Opens Dec 1.Officially operating from December 1. We are still in development and testing. Get notified when orders open:
Methodology
A number with no method behind it is decoration. This page says exactly what goes into a PollyX report, how the 0 to 10 score is arrived at, and - just as importantly - what it cannot tell you. If something here contradicts what you see in a report, the report is wrong and we want to hear about it.
PollyX searches millions of sources, including global news from 100+ countries; X, TikTok, Instagram, YouTube, Reddit, LinkedIn, Threads, Bluesky, Telegram and Twitch; Trustpilot, Google Reviews, Amazon, the App Store, Google Play, Tripadvisor, Booking.com, Airbnb, Yelp, Deliveroo, G2, Capterra, eBay, Etsy and Steam; Hacker News, Substack, Medium, GitHub, Stack Overflow and Product Hunt; Glassdoor, Indeed and kununu; podcasts; and local platforms such as Bilibili and Zhihu in China, PTT in Taiwan, Kaskus in Indonesia, Reclame Aqui and Mercado Libre in Brazil and Zomato in India. Every report draws on several independent layers, gathered in parallel. Which of them run depends on the brand's industry and market, and every report lists exactly the sources it read:
The total is whatever survives that cleanup. It varies by how widely a brand is covered, which is why every report and every score on this site displays its own actual source count and generation date rather than a marketing number.
When PollyX reads a website itself it identifies as PollyXBot and follows the site's robots.txt: how PollyXBot works.
A note on scale first, because source counts are where sentiment tools mislead most. Some tools advertise tens of millions of sources; PollyX searches the same web through global news and web indexes, so the reach is comparable. Think of a library with a million books. Ask what they say about your brand and you do not read a million books: only a few hundred mention your brand at all. The rest are about cooking and dinosaurs. We search the whole library; your report reads the shelf that matters. That is why a PollyX report cites hundreds of sources rather than millions: millions is the size of the library, hundreds is what actually says something about your brand once everything below has been removed. Every report shows its own real count.
Brand-owned content is excluded from sentiment. Press releases, newswire syndication and the brand's own posts tell you what a company says about itself, not what the public thinks of it. Where that material is interesting it may be noted as context, but it never counts toward a score, a theme or a quote.
Sources that turn out to be about a different entity with the same name are dropped. If you supply "also known as" or "not about" terms when creating the report, those feed the search plan and this relevance filter directly.
Other markets, when you ask for one market only. By default a report covers the brand worldwide with its home market foregrounded. Choosing this market only when you create a report is a hard restriction rather than a preference: news is filtered by the publisher's country, web and video search are run against that country, Amazon ratings come from that country's marketplace (and are skipped where it has none), App Store and Google Play ratings and reviews come from that country's store, and coverage of the same brand elsewhere is discarded rather than down-weighted. Reddit, X, TikTok, Instagram, Bluesky, Mastodon and Hacker News posts carry no country, so those are judged individually on whether the poster is writing about your market. A restricted report is usually a thinner report - that is the trade-off, and where the evidence is thin it will say so instead of quietly refilling from elsewhere. Restricting requires a location; without one the report stays worldwide.
This is the part most tools are vague about, so plainly: there is no fixed formula and no fixed weighting between the news, video, web and community layers. The score is not a weighted average, and GDELT's -10 to +10 tone values are not arithmetically rescaled onto 0 to 10.
What actually happens is that all five layers - including GDELT's tone values and article volumes - are supplied as evidence to an AI analyst, which judges the balance of positive against negative opinion and places the brand on this scale:
It weighs how many people are talking, how strongly they feel, how recent the coverage is, and how credible the source is. One deliberate correction is applied: review platforms and complaint forums heavily over-represent dissatisfied customers, because unhappy people write reviews far more often than satisfied ones. A handful of complaint pages is not allowed to drag down a brand that is broadly well regarded.
The honest consequence of a judged rather than computed score is that it is reproducible in direction but not to the decimal. Two runs a day apart on the same brand should agree on whether sentiment is positive, mixed or negative, and should broadly agree on the themes. They may differ by a few tenths. Treat the first digit as the signal and the decimal as resolution, not precision.
Every score is published with the number of sources behind it and the date it was generated. A score with neither is not a measurement, so we do not show one.
The news window reaches back 365 days at its widest, with the most recent coverage weighted most heavily. Web and community results reflect what was publicly visible at generation time.
Each report also carries a confidence level of high, medium or low, reflecting how much real public evidence backs it. A brand with thin coverage gets a report marked low confidence rather than a confident-looking number. Where there is not enough public data to produce an honest report at all, we say so and refund you rather than generating something that reads authoritative and is not.
AI Visibility checks ask each of the 12 buyer questions once per assistant, live at the moment the check runs and never from a cache or an earlier check: eight assistants through the provider's official API with web search switched on, not through the consumer app, which can answer differently, and Google AI Overviews, Google AI Mode and Microsoft Copilot, which offer no API for this, straight from their live results (DeepSeek is asked in its standard non-thinking mode, the one its app uses by default). Assistants are non-deterministic, so a single ask is one sample. The check therefore prints a margin of error beside its headline share, a 95% Wilson interval over the answers collected; a brand's questions are saved on the first check and re-run word for word so later checks compare like with like; and a tracked brand's page shows consistency across its recent checks, how often each question led and how often each assistant named the brand first, rather than treating any one check as a verdict.
If a report contains a factual error, a misattributed quote, or a source that is about a different company, email contact@pollyx.org with the report link. We will correct it and, where the error was material, regenerate the report at no charge.
Millions of sources searched per report, the ones that talk about the brand read and cited. Last reviewed 30 October 2026.