimal personification. Helpful, knowledgeable, the Ordinary logo as its only face.

Research shows users often don't know what a bot can do, which leads to abandoned sessions.

The "gulf of envisioning," an HCI concept, describes the cognitive distance between a user's goal and their ability to formulate a prompt that gets them there. Three named entry points are the answer: they make the possible paths visible before users have to ask.

Not every input lands cleanly. Users go off-script, ask questions outside scope, or combine intents the bot wasn't trained to handle. Escape hatches are the design response to that: deliberate exit routes mapped into the flow.

Magnify

Users don't describe their skin the way dermatologists do.

The training data work surprised me. I expected it to be mechanical: write the intents, list the utterances, done. Instead, it forced a more basic question: how do people actually talk about their skin? Not how they'd phrase it on a form, but how they'd text a friend.

Sample scripting is how a conversational design gets stress-tested before it exists. Every script asked the same three questions: is the voice consistent, does the user have enough guidance to keep moving, and is there a real response for every input, even the ones you didn't anticipate.

Personality

Scripts had to reflect the bot's voice so every message matched the brand users already knew.

Turn-taking cues

The bot needed to guide the conversation and signal what comes next, not just respond.

Responding to needs

If no suitable answer existed, the script needed a redirect rather than a dead end.

What I learned building a rule-based chatbot applies directly to how AI systems are designed today.

When the system can't infer anything, you have to define everything. That discipline transfers directly to designing the guardrails, prompts, and fallbacks that make AI systems work.

The technology changed. The design problems didn't.

Rule-based logic forces every decision into the open. What I designed around here: multiple intents, vague input, no graceful fallback. All still unsolved in most AI chatbots.

What I took from this: the voice, the pacing, and the recovery paths determine whether a user trusts a conversational system. Not the model underneath. An LLM can handle more intents. It still needs someone to decide what the system should sound like, what it should stay out of, and what happens when it gets it wrong.

If I returned to this project, I'd layer in AI to expand what the bot can understand and pair guided entry points with free-text input. But the personality definition, the scoping decisions, and the fallback logic would stay.