What our data represents
Our scores summarize publicly available online discussion about a product. They are a reflection of what people are saying in public forums, not a statement of fact about the product, not an audit, and not our own opinion of its quality. A high or low score means the public conversation skewed positive or negative, nothing more.
Where the data comes from
Our current sources are Hacker News, accessed through its public search interface, and YouTube, where we collect top public comments on review videos about each product through the official YouTube Data API. For YouTube comments, the title of the video they respond to is stored alongside the comment for context. Each product page lists the exact sources and mention counts behind its data. We are integrating additional public sources, and as each one comes online this page and the per-product source breakdowns will reflect it. We store short excerpts and links back to the original posts for context. We do not republish entire threads or reproduce large portions of any source.
How mentions are collected and filtered
We search each source for mentions of every tracked product, including common alternate spellings, and apply filters to exclude obvious mismatches such as unrelated uses of the same word. Each remaining mention is then checked by an automated language model for relevance. Mentions judged not to be about the product are recorded but excluded from all scores. Very short mentions with little signal are also excluded from scoring.
Sampling
For widely discussed products, we classify a sample of up to 200 mentions per product per week rather than every mention, prioritized by engagement such as upvotes, with longer substantive posts preferred among comments. The full set of collected mentions is retained in our archive. Sample sizes are disclosed on every product page.
How mentions are classified
Each sampled mention is analyzed by an automated language model that estimates its tone (positive, negative, neutral, or mixed), assigns a sentiment value between -1 and 1, and tags the topics it discusses, such as pricing, ease of use, performance, or reliability, from a fixed vocabulary. We record the model and prompt version used for every classification so that scores remain comparable as the system evolves. Classification is automated and probabilistic, so individual classifications can be wrong.
How the Pulse Score is computed
Scores are computed weekly, per product, from complete calendar weeks only. An in-progress week is never used for a published score, because partial weeks have small samples that swing wildly. For each complete week, we average the sentiment values of that week's relevant sampled mentions, with each mention counting equally, and map the average linearly to a 0-100 scale: a week where sentiment averaged neutral scores 50, uniformly positive approaches 100, uniformly negative approaches 0. The score shown on a product page is the most recent complete week. We maintain one canonical page per product, so all discussion about it is reflected in one place.
Minimum sample sizes
We do not publish a score for a product until it has at least 30 relevant mentions in our archive. Below that threshold, the product page says so plainly and shows the current count instead of a score. Small samples produce unreliable numbers, and we would rather show no score than a misleading one.
Known limitations
Online discussion is self-selected and skews toward people motivated to post, so it may not represent the full user base. Different communities skew differently: technical forums are often more critical than general ones, and scores reflect the mix of sources listed on each product page. Sentiment can be affected by sarcasm, bots, coordinated campaigns, or a single viral thread. Automated classification makes mistakes. Scores shift in response to launches, outages, and news. Volume varies a lot between popular and niche products. We surface these numbers to describe the conversation, not to certify or rank product quality, and they should be read with that in mind.
Corrections and business feedback
We want this to be fair to the products we cover. If you represent a product and believe the data misrepresents it, or you want to add context, you can send us feedback. We read every submission and review them at our discretion. If you identify a factual error in how we describe a source or method, let us know and we will review it.
Updates
New mentions are collected daily and scores are recalculated weekly as complete weeks close. This methodology is versioned; material changes will be reflected on this page with the version noted above.
This page describes our approach in good faith for transparency. It does not create any warranty or guarantee as to the accuracy, completeness, or fitness for any purpose of the data, which is provided as described in our Terms of Use.