Methodology

Pulvian aggregates publicly available information from the sources listed below. No private data, no scraping of non-public content, and no paid insider information is used. Each data point is timestamped and stored as a snapshot — the score reflects conditions at a specific point in time, not a real-time feed.

Company metadata (founding year, headcount range, industry) is sourced from Wikidata and supplemented by a curated seed dataset. Metadata is not used in scoring — only in filtering and display.

Note on Trustpilot data: Trustpilot reviews reflect customer experience, not employee experience. They are included as a supplementary risk proxy only — sustained low customer trust (≤ 2.5 / 5) can indicate operational or reputational dysfunction. Trustpilot signals are weighted conservatively and are never the sole basis for a high stress or risk classification.

Every company snapshot produces three scores, each on a 0–100 scale:

Confidence is the most important score to look at before interpreting stress or risk. A stress score of 75 backed by 3 reviews is very different from the same score backed by 800 reviews across multiple sources. The confidence score encodes that difference.

When confidence is below ~30, treat the directional finding as a weak or early signal. When confidence is above ~70, the finding is more reliable — though still not a guarantee. The coverage badge visible on each company page gives a plain-language summary (Strong / Moderate / Limited) derived from source count, review count, and confidence score.

Not all signals carry equal weight over time. Pulvian applies exponential decay to older data points using a 90-day half-life: a signal that is 90 days old contributes roughly half as much to the score as an equivalent signal collected today.

This means recent spikes in negative sentiment matter more than a stale review from two years ago — which better reflects how workplace conditions actually evolve. The snapshot date shown on each page reflects when the most recent data was ingested, not when the company was first added.

In addition to numeric scores, each snapshot carries a categorised level (low / medium / high) derived from fixed thresholds applied to the numeric score. These thresholds do not currently vary by industry or company size.

Future versions may introduce industry-relative benchmarking, which would make the level more contextually meaningful.

Where available, snapshots include a short summary of the signals found. Summaries are currently generated by a deterministic, rule-based system that selects phrases based on score levels and signal content. Each summary is labelled with the model that produced it. They are supplementary context, not editorial opinions.

Positives and negatives listed in a summary reflect patterns in the source data. When no summary is available, the scorecard is still shown — scores do not depend on summary generation.

Signals are extracted from public documents using a deterministic, keyword-based rule engine. Each rule maps a set of phrases (e.g. "layoffs", "hiring freeze", "record revenue") to a sentiment value. When a keyword is matched, the surrounding context is checked for negation cues — words like "avoids", "no", "not", or "despite" that reverse the meaning. Negated matches are suppressed to reduce false positives.

This approach is transparent and reproducible: the same document always produces the same signal. It does not use machine learning or large language models for scoring.

The Similar companies section on each company page shows other companies in the same industry, ordered by score proximity. This helps you explore related organisations and notice patterns across a sector.

The Compare feature lets you view two or more companies side by side. Scores, levels, and metadata are presented in parallel columns so differences are immediately visible. Comparison does not introduce any additional scoring logic — it is a presentation layer over the same underlying snapshots.

• Uneven coverage. Publicly traded and large established companies typically have higher review volumes and therefore stronger confidence scores. Startups and smaller companies are tracked wherever public data exists, but confidence scores will reflect the available coverage — always check the coverage badge before interpreting the score.
• No private data. Pulvian only sees what is publicly posted. Internal culture that is never written about online is invisible to the model.
• No real-time feed. Snapshots have a date. Significant events (layoffs, leadership changes) that occur after the last snapshot are not reflected until the company is rescored.
• Fixed thresholds. Stress and risk levels use static cut-offs, not industry or region-adjusted benchmarks.
• English-language bias. Review platforms skew toward English-language content. Companies whose employees primarily write in other languages may be underrepresented.

• Not a replacement for due diligence or professional HR advice
• Not a ranking of "best" or "worst" employers
• Not derived from proprietary, private, or leaked data
• Not a real-time service — always check the snapshot date