The question tree

The editor, the answers that lead to other questions, the results that return products, and what the validator flags.

The tree is the deterministic half of the assistant: closed questions that lead to a result. It doesn't spend AI allowance per visit, it always answers the same way, and it's what you use for the paths that repeat ("a gift", "for a beginner", "for sensitive skin").

How it's built

The first question is the way in. From there, each answer leads to another question or to a result. Each answer has its own label and an optional microcopy, and can be marked to stay out of the statistics, for sensitive choices.

Three questions are usually enough. Five is nearly always too many: the drop-off per question, in the "Where they give up" table, says exactly where the tree got long.

The results

A result defines what's shown at the end: title, description and categories chosen by name, with autocomplete from the real vocabulary of the index, never identifiers typed by hand.

In the advanced block: the sorting (best sellers, new arrivals, price, relevance), products pinned at the top and a technical filter.

The essential point is that the products of a result are searched for at that moment, in stock, and not taken from a fixed list that ages. A tree written in January still shows today's catalogue.

The validator

The editor validates as you write, and flags:

• questions with no answers;
• answers that lead nowhere;
• empty results;
• orphan branches, with the "nobody gets here" badge.

A tree with warnings still saves. But what's flagged is what a customer will run into as a dead end, and it's worth clearing up before switching the status to Active.

Draft and active

An assistant in Draft is invisible in the shop. It's the state where you build calmly, experiment and get things wrong. Active puts it in front of customers, and from then on every change is immediate: there's no two-step publishing.

Reading the numbers in the editor

Each question and each result shows its statistics inside the editor itself. It's deliberate that they're there and not only in a report: the question that loses 60% of people is fixed in the place where the number appears, and not three screens away.

The usual reading:

• High drop-off on a question: either it's one question too many, or it's worded in the shop's language and not the customer's.
• A result seen a lot and clicked little: the categories it returns don't match what the answer promised.
• A branch with zero passes: nobody chooses that answer. Either the option doesn't exist in the buyer's head, or it's written in a way nobody recognises themselves in.