When an AI answer engine recommends software, where does it actually get the answer? We logged every source three engines cited across 20 go-to-market buying questions. One pattern holds everywhere, and one famous one does not.
This is part two of a two-part program. Part one measured what the AIs recommend and how far apart they land. This one asks where those recommendations come from.
Across roughly 464 citations from three engines, 86 percent point to a third party, an independent roundup, review blog, comparison aggregator, or forum thread, rather than to the recommended vendor's own website. Only 14 percent are the recommended vendor citing itself. This holds on every engine we measured.
Where Perplexity's citations concentrate (share of its 20 answers each domain appeared in)
Reddit is the single most-cited domain in Perplexity's set, appearing in 15 of 20 answers. Beyond the top three, citations spread thin across more than two hundred distinct domains, most appearing in only one query. The concentration is in coverage, one community cited almost everywhere, not in a single dominant publisher.
The new depth in this edition. We captured full or near-full citation sets from three engines: Perplexity's native source list, Gemini 2.5 Flash's grounded search API, and Google AI Overviews' domains. Two things stand out. The third-party lean is universal. The Reddit habit is not.
The real, classified source-type mix, no longer an estimate. Every distinct cited domain across all three engines was classified by a published rule (forum, aggregator, recommended-vendor first-party, else third-party review). The full rule and the domain-to-type table live in _citation-analysis.json.
Vendor first-party pages, the recommended tool citing its own site, are a small minority everywhere. The bulk is independent third-party writing. On the two search surfaces a real slice is raw forum and video UGC; on Gemini that slice all but vanishes and third-party editorial swells to fill it.
We pulled one of the most-cited pages, a Zapier comparison roundup, and coded it directly. This is the archetype of what earns a citation in this field: structured, priced, schema-marked, and recently dated.
What is almost entirely absent from the cited set: original first-party measurement.
The cited pages overwhelmingly re-organize the same publicly available vendor facts into a clean, dated, scannable table. When an entire citation pool is consensus-aggregation, a page that runs an original count or test is structurally different from everything around it. We are not claiming a measured citation lift. We are pointing at a gap in what currently gets cited.
Cross-engine agreement across the 14 category and intent queries, comparing the top tool each engine named. The six head-to-head questions are consensus by construction and excluded here. This is the same disagreement thesis as the companion study, seen through the source layer.
Every query we ran, with what each engine recommended. Click any question for the full receipt: each engine's top pick, the ordered list it named, and the actual source domains cited by Perplexity and by Gemini (Reddit highlighted). ChatGPT is tools-only by design; see Method.
FULL all three engines named the same top tool · 2 OF 3 two agreed · SPLIT three different tools · H2H head-to-head question.
Method. On 2026-06-07 we submitted 20 GTM buying-intent queries, drawn from the verticals in a companion 30-tool GTM index and balanced across marketing, SEO/GEO, and sales, to three consumer answer surfaces: Perplexity (default web search), ChatGPT (GPT-5.x, web-search enabled), and Google AI Overviews (default Google Search). For each answer we logged the tools named in order, and for Perplexity the full cited-source URL list. On 2026-07-08 we added a fidelity-upgrade capture: the same 20 queries run through Gemini 2.5 Flash with the grounded google_search tool via the free API, extracting cited source domains from groundingMetadata.groundingChunks (resolving redirect URLs by HEAD request). That gives two engines with high-fidelity full citation sets (Perplexity and Gemini) plus Google AIO domains, roughly 464 classified citations in total. Every distinct domain was classified by a published rule and the domain-to-type table plus the tool-to-domain map are open in _citation-analysis.json. One cited page (zapier.com) was visited and coded directly as a structural exemplar. Raw per-engine captures are retained so any standing here can be reproduced or challenged.
Data. Open under CC-BY-4.0. DOI 10.5281/zenodo.20632767. Base snapshot 2026-06-07; Gemini citation capture added 2026-07-08. AI answers are volatile and will change; re-run the published protocol to reproduce.
Limitations. This is a dated snapshot, not a longitudinal average, and the Gemini layer was captured one month after the base (2026-07-08 vs 2026-06-07), so cross-engine comparisons mix two adjacent dates, disclosed here rather than hidden. Citation fidelity differs by engine and is stated per engine: Perplexity exposes a native numbered source list (high fidelity); Gemini grounding chunks are read directly from the API (high fidelity); Google AI Overviews collapses and intermixes its citations with organic results, so its domains are best-effort and likely over-count forum and Reddit hits; ChatGPT renders citations as in-product chips that were not captured, so it is tools-only and excluded from the source-type analysis entirely. Source-type proportions are now real classified counts, not estimates, but classification of a long-tail domain rests on a documented heuristic (see the published rule). Vendor first-party is defined query-relatively: a citation counts as first-party only when the cited domain belongs to a tool recommended in that same answer, so a vendor's roundup that recommends competitors is treated as third-party, which is the honest reading. No claim is made that Lucreya itself is cited by any engine. Read the numbers as a measurement of a moment, not a permanent ranking.