Redefining Home Improvement Experiences with Generative AI

Improving Platform Interaction Experiences with GenAI to Boost Conversions and Strengthen the Bottom Line

Houzz is an online platform that has transformed the home remodeling and design industry. Houzz connects homeowners with a vast network of home improvement professionals, while offering millions of high-resolution photos for inspiration, a marketplace for purchasing home products, and powerful software for 3D renderings, estimates, proposals, invoicing, and project management.

Houzz’s business model is straightforward:

Provide the best possible experience for finding design inspiration, researching and hiring professionals, and discovering products to complete projects. In this model, user experience drives engagement, leading to better conversions and sales.

Houzz recognized the need to enhance user experience on their platform and boost conversions. One key strategy was to improve the accuracy and relevance of the search results and recommendations. Houzz’s current search engine was powered by an NLP model that struggled to understand search intent in user queries and effectively rank search results. While the engine can easily handle direct-match searches, its accuracy dramatically drops when processing more specific, detailed, and varied search queries (i.e. long-tail search queries). As a result, customers are often directed to pages that do not align with their search intent, or to pages that are not properly optimized for conversion.

Houzz needed to better understand search intent to help them:

• Enhance user experience by providing relevant and accurate search results to platform users
• Boost purchase rates by quickly matching user queries with the search results optimized for conversion
• Leverage sales & marketing resources effectively, reducing ad spend and customer acquisition costs while allowing them to focus on more qualified leads and complex sales

The leaders at Houzz saw an opportunity to leverage generative AI technology to enhance their search capabilities. Their idea was to enhance the performance of their search and recommendation engine with GenAI by more accurately identifying and interpreting categories and attributes in all types of user queries, ensuring that each search leads to precise results, whether it is an article, a forum topic, a professional, or a product.

Realizing the need for expert assistance with their GenAI project, Houzz reached out to Provectus, an AI-first consultancy with AWS Generative AI Competency. The companies agreed to collaborate on developing a prototype of the generative AI solution for semantic query understanding, primarily focused on product searches during the first stage.

This effort served as an experiment with generative AI, boosting the client’s confidence in the technology and paving the way to scale the solution to cover other search categories on the platform, including articles, photos, forum topics, professionals, and more.

Leveraging the AWS Generative AI Stack and Anthropic’s LLMs to Improve Houzz’s Search Engine

At Provectus, we prioritize GenAI use cases that have the potential to deliver the most value to businesses, without requiring major operational overhauls. Our process includes:

• Starting with a discovery phase to identify the client’s most impactful GenAI use case(s)
• Building and deploying a solution prototype to evaluate the viability of the selected GenAI use case
• Scaling the MVP across the client’s organization to measure success and drive value, enterprise-wide

This strategy ensures that our clients can quickly achieve significant, real-world impact with minimal disruption to their ongoing business and technology processes.

With this strategy in mind, Provectus engaged in detailed discovery sessions with Houzz. We spent considerable time researching their challenges and objectives, and studying their business model. During the discovery phase, our goal was to assess the current state of Houzz’s search and recommendation engine, and to finalize our approach to the development of the GenAI solution’s prototype.

Several Major Challenges Were Identified

1. Houzz’s NLP search engine model could accurately process less than 40% of long-tail queries. The engine could neither rank products nor apply filters as effectively as was required. We proposed implementing advanced semantic understanding, to improve the accuracy of query analysis while ensuring low-latency processing.
2. Houzz wanted to avoid collecting and preparing an extensive dataset of ground truth data. To address this, Provectus leveraged synthetic query generation to prepare a wide range of data combinations of categories and attributes, emulating user queries. This synthetic data was then used to train the NER model and classifiers.
3. Detecting residual parts of queries that are neither categories nor attributes was an important requirement. Provectus proposed training a NER model to classify parts of queries as either categories and attributes, or residual, to improve search accuracy and relevance, while meeting the latency requirement.

Named Entity Recognition (NER) is an NLP technique for identification, classification, and categorization of specific text entities, like search queries, into predefined categories. NER is useful for such tasks as information retrieval, summarization, and cognitive search.

For Provectus, these combined challenges and requirements meant that:

• Out-of-the-box LLMs could not be used to process queries with a required latency
• Queries had to be converted into embeddings first for ease and speed of processing
• Several ML models needed to be trained to identify categories and attributes in queries converted into embeddings as step one
• A specific NER model had to be trained to identify residual parts of converted queries as step two
• All models had to be trained predominantly on synthetic data

The major constraint to using out-of-the-box LLMs for query processing was the sub-second response time requirement — none of the tested LLMs could produce categories and attributes for a given query within a required timeframe. Instead, Provectus utilized Amazon Titan Text Embeddings, hosted on Amazon Bedrock, to generate query embeddings. Next, simple classifiers were trained on query embeddings to identify categories and attributes in search queries. Finally, these “embeddings” models were used during inference for feature extraction for downstream ML models built on top.

