Yelp

Yelp曾长期被指控通过算法操控评分来推动商家购买广告服务

评分通胀（Rating Inflation）现象在学术界已有大量研究，Amazon、Yelp等平台均曾面临类似问题

UGC (user-generated content) is the core asset of local life platforms such as Yelp, referring to reviews, ratings, photos and posts contributed spontaneously by users.

Yelp has embedded new AI features in its app, including AI calling restaurants on users' behalf to ask about wait times.

Yelp's new AI features also include a personalized recommendation system that can remember user preferences.

Yelp's traditional role is as an aggregator of local life information: users search, browse reviews, make their own decisions, and then manually call or visit the store.

Yelp's AI phone assistant requires the model to understand user intent, initiate the call, parse the merchant's reply and report the result back.

Yelp's newly introduced memory-based recommendation is a combination of content-based personalization and conversational memory, where the system explicitly stores user preference signals and continuously calls on them in subsequent interactions.

Yelp's core assets are its local merchant relationships and review content, and AI agents can activate both ends: one end saves users communication costs, while the other brings merchants more precisely targeted foot traffic.

It is currently unclear what range of merchants the Yelp AI phone assistant feature covers and in which regions it is available.

Yelp currently only accepts AI Agents that have signed a paid data licensing agreement, and unsigned Agents fail outright when they try to make a reservation or join a waitlist.

In the Yelp review task, the binary labels come from star-rating thresholds, but the question asks 'based solely on the review text, is the reviewer clearly positive overall'; the two do not necessarily express the same judgment, causing the model to receive contradictory supervision signals.

After the researchers reformulated explicit criteria, 38 problematic questions were ultimately excluded from supervision, and the fixed training used 12,520 records and 15,476 questions.

On 194 Yelp text judgments, B1 (0.8B) got 184 correct and B2 (4B) got 192 correct; the average NLLs were 0.09620 and 0.04024, respectively.

Existing tools such as Google Maps and Yelp can only provide vague information like 'usually a bit busy at this time' and cannot provide real-time queue wait times.

Platforms such as Google and Yelp have algorithms that detect anomalies (such as user reports and sudden drops in visitor check-in data), but these signals themselves have latency.

OpSpam, Amazon, and Yelp are the three most widely used benchmark datasets in the field of fake review detection, but their label sources differ.