How Narrative Compass Works

Narrative Compass is a media-literacy tool. It watches the public RSS headline feeds of 8 far-right / hyper-partisan outlets and 10 mainstream baseline outlets, finds stories covered by both groups, and compares how each side frames the same event.

The pipeline has five stages:

1. 1

   Collect

   Pull the latest headlines from each outlet's public RSS feed.

2. 2

   Cluster

   Group headlines that share key terms into candidate narratives.

3. 3

   Score

   Compute a 0–100 Velocity Score from outlet saturation, Google Trends, and Reddit.

4. 4

   Analyze

   Fetch the article text and ask an AI model (Google Gemini) to compare framing.

5. 5

   Present

   Show the results with the original headlines and sources so you can verify them.

All inputs are publicly available. No private data, paywall bypassing of live sites, or user tracking is used. For paywalled mainstream outlets, the tool tries the Internet Archive's public copy of the page; if no text can be retrieved, the comparison falls back to the headline and RSS excerpt only.

A “narrative” here is a news event that several partisan outlets are covering at the same time and that mainstream outlets also cover. The selection is automated and rule-based:

1. 1

   Keyword extraction

   Each headline is lowercased, punctuation is removed, and it is split into words. Words of 4 letters or fewer and common stop-words (e.g. “the”, “after”, “said”) are discarded. The remaining words are the headline's keywords.

2. 2

   Clustering on far-right feeds

   Every keyword becomes a cluster. Each far-right headline is added to the cluster of every keyword it contains (duplicates by link are ignored).

3. 3

   Matching mainstream coverage

   Mainstream headlines are matched to the same keyword clusters. Mainstream headlines never create new clusters, so the tool only follows stories the partisan outlets are pushing.

4. 4

   Eligibility filter

   A cluster qualifies only if at least 2 far-right headlines and at least 1 mainstream headline share the keyword.

5. 5

   Ranking and cut-off

   Qualifying clusters are ranked by the number of far-right headlines. Only the top few are analyzed in a live scan (the number is capped to keep AI usage limited).

6. 6

   Article pairing

   For each selected cluster, the first far-right article and the first mainstream article in the cluster are compared side by side.

The Velocity Score estimates how strongly a narrative is being amplified right now. It is the sum of three components, capped at 100:

The score measures attention and amplification, not truth. A high score means a story is widely pushed, not that it is false. If Google Trends or Reddit cannot be reached, that component falls back to its baseline (10 for Trends, 0 for Reddit).

“Opinion bleed” means editorial opinion, emotional language, or framing mixed into what is presented as straight news. This part of the analysis is performed by an AI model (Google Gemini), not by a fixed formula.

1. 1

   Text retrieval

   The article page is downloaded, scripts, styles, navigation, header and footer are stripped, and up to the first 15 substantial sentences (over 60 characters) are kept. The text is further limited to 8,000 characters before being sent to the model.

2. 2

   Structured prompt

   The model receives both articles and is told to act as a nonpartisan media analyst, to treat article text strictly as data (ignoring any instructions inside it), and to return a fixed JSON structure.

3. 3

   Fields produced

   Topic title; the narrative being pushed; the person or group framed as the antagonist; a fact-check anchor (the verifiable core of what happened); for the mainstream article — focus, factual-tone rating, and loaded rhetoric; for the partisan article — a News vs. Opinion ratio, quoted loaded phrases with an explanation of each, and facts that were omitted.

4. 4

   Validation

   The response is checked against a strict schema. If it cannot be parsed or validated, it is discarded rather than shown.

The Trending Narrative view is designed to show the top 3 stories dominating the Trending and Top Stories areas of far-right outlets, and then answer: how do all covering outlets frame it, and how does that differ from the mainstream baseline?

• Ranking: stories are ranked by the number of partisan outlets covering them and by Velocity Score (same formula as above).
• Far-right panel: consensus framing across outlets, dominant rhetoric, the shared “villain”, and the average opinion-bleed estimate.
• Mainstream panel: the baseline's factual focus, tone, and consensus framing.
• Divergence analysis: where the two sides differ and which context is omitted.
• Overview: recurring themes, rhetorical tactics, and what is absent across the top stories.

• Scheduled refresh: data is refreshed automatically three times a day, at 8:00 AM, 12:00 PM and 4:00 PM Eastern Time. Visitors cannot trigger a new scan; the “Latest Refresh” button simply shows when the data was last refreshed.
• AI usage limits: each scan analyzes only the top few narratives, identical article pairs are cached for 6 hours, and a daily cap on AI calls applies. When the cap is reached, the site shows clearly labeled demonstration data instead.
• Demo data is labeled: a banner appears whenever demonstration data is shown instead of live results.
• Not a fact-checker: the tool highlights framing differences and attention levels. It does not rule on whether any outlet is right or wrong.
• Known limitations: keyword clustering is approximate; one article per side is compared; RSS feeds only expose recent headlines; paywalled or blocked pages may fall back to headline-level analysis; Reddit is sampled from a single community; AI output can be imperfect.