Nearbia
A technology team focused on the collection and classification of local business website data, using browser rendering and local LLM pipelines to handle website determination tasks.
Timeline (last 90 days)
The Nearbia team used a single HTTP request to check 200 local business websites, and the result missed more than 30% dead sites and a large number of marketing sites.
The initial HTTP request checker sent one request to each homepage, captured the first 400KB of raw HTML, and searched for city names and plumbing-related words.
When the check was launched from a computer in the Netherlands, 47 of the 200 sites did not respond; but when retried from a U.S. IP, 15 of those 47 loaded successfully immediately.
Any pipeline that takes "whether a site is alive" as an input must fix the probing node in the target market's location, otherwise normally accessible local businesses will be misjudged as dead sites.
The new pipeline has four steps: rendering, labeling, verification, and localization; rendering uses regular Chrome with JavaScript enabled, with 5 browsers in parallel, and the 171 responsive homepages took about 40 minutes.
In the labeling step, a language model deployed on Nearbia's own servers reads each captured page and assigns a site type: local company's own site, chain branch page, national chain, directory site, lead-generation site, irrelevant site, or dead site.
The verification step requires the model to cite the original page text to support the label, the code confirms that the citation actually appears in the page, and labels with confidence below 0.6 are forwarded for manual review.
Location evidence priority from high to low: 1. the site's phone matches the phone in the source record; 2. the area postal code in the address or structured data; 3. a city name adjacent to "FL" or "Florida"; weak evidence is an area name alone or only the 727 area code.
Phone is regarded as the strongest signal, because lead-generation sites and directory sites usually list only a national customer service or 8xx number, and 8xx is explicitly excluded because chains share it across locations.
In the final classification of the 200 domains, 88 local plumbers with location evidence form the corpus, and 36 are dead sites/parked/no DNS resolution.
3 more timeline events
All Facts (13)
The Nearbia team used a single HTTP request to check 200 local business websites, and the result missed more than 30% dead sites and a large number of marketing sites.
50%UnverifiedThe initial HTTP request checker sent one request to each homepage, captured the first 400KB of raw HTML, and searched for city names and plumbing-related words.
50%UnverifiedWhen the check was launched from a computer in the Netherlands, 47 of the 200 sites did not respond; but when retried from a U.S. IP, 15 of those 47 loaded successfully immediately.
50%UnverifiedAny pipeline that takes "whether a site is alive" as an input must fix the probing node in the target market's location, otherwise normally accessible local businesses will be misjudged as dead sites.
50%UnverifiedThe new pipeline has four steps: rendering, labeling, verification, and localization; rendering uses regular Chrome with JavaScript enabled, with 5 browsers in parallel, and the 171 responsive homepages took about 40 minutes.
50%UnverifiedIn the labeling step, a language model deployed on Nearbia's own servers reads each captured page and assigns a site type: local company's own site, chain branch page, national chain, directory site, lead-generation site, irrelevant site, or dead site.
50%UnverifiedThe verification step requires the model to cite the original page text to support the label, the code confirms that the citation actually appears in the page, and labels with confidence below 0.6 are forwarded for manual review.
50%UnverifiedLocation evidence priority from high to low: 1. the site's phone matches the phone in the source record; 2. the area postal code in the address or structured data; 3. a city name adjacent to "FL" or "Florida"; weak evidence is an area name alone or only the 727 area code.
50%UnverifiedPhone is regarded as the strongest signal, because lead-generation sites and directory sites usually list only a national customer service or 8xx number, and 8xx is explicitly excluded because chains share it across locations.
50%UnverifiedIn the final classification of the 200 domains, 88 local plumbers with location evidence form the corpus, and 36 are dead sites/parked/no DNS resolution.
50%UnverifiedFive domains associated with St. Petersburg actually belong to plumbers in Colorado, Idaho, Mississippi, Broward County and Orange City, Florida.
50%UnverifiedOf the 45 in-city plumbers in the corpus, 21 were not included in the initial database, and Google alone found 11 of them.
50%UnverifiedThe model once labeled a sewer pipe lining company as window cleaning, indicating that the "other industries" group must be manually reviewed.
50%