User Generated Content Quality and Quantity Research Experiments

• My role: Senior Product Designer
• Deliverables: In-depth Research, UI/UX/IxD, User Testing
• Audience: Private & SME Sellers
• Platforms: Web, Mobile

Listing editor

Context

Avito is a leading classified platform, offering ads for goods, transportation, real estate, jobs, and services from both individuals and businesses. Since 2021, it has become the world's most visited classified, surpassing Craigslist in daily active users (DAU), with over 42M registered users and 150M listings.

However, with increasing popularity, companies like Yandex and Yula launched competing services, focusing on a more refined content approach. This posed a risk of user churn, making it critical to enhance the quality of Avito's content while maintaining the growing volume of listings.

To address this challenge, we initiated the "Content Quality" product stream. Over three years, we conducted a comprehensive research, ran a dozen experiments, defined high-quality content and developed strategies to encourage users to create better listings.

Challenge

Growing the number of listings is straightforward. You fix the posting flow funnel: refine critical steps to improve conversion rate, and pour more traffic at the top. A clear metric and approach to work.

However, what constitutes quality is not as clear. On an empirical level, we know that a listing should have many good photos, a detailed description, and filled-out parameters. The question is how to quantify this into a usable metric.

This challenge is compounded by the fact that posting flow varies across categories and platforms, containing different parameters and even features. Additionally, more and more designers are joining business verticals and contributing to the posting flow, further extending the information model and adding new services.

Issues with UGC

We organized a workshop to capture a wide range of viewpoints on the issue, bringing together designers, researchers, and product managers from all the business verticals.

Problems with UGC

A significant portion of our content (over 30%) is contributed by private sellers, often resulting in listings that lack visual appeal. This inconsistent quality reinforces a perception of the platform as a flea market.

Product vision

Content influences the perception of a platform, but the platform can also influence the type of content displayed on it. Our strategic goal was to move towards content similar to Airbnb's, in order to attract a new, younger audience of buyers, followed by new professional sellers, and to justify investments in developing new services for them.

Target vision for the main recommendations page

Product research findings

Competitive analysis

My team and I researched similar platforms and examined how they work with UGC. We didn't find any groundbreaking technologies. Everyone had some combination of content guides, static prompts in the interface, and simple AI models for improving photos and descriptions.

Competitive analysis

User interviews

We regularly spoke with users across all segments and compiled a list of issues that prevent them from creating quality content:

1. Beginners have no idea what content is required of them.
2. Experienced users do not want to spend a lot of time filling out listings.
3. They are frustrated by having to manually select the category for their listing.
4. Users dislike that the form does not remember recurring parameters such as phone number and addresses.
5. People are unsure what to write in the description after filling in numerous parameters.
6. They are not sure how to take photos and what is important to show.

The key idea is that users do not care about the process of filling out a listing. It has no intrinsic value and merely serves as an obstacle to their ultimate goal — selling their goods or services. Therefore, it should be as simple and unobtrusive as possible.

Analytical approach

To digitize the content quality metric, we came up with a hypothesis that listings with good content should have higher liquidity. To test this, we conducted a buyer experiment, comparing similar products with different content.

The results surprised us. It turned out that listings with poor content sold better; they had a 25% higher click-through rate (CTR). We explained this by the fact that users got used to buying used items from private sellers, often ignoring the professional content created by pro sellers.

Based on this experiment, we developed a holistic metric for quality content. It varied by category, but generally included the following criteria:

• Correctly filled category
• Properly filled strict parameters, especially category-specific ones like VIN
• At least 4 well-cropped photos formatted for search snippets with proper color correction
• Ideally, a product video
• At least 200 characters in the description
• Filled out contact methods

Solution hypotheses

Based on the identified problems, our team brainstormed ideas to solve them. We developed two distinct approaches to working with UGC, which we ultimately combined into one.

Approaches to Working with UGC

If we simplify the posting flow for private sellers, it will help them create quality content, bringing them closer to professionals and positively impacting the entire platform.

Solution 1. Category suggest

App category suggest

The easiest implementation was assistance with category selection. The user selects a top-level category. Then, depending on the chosen category, we suggest uploading photos (for goods) or filling in a title (for services and jobs). The AI model recognizes the sub-category with 90% accuracy for all popular listings. For unique items, users still have to select category manually.

We rolled out this solution across all platforms and achieved a +1.2% increase in successful posting conversion rate and a -4% decrease in abandoned drafts.

Solution 2. Item posting flow redesign

Posting flow sample and developer guide

Over the course of three years, I conducted several redesigns of the posting flow, each with varying degrees of success. The most effective outcomes occurred when we found synergy with developers in optimizing product support while simultaneously implementing a visual facelift.

Key accomplishments:

• Moved authorization out of the flow and eliminated unnecessary steps.
• Standardized step-by-step flow across mobile apps and desktop, addressing UX issues with navigation.
• Standardized and added about 10 specific posting flow components to the design system.
• Performed a visual uplift of all posting flow components.
• Created guidelines for adding standard parameters through the constructor for business vertical designers and product managers.

Overall, after dozens of iterations of testing, we achieved a +2.2% increase in conversion rates and a 6% reduction in posting flow restarts.

Solution 3. Category specific tips and guides

Category specific tips and guides

To help users understand the type of content we expect from them, we displayed category-specific hints. We know from UX research and interviews that users dislike reading large blocks of text in the interface. Therefore, we aimed to break it down into small interactive hints to capture their attention.

