Chrome is preparing an API that turns what your visitors type into filters: 38% to 85% right depending on the engine

Get ready by writing your questions and answers now

In practice, the site’s developer adds a small script to the page. The script describes the questions in a “schema”, a JSON-like object: for each question, its wording and its possible answers. When the visitor runs a search, the script passes the typed text to Chrome, which returns the chosen answer and a confidence score for each question; the script then ticks the filters. Example for a car site:

// The schema: the questions Chrome will have to settle
const schema = {
  context: "",  // left empty
  questions: [{
    id: "body_type",
    type: "choice",
    prompt: "Which body type is the visitor looking for?",
    options: [
      { label: "city car", description: "small urban car, easy to park" },
      { label: "sedan",    description: "family car with a large trunk" },
      { label: "suv",      description: "tall, roomy vehicle, 4x4 style" }
    ]
  }]
};

// The text the visitor typed goes to Chrome, which picks an answer
const model = await DecisionModel.create(schema);
const answer = await model.decide(typedText);
// answer.body_type: chosen answer + confidence score from 0 to 1

In our tests, four adjustments to this schema take Chrome’s current model from 38% to 58% correct answers:

• Fill in the “description” field of each possible answer (+13 points).
• Leave the “context” field, where the site can introduce itself, empty (+10 points on multiple-choice questions).
• Prefer lists of described choices to yes/no questions, and avoid catch-all answers such as “other”.
• In the script, tick nothing automatically: with the current model, the confidence score is not reliable. Suggest the filters to the visitor instead.

This tuning has to be redone for each model: the same schema drops to 45% with EmbeddingGemma 2.