The Sentiment page no longer pulls full scan rows to populate the summary cards. It now fetches only the lightweight fields those cards actually read, which cuts several megabytes of LLM response data per row out of the network payload. Pages that were noticeably slow to load should feel much snappier.
The Yelp advertiser study has gone through several rounds of corrections this cycle. The size-controlled regression has been removed entirely — the dataset lacked review counts for non-advertisers, so the coefficient was not meaningful and has been retracted. What remains are directly-counted findings: roughly 87% of AI local recommendations are Yelp-listed businesses, and roughly 13% are confirmed Yelp advertisers (a floor, not a ceiling). The page now makes no causal claim. Off-category ad injections — advertisers Yelp places into the wrong category's slots, which made up 41% of what had been counted as "advertisers" — are now excluded from all counts.
The study has also been rebuilt around a per-platform comparison across Claude, ChatGPT, and Perplexity, with a per-industry breakdown and a per-city drill-down within each industry. Business names now link directly to their Yelp pages. The page payload dropped from 2.3MB to 0.7MB after removing tabs that drew from incomplete early-batch data where a third of LLM responses were refusals counted as non-mentions.