YandexGPT Chatbot

The Yandex Realty Marketplace team, in partnership with the YandexGPT team, has created a product that lets users submit a text description of the apartment they’re looking for. The system then presents different options to the user and explains why each one might be a good fit.

The search function is only available for newly constructed buildings. Existing AI cannot engage in dialogue with users.

Yandex has many products that are starting to integrate GPT technologies into their services. Before designing our chat, I wanted to see how colleagues from related products were doing: Yandex Market (similar to Amazon), Lavka (Uber Eats), Yandex. Auto, Yandex Search (Google), and Alice voice assistant (Siri).

I looked at various ways to interact with users in different contexts and the entry points for chatbots. I also paid attention to the layout of the browser and mobile versions. We decided to use a classic chat format with bubbles and cards, as it is the most familiar experience for users to communicate with support chat.

During the presentation of the project, the product manager showed the user flow diagram. Based on this, I began to think about various possible scenarios and states. This type of idea demonstration saves a lot of time. I also talked with the developers to consider all possible pitfalls in the early stages of development.

To begin, I created wireframes to get a better understanding of the overall picture. The bird’s-eye view helped me see all possible scenarios. In addition, I made the layouts interactive so that I could think about the animations for each element. This approach allowed me to immediately identify potential errors in the scenarios that I might have missed during the user flow design stage.

To save time in the future, I scheduled several meetings with GPT designers. We discussed the possibilities of using their logo and brand name, as well as asked them to evaluate the first detailed sketches.

I analyzed individual screens and interface elements through usability testing. Yandex has its own platform for this purpose, called 'Asks'. There are two common scenarios:

First Click Test — users can click on the desired element, and the frequency of clicks will be presented in a heat map. In another case, they can answer questions about how they understand a particular element, and you will receive a list of responses that can be ranked.

This study involves 100 respondents and takes up to 5 minutes to complete. It is a valuable tool for making decisions.

The question for the test was: 'You have contacted for help in choosing Residential complexes. We have sent you three options, but you would like to see more. Where can you click to view more new complexes?'

Based on the results of the test and my personal observations, I have chosen the first option. This choice is based on the following reasons:

1. The first option has a logo, which is a trust element for the user.
2. The chevron arrow in the first option is visually lighter and more attractive than the button in the other options.
3. The quantitative indicator (73%) is the average of the tested options. Considering the previous advantages, this is the most optimal choice.

The next stage was the creation of the final, detailed layouts. I also worked with the editor on the texts. We tried to write concise texts that were short, but still provided context for the AI assistant. We considered ToV when communicating with our users.

The video demonstrates the main scenario for interacting with a chatbot. You can give it (only Russian is supported).

Some intermediate interface solutions are presented below: snippets-bubbles and entry points to the YaGPT chat.

During the two months of our chatbots working, we’ve received over 9k unique requests from around 32k users.

About 60% of the questions concerned new buildings, while the remaining 30% covered other topics. Users were probably curious and imaginative about the new product, which means that most of them understood the context and purpose of the chatbot.

We had planned to introduce the assistant feature in the app as well, but after six months of analysis, we realized that the conversion from chat to calls with developers was low. Perhaps this feature was a bit ahead of its time, as people prefer traditional search methods like filters. As a result, we decided to pause the development of this service for now.

This was the first major project for me after joining Yandex. I worked on the YaGPT chat functionality in parallel with my onboarding at the company. During this project, I learned a lot of new things.

What could I do to improve this assignment?

It was important for me to not only implement other people’s ideas but also to contribute my own ideas, bring innovative solutions, improve the product, and approach business with passion.