• by Laurie Sullivan  @lauriesullivan, October 29, 2023

Yelp for years invested in building machine learning (ML) and artificial intelligence (AI) models to power its products and features, from XGBoost to neural networks and more recently, large language models (LLMs). Now it is prepared to face some unlikely competition throughout the industry.

An advanced bidding system is one platform Yelp is benefiting from by advancing AI technology. The company recently announced the expansion of its neural networks to improve ad matching, search results, photo classification and Waitlist, among others.

Advertisers might not be interested in Yelp's return on capital per employee, but investors watch the metric. And for Yelp, analysts say, it is rising.

ROCE is a measure of a company's yearly pre-tax profit -- its return -- relative to the capital employed in the business. 

Reports suggest that Yelp has moved from loss to profitability during the past three years of share-price growth, but the company is running into competition aside from Google Reviews.

With a pulse on AI technology advancements, any obstacles in Yelp's path may simply come down just that -- competition. Musician John Legend launched his first tech startup in March, a social app called It’s Good, aimed at giving users personalized food and travel recommendations. 

The company raised $5 million in funding this month from investors including the Silicon Valley venture fund Lightspeed Venture Partners.

Yelp seems prepared to balance any competition it faces on the horizon. Media Daily News caught up with Sam Eaton, CTO at Yelp, to ask about the specific changes that have propelled advancements at the company.

Media Daily News: How have advancements in AI accelerated Yelp's product release and services?

Eaton: Neural networks and other advanced AI models, such as LLMs, enable a faster pace of innovation. The use of AI models allows us to quickly parse and analyze large datasets, including reviews, photos and other business information. All power innovative and helpful features for consumers and businesses.

For example, our investment in LLMs has enabled us to create a product called review highlights, which surface relevant information from reviews in search results for nuanced queries. Similarly, neural networks have accelerated our advertising bidding system, enabling us to analyze many more signals about users’ interactions with ads to provide ads that are better matched to their intents.

MDN:  When did Yelp start using neural networks, and why only now has the industry seen these advancements to support services?

Eaton:  We’ve been using neural networks to power features and functions across our platform for almost a decade.

Neural networks power the photo categorization and identification systems that inform our Popular Dishes feature, which we first announced in 2018. We have continued to use neural networks across Yelp, enhancing advertising and Waitlist functionalities. What has changed more recently is a combination of both advances in the technology itself, allowing new classes of problems to be addressed, like the explosion in LLM capabilities and usage, as well as more awareness and interest in the usage of AI.

MDN:  How quickly can the platform process years worth of data to better target ads or serve more accurate search queries?  How does this compare with prior?

We see approximately a 2x speed-up in actual training speed with our neural network-based systems.

But even more significantly, we see increasing returns in model performance with larger data set sizes, as opposed to our older systems, which didn’t scale as well with larger volumes.

MDN:  What's the future look like for Yelp with neural networks technology?

Eaton:  We see neural networks driving even more growth for Yelp as we continue to invest in new capabilities, like an even more helpful and relevant search experience that understands the nuance of natural language, and more advanced ad systems that drive relevant leads to advertisers.

We continue to develop our own specialized internal neural net systems based on unique datasets, but also build on top of the growth in open-source and commercial pre-trained models, allowing us to develop new features even more quickly.