Join our daily and weekly newsletter for the latest updates and exclusive content on the top AI coverage. Learn more
The app review Yelp has provided benefits to eateries and other consumers for decades. It has been experimenting with the study of the machine since the early years. During the recent explosion of AI technology, it still encountered obstacles as it worked to use modern large language models to empower certain features.
Yelp realizes that customers, especially those who occasionally use the app, have a problem connecting its AI featuresLike the ai-powered assistant.
“One of the obvious lessons we've seen is very easy to build something that looks cool, but it's very difficult to build something that looks cool and very good -benefit,” Craig Saldanha, chief product official at Yelp, told Venturebeat in an interview.
Surely this is not easy. After launching the Yelp assistant, the AIs' assistant in search of the AI service enabled, in April 2024 with a broader customer swathe, Yelp saw figures of use for AI tools that were actually starting to go down.
“What surprised us was when we launched it as a beta to consumers – some users and people familiar with the app – [and they] It is expensive. We got a strong signal that it would be successful, and then we combined it with everyone, [and] The performance just fell, ”Saldanha said. “It took us a long time to find out why.”
It became more casual Yelp users, those who occasionally visit the site or app to find a new tailor or plumber, are not expected to immediately talk to an AI representative.
From simple to more involved features of AI
Most people know Yelp as a website and app to look for restaurant reviews and menu photos. I use Yelp to find food pictures with new eateries and to see if others share my feelings about a particular bland dish. It is also a place that tells me if a coffee store I plan to use as a workspace for the day has a wifi, plug and seating, a rarity in Manhattan.
Saldanha recalls that Yelp was invested in AI “for the better part of a decade.”
“Back to when, I will say in the 2013-2014 timeline, in a different AI generation, so our focus is on developing our own models to do things like understanding the query. Part of the job of making a significant connection is to help people refine their own search goals,” he said.
But as AI continued to change, so did Yelp's needs. It has invested in AI to identify food with photos submitted by users to identify popular dishes, and then it launches new ways to connect with entrepreneurs and services and services and Help Guide to Platform Users' Searches.
Yelp Assistant helps Yelp users find the right “pro” to work. People can tap the chatbox and use any of the signals oi -type the work they need to do. The assistant then asked follow-up questions to narrow down potential service providers before forming a message to pros that could bid for work.
Saldanha said Pros is encouraged to respond to users themselves, though he recognizes that larger brands often have call centers holding messages generated by Yelp's AI assistant.
In addition to Yelp Assistant, Yelp launched insights on review and highlights. LLMs reviewed the user's and examiner's sentiment, which Yelp collects on emotional marks. Yelp uses a detailed GPT-4O Prompt To generate a dataset for a list of topics. Then, this is well with a GPT-4O-mini model.
The featured highlights of the review, which present information from the reviews, also uses an LLM prompt to generate a dataset. However, it is based on the GPT-4, with fine-tuning from the GPT-3.5 turbo. Yelp said it will update the feature with GPT-4O and O1.
Yelp joins many other companies Using LLMs to improve the benefit -benefit of reviews by adding better search operations based on customer comments. For example, for example, Amazon launched Rufusan Ai -enabled Ai that helps people find recommended items.
Large model and performance needs
For many of its new AI features, including AI's assistant, Yelp turned to the GPT-4O of Openai and other models, but Saldanha noted that regardless of the model, Yelp's data was the secret sauce for its assistants. Yelp doesn't want to lock herself in a model and keeps an open thinking about which LLMs will provide the best service for its customers.
“We use models from Openai, Anthropic and other models in AWS Bedrock,” Saldanha said.
Saldanha explained that Yelp created a rubric to test the performance of models in accuracy, relevance, awareness, customer safety and compliance. He said “these are really the leading finishing models” that are best performed. The company runs a small pilot with each model before considering the cost of diverting and responding to latency.
Instruction users
Yelp also began with a combined -with -thee effort to educate both the casuals and power of users to be comfortable with new AI features. Saldanha said one of the first things they realized, especially to the AI's assistant, was that the tone had to feel human. It cannot respond very quickly or too slowly; It cannot be too much encouraging or too brusque.
“We put an effort to help people be comfortable, especially in the first response. It took us about four months to get this second piece. And once we did, it was obvious and you could see the hockey stick in contact,” Saldanha said.
Part of that process is involved in the Yelp assistant training to use a few words and be positive. After all the fine tuning, Saldanha said they were finally seeing higher use numbers for Yelp's AI features.
