particle_podcast_compare_share_of_voice

Compare how two to five companies, people or entities are spoken of on podcasts, side by side. It answers “how is NVIDIA spoken of vs AMD vs Intel?” in one call. Each subject’s mention lines over the same window are grouped by their label on one speaker statement characteristics axis, and each label’s share is of that subject’s scored lines.

by="sentiment" (the default) is how negatively or positively each subject is spoken of. certainty, specificity and temporal_orientation work the same way.
By default it counts guests’ and panelists’ lines, as particle_podcast_get_share_of_voice does. speakers="all" adds hosts.
breakdown splits each label group by a second axis: by="temporal_orientation" with breakdown="certainty" is how sure each subject’s predictions are.
speaker_statement_characteristics counts only lines said a certain way, and podcast_slug or publisher_slug narrows to one show or network. Both apply to every subject.

Each subject is exactly what particle_podcast_get_share_of_voice returns for it with the same arguments and by set to the axis, so the numbers agree. Not every line carries labels, so a subject’s shares are of its scored lines, and a subject with fewer than 20 is marked low sample: read its shares as examples, not a distribution. Ad reads are left out unless include_ads is set. The subjects are split at the same time, so a call over five takes about as long as the slowest one. For which professions say it, use particle_podcast_get_share_of_voice one subject at a time. To read the lines behind a cell, pass the subject and the label as speaker_statement_characteristics to particle_podcast_find_mentions. For mention volume over time, use particle_podcast_get_mention_trends.

Inputs

Field	Type	Required	Default	Description
`company_slugs`	string[] (2–5)	one of	—	Company slugs or domains from `particle_company_resolve`, in the order to show them (e.g. `["nvidia", "amd", "intel"]`). Exactly one of `company_slugs`, `person_slugs` or `entity_slugs` is required.
`person_slugs`	string[] (2–5)	one of	—	Person slugs from `particle_person_resolve` (e.g. `["sam-altman", "dario-amodei"]`).
`entity_slugs`	string[] (2–5)	one of	—	Knowledge-graph entity slugs from `particle_entity_resolve`, for anything that isn’t a person or company (e.g. `["germany", "france"]`).
`by`	string	no	`sentiment`	The axis whose labels each subject’s lines are grouped by: `sentiment`, `certainty`, `specificity` or `temporal_orientation`.
`breakdown`	string	no	—	Split each label group by a second axis. Must differ from `by`.
`speakers`	string	no	`guests`	Whose lines count: `guests` (guests and panelists) or `all` (hosts included).
`podcast_slug`	string	no	—	Only mentions inside this podcast (slug from `particle_podcast_resolve`).
`publisher_slug`	string	no	—	Only mentions on podcasts of this publisher (network), by slug.
`include_ads`	boolean	no	`false`	Count mentions inside ad reads too.
`published_after`	string	no	30 days before `published_before`	Only episodes published on or after this date (`YYYY-MM-DD`). The window may span at most 365 days.
`published_before`	string	no	now	Only episodes published on or before this date (`YYYY-MM-DD`).
`speaker_statement_characteristics`	object	no	—	Count only the lines said a certain way, for every subject: `{"temporal_orientation": ["future"]}`.

Passing no subject list returns missing_parameter, and passing two lists returns invalid_input. Fewer than two subjects or more than five, an occupation or field as by, or a window longer than 365 days returns invalid_parameter. A slug that matches nothing returns unresolved_reference naming it. A company or person with no linked knowledge-graph entity, a slug named twice, and two slugs that resolve to the same entity are refused by name too, so a missing subject is never shown as one nobody mentions.

Output

A markdown document headed ## Share of voice compared — by <axis>, <guests and panelists|every speaker>, then a **Window:** row (start to end (end excluded)) and a table with one row per subject, in the order given:

Subject — Name (slug).
Scored — the subject’s lines that carry labels, which its shares are of, followed by (low sample) below 20.
One column per label in the axis’s order, each share (lines). A zero share shows as -.

With breakdown, a ### Each <by> label by <breakdown> section adds a second table with one row per subject and label, its scored lines, and one column per breakdown label. Closing paragraphs restate what the shares are of and point to particle_podcast_find_mentions with the label as speaker_statement_characteristics. With output_format="json", the response is {by, breakdown, start, end, speakers, characteristics_version, subjects}, where each subject is the full particle_podcast_get_share_of_voice JSON for it. characteristics_version is present only when every subject that read labels read the same version. Sample (company_slugs=["nvidia", "amd", "intel"], defaults):

## Share of voice compared — by sentiment, guests and panelists

**Window:** 2026-09-27 to 2026-10-27 (end excluded)

| Subject | Scored | negative | neutral | positive |
|------|------|------|------|------|
| Nvidia (nvidia) | 3928 | 5.20% (205) | 82.30% (3232) | 12.50% (491) |
| AMD (amd) | 495 | 1.60% (8) | 82.60% (409) | 15.80% (78) |
| Intel (intel) | 1112 | 2.80% (31) | 88.70% (986) | 8.50% (95) |

Shares are of each subject's scored mentions, the lines that carry labels; a subject with fewer than 20 is low sample, so read its shares as examples.

Read the lines behind a cell with particle_podcast_find_mentions, the subject, and speaker_statement_characteristics {"sentiment": ["<label>"]}.

Get started

Tools

particle_podcast_compare_share_of_voice

Inputs

Output

​Inputs

​Output

Inputs

Output