Methodology

An index is only useful if you can audit it. Everything below is the actual calculation — inputs, weights, and limits — so you can judge whether the numbers mean what you need them to mean.

Pipeline

1

Ingest

Episodes are pulled from public RSS feeds on a weekly cycle. The source panel is fixed per vertical and versioned; panel changes are recorded with the date they took effect.

2

Transcribe

Audio is transcribed, then the audio file is deleted. No source recordings are retained.

3

Resolve entities

Entity detection uses a curated alias and nickname dictionary with word-boundary matching, plus pronoun-based attribution for follow-on references.

4

Derive signals

For each entity and episode: mention count, talk time, and context. Transcripts are retained in a private internal store used for deduplication and re-scoring; they are never published or redistributed.

5

Score

Signals are combined into JORDN, PIPN, and RODMN, then written to the database and exposed via API.

JORDN — attention

Measures who is being discussed right now. Combines podcast mentions, talk time, social signals, and search interest.

ComponentBase weight
Talk time30%
Mentions25%
Google Trends20%
Reddit mentions15%
Wikipedia views10%

Weights apply to standardised inputs. Each component is z-score normalised against the population before weighting, so these percentages are weights on standardised deviations from the population mean — not shares of a raw total.

Scale. Scores are rescaled so the population mean is exactly 100. There is no fixed floor and no fixed ceiling. A score of 140 means 40% above the population average for that period; it is not an absolute quantity and is only comparable within the same population and period.

Weighting adapts to data completeness. A component with sparse coverage in a given period is down-weighted, after which all weights are bounded and renormalised. Weighting responds to how much data exists, not to observed outcomes — nothing about the model is trained or self-adjusting.

Interaction effects. The model detects combinations rather than treating components as independent. Three rules apply a multiplier to the weighted base score:

  • Talk time and mentions both more than one standard deviation above the mean → ×1.3
  • Google Trends more than one standard deviation above the mean with below-average talk time → ×1.2
  • Reddit mentions more than one standard deviation above the mean → ×1.15

PIPN — attention efficiency

Measures whether an entity receives more or less discussion than their social media reach would predict.

PIPN = JORDN percentile − Reach percentile

Scale. −100 to +100. Zero means conversation volume is in line with reach. Positive means more discussed than reach predicts; negative means less.

Reach baseline. Follower counts from Instagram, TikTok, and Twitter/X, combined on a log scale so that additional followers count for progressively less:

Instagram1.5 × log(followers + 1)
TikTok1.0 × log(followers + 1)
Twitter / X0.75 × log(followers + 1)
Combined total0.3 × log(all followers + 1)

The resulting reach score is percentile-ranked within the population, and refreshed weekly as part of each analysis run.

Exclusions. PIPN is only produced for people. Entities with no social media data are excluded rather than scored as zero.

Read it as descriptive, not evaluative. A negative PIPN says an entity is discussed less than their following would suggest. It does not say they deserve less attention, and it is not a judgment about the entity.

RODMN — divisiveness

Measures how polarised discussion of an entity is, on a 0–10 scale. Independent of popularity — a little-known entity can score high.

RODMN describes the coverage, not the entity. A high score means commentators disagree about someone; it is not an assessment of that person's conduct or character.

Full RODMN methodology is documented for clients on request.

Data sources and rights

Access
All feeds currently ingested are public podcast RSS feeds requiring no authentication. No paywalled sources.
Audio
Transcribed then deleted. Never retained or redistributed.
Retention
Derived data — entity, timestamp, talk time, context — is what powers the scores. Transcripts are retained in a private internal store used for deduplication and re-scoring, and are never published or redistributed.
Opt-out
A publisher can request exclusion by contacting us; the feed is removed from the source panel.
Output
Counts, durations, and scores. No source content republished.

Known limitations

Panel-bounded

An entity discussed only on shows outside the source panel won't register. Coverage is a function of the panel, not of the world.

Transcription error

Names, nicknames, and non-English pronunciations are the most common failure mode.

English-language only

Non-English sources are not ingested.

Attention, not sentiment

JORDN counts discussion. It does not distinguish praise from criticism.

Relative, not absolute

Scores are normalised within a population. Cross-population comparison requires care.

No causal claim

The metrics describe what is being discussed. They do not establish why, or what follows.

Validation

Independent validation is in progress. We publish our full methodology so that scores can be assessed directly against your own outcome data.