How Perplexity picks its sources: a 1,000-query analysis

Analysis of 1,000 sector-specific Perplexity queries in French-speaking B2B verticals. Five source families dominate, in a deterministic order every CMO should know.

1. Analysis methodology

1,000 queries were run between 8 January and 22 May 2026. Five hundred on the GEO Rocket portfolio (coffee, wine, alcohol-free, premium hospitality, niche perfumery), five hundred on control verticals outside the portfolio (B2B software, consulting, private banking, fashion, contemporary art). Each response was parsed to identify cited sources, their nature, rank, and verbatim word counts.

The goal was not to audit our portfolio, but to understand the macro distribution of Perplexity citations on premium French-speaking verticals. Conclusions match in five cases out of six with the Yang reference study on arXiv 2507.04881 (July 2025, 366,000 citations across 65,000 LLM responses).

Across 1,000 sectoral B2B French-speaking Perplexity queries, five source families capture 82% of citations. The distribution is strikingly stable from one vertical to another.

2. Family 1: Wikipedia and Wikidata

Wikipedia appears in 13% of responses, Wikidata indirectly in 54% via sourced entities. The most predictable family: on a definitional or contextual question, Perplexity hits Wikipedia almost systematically.

The implication for a French-speaking B2B brand is threefold. First, creating a Wikidata entity is absolute priority. Second, aiming for a Wikipedia page (respecting eligibility criteria) mechanically doubles citation rate on definitional queries. Third, controlling sameAs links between the two is what lets LLMs tie the brand to its category, sector, leadership.

3. Family 2: forums and community platforms

Reddit, Stack Exchange, Quora and their French-speaking equivalents (sectoral professional forums) appear in 12% of Perplexity responses. The least manipulable family, but the most profitable when a product or service gets mentioned spontaneously.

A French-speaking B2B brand that does not follow Reddit, especially English-language sectoral subreddits, ignores roughly one fifth of the raw material Perplexity uses to answer. PROEMA standard recommends weekly monitoring of five sectoral subreddits with brand mention alerts.

4. Family 3: verifiable tier-1 press

Reuters, Bloomberg, The Guardian, AFP, and French-speaking equivalents (Le Monde, Les Échos, L’Echo in Belgium, Le Soir, La Libre, Paris Match) represent 14% of citations in our sample. These sources are favoured because they hold knowledge graph-recognised domain authority and a documented fact-checking history.

For a brand the stake is dual: obtaining tier-1 press mentions (at least one or two per year), and marking them in Wikidata via P973 (described at URL). LLMs consult this property when enriching entities.

FamilyCitation shareManipulable?Priority lever
Wikipedia + Wikidata13% + 54% via entitiesPartiallyCreate entity, target page
Community forums12%HardlyMonitoring + product quality
Tier-1 press14%IndirectlyPR + P973 marking
Institutional sites11%NoOfficial partnerships
Brand and editorial32%DirectlyGEO-native site + dense Schema

5. Family 4: institutional sites

.gov, .edu, .europa.eu domains, and French-speaking equivalents (.be for Belgian institutions, .fr for French institutions) appear in 11% of responses. Nearly non-manipulable, but they strongly orient the selection of other families: a brand cited in a sectoral report by an official body sees its probability of appearing in the other families multiplied by three.

6. Family 5: brand sites and specialised editorial

The only directly steerable family. 32% of Perplexity citations come from brand sites or specialised editorial (sectoral blogs, vertical media like Decanter on wine, Sprudge on coffee, Hotelmanagement.net on hospitality). This is where GEO plays out.

A sectoral GEO-native site can, in six to twelve months, hold a dominant share of this fifth family for its vertical. That is exactly the GEO Rocket portfolio strategy: expertcafe.be on coffee, zeroproof.one on alcohol-free, expertvin.be on wine. Each installs itself as default reference on its category.

Perplexity cites five source families in a deterministic order. Four are inherited or indirect. The fifth is built. That is where GEO is won or lost.

source: Proema Insight · 1,000-query study Q2 2026

The PROEMA GEO diagnostic measures Perplexity citation distribution across the five families for a given brand. Read in five minutes, prioritised action plan on the three directly steerable families.

Get started

Is your brand citable by AI?

Receive your free GEO score and then request a full GEO diagnostic: what ChatGPT, Perplexity and Gemini do (or don’t) see about you today.