Over time, we gathered more data on user behavior and developed personalized hints to increase liquidity. We verified several display types in SBS tests and finally arrived at this set of in-product communications:

• Price recommendation based on market analytics in the category.
• Hints on how to take the right photos in the category, using placeholders and a full guide on a landing page.
• Sellers coach hints to increase listing liquidity.
• Built-in guides on how to sell properly.

After A/B testing and launching these hints, users in targeted categories began to upload an average of 2 additional photos.

Solution 4. Smart camera

App category suggest

To help users take better photos, I got inspired by new camera recognition technologies in Android smartphones and packaged these ideas into a concept:

• A large widget in the posting flow to draw attention to the camera.
• A guide with basic tips on how to prepare for shooting.
• Object recognition and subcategory refinement.
• Autofocus, centering assistance, and different formats for key categories.
• A step-by-step, category-specific guide to photographing the item from correct angles.
• Automatic color corrections based on lighting conditions.

Smart models in the gallery concept

90% of the users come with ready-made photos, so I thought about integrating smart models into the gallery:

• Animated counter for added photos to motivate users upload more.
• Automatic photos arrangement in the most appealing order for buyers.
• Automatic color correction and basic filters.
• Highlighting duplicate and very similar photos.
• Highlighting poorly cropped, overexposed, and low-resolution photos.
• Placeholder hints for missing product angles.

Camera redesign specification

I brainstormed the concept with our ML engineers, and it turned out that the company lacked the expertise to develop neural networks on devices. Together with an initiative team, we tried to integrate open models to improve photo quality during a hackathon, but we didn't achieve production-ready quality.

I redesigned the camera layout to prepare it for the future appearance of smart hints. We launched it with gray metrics. The next step was to increase the visibility of the camera in the product to boost content quality and quantity metrics.

Solution 5. AI-generated listing description

AI-generated listing description

A good product description is concise, informative, and address common customer queries without duplicating existing product information. To simplify the process of writing product descriptions, I developed a concept using LLMs:

1. Training LLMs. We analyzed high-quality descriptions and FAQs in each category to train LLM models, ensuring they understand what makes a description effective.
2. Guiding sellers. Sellers were prompted with frequently asked questions from buyers in their category. This helps ensure that descriptions address common customer queries.
3. Speech to text. Sellers responded by dictating their answers in free form, and we converted their speech to text in real-time.
4. AI-Generated description. Sellers' responses, combined with product and category knowledge, were used to generate descriptions via LLMs. This process transforms data into a readable format that tells a compelling story about the product.
5. Post-publication enhancement. After listing was published, sellers had an option to enhance the description by answering new questions from real buyers, ensuring the description remains relevant and addresses evolving customer needs.

We refined the concept with AI engineers and released an MVP with simple description generation, requiring no upfront input from sellers. This approach resulted in a 20% increase in the average length of descriptions, indicating more comprehensive and detailed product information.

Solution 6. Listing quality index

Listing editor

Over the course of three years, I conducted several experiments to integrate gamification into posting flow. I tested a ten-point scale and qualitative assessments with different wording, ultimately settling on a simple percentage as the most user-friendly option. This percentage increased based on the completion of strict parameters, the number of photos, and the length of the description.

To maintain consistency across interfaces, I also implemented this feature in the listing card. I redesigned it, organizing product data into blocks with quick editing, and integrating quality index along with tips for improvement.

This was my final project at the company. I validated the concept through usability tests and handed it over to the team for further implementation.

Final solution

After numerous iterations and experiments, we successfully developed a consistent content posting flow that integrates smart models and best practices to further improve content quality:

• Standardized posting flow with a userguide for managers and designers (released on mobile and desktop).
• Category suggest (released on mobile and desktop).
• Category specific guides and tips (partially released).
• Smart camera (concept) and photo flow redesign (released on mobile).
• AI-generated description (MVP released on mobile).
• Listing quality index (concept).

Outcome & impact

• Content quantity. The conversion rate to successful publication increased by 3.4%.
• Content quality. The number of listings with more than four photos increased by 4%, and the average description length grew by 20%.
• Impact. We developed best practices for managing content quality and shared them with vertical teams, enhancing capabilities in their domains.

Reflections & learnings

Long-term vision. It's beneficial to develop a comprehensive product vision that extends far beyond six months, even if current technical capabilities don't allow immediate implementation. This approach broadens the perspective and motivates the team by working towards something larger and more meaningful.

Resource management. Ideas often outnumber the resources available for their implementation. It's crucial to collaborate with other teams, leverage external resources, break projects into manageable pieces, and delegate specific solutions to achieve synergy.

Team credits

During my work on the posting flow, the team changed several times. While it's not possible to mention everyone, I would like to express my gratitude to those who have had the most significant impact:

• Senior Product Designer: Paul Sikorsky
• Art Director: Alexander Stogov
• Product Designer: Elena Gavrikova
• Product Designer: Valeria Smirnova
• Product Designer: Matvey Grab
• Lead Product Manager: Margo Alumyan
• Senior Product Manager: Adeliya Sharafeeva
• Lead UX Researcher: Evgeniya Kritchko
• Senior UX Writer: Ekaterina Eliseeva
• Senior Product Analyst: Alina Dranga
• iOS Engineer: Kirill Davidov
• FE Engineer: David Kagirov

I would love to do more AI-related projects

I thoroughly enjoy integrating AI models into products to make users' lives easier. If you're interested in collaborating on similar projects, feel free to reach out!

Telegram, LinkedIn, p@sikorsky.design.