The NER model of choice was powered by Flair, a framework for state-of-the-art NLP. Flair has simple interfaces that allow for using and combining different word and document embeddings.

Because Houzz was reluctant to engage in any complex, costly, and labor-intensive data labeling effort, Provectus leveraged Anthropic’s Claude 3.5 Sonnet LLM to generate synthetic data. This synthetic data was used to train both the NER model and downstream ML models for the classification of product categories and attributes.

All models were selected and optimized using Weights & Biases, an MLOps platform that helps AI developers streamline their machine learning workflows from end to end.

As a result of the development, Houzz received a solution consisting of two major parts: 1) A bundle of ML models for enhanced semantic search understanding; and 2) Infrastructure for generating synthetic data for new product categories and attributes on their platform.

With an infrastructure for synthetic data generation, Houzz has a clear path forward for further improvements. By generating more synthetic data, and retraining the NER model and the category and attribute classifiers, Houzz can easily add new product categories or improve the coverage of existing ones, thus improving search accuracy and relevance.

Enhancing Search Accuracy and Improving User Experience: The Impact of Generative AI on Houzz

Provectus developed a working prototype of the Houzz GenAI solution for semantic query understanding within two months. The application of advanced semantic understanding perfectly positioned Houzz to enhance their search capabilities while addressing their primary business goals: improving user experience and increasing conversion rates on product pages.

Improved User Experience

1. More accurate and relevant search results. Customers can find desired products more quickly and easily, reducing frustration and improving overall satisfaction
2. Enhanced marketplace value for merchants. Improved search accuracy increases the visibility of products (and, potentially, services of home improvement professionals) to interested buyers
3. Smarter category redirection. Customers are guided to the most relevant product categories and pages, based on their search intent, streamlining the shopping journey

Increased Conversion Rates

1. Targeted upselling opportunities. The improved search accuracy allows for more precise product recommendations, increasing the average order value through relevant upselling
2. Reduced cart abandonment. By quickly guiding customers to products that match their intent, there is less chance of frustration leading to abandoned carts
3. Improved seasonal campaign effectiveness. The enhanced search capabilities enable better alignment of search results with promotional campaigns, boosting conversion rates during key sales periods

Thanks to Provectus enhancements, Houzz’s search and recommendation engine can now understand various customer query types, regardless of their length, structure, vocabulary, or search style. Search intent, rather than direct-match keywords in the query, is now the foundation of accurate searches. By efficiently connecting customers with relevant products, Houzz is poised to see an increase in conversions and purchase rates, benefiting not only Houzz and their customers, but also their merchants.

Technology-wise, the new GenAI-powered semantic understanding model outperformed Houzz’s NLP model. Search category and attribute identification accuracy saw an increase from 52.94% to 78%, representing a nearly 50% improvement in the engine’s ability to correctly understand customer queries. The model’s recall showed substantial growth, rising from 66.98% to 85%, which means that users are now much less likely to miss relevant products in their search results. While achieving these gains, the model also maintained a high precision of 79%, ensuring that increased recall does not come at the cost of result relevance.

A key feature of the Provectus solution is its synthetic data generation capability, which allows Houzz to easily expand and refine search categories as market trends and user needs evolve. The solution is designed for low-latency processing, ensuring that the enhanced capabilities do not come at the cost of user experience.

These advancements lay a strong foundation for Houzz to strengthen its position as a leading platform in the home design and improvement space. By continually improving its search functionality, Houzz is well-positioned to meet the growing demands of both consumers and merchants in the competitive home improvement market.

By addressing the critical challenge of understanding users’ search intents, Houzz was able to improve its platform’s technology foundation by reinventing how users interact with and find value on its vast product and service marketplace. This advancement aligns perfectly with Houzz’s mission to provide the best possible experience for home design enthusiasts and professionals alike. The team at Provectus was happy to support Houzz on this transformative journey.

Moving Forward

1. Learn more about GenAI’s potential from business and technology perspectives in The CxO Guide to Generative AI: Threats and Opportunities
2. Assess the readiness of your organization to adopt GenAI solutions with Generative AI Readiness Assessment with Provectus
3. Develop high-value generative AI use cases with “Generative AI by Provectus” offering