Prompt Instructions Cannot Replace Missing Market Evidence

What you'll learn#

This article will help you:

• understand why prompt instructions cannot compensate for missing source coverage;
• recognize official, review, community, and competitor-shaped evidence gaps;
• decide what different source gaps imply for content, customer proof, and third-party market presence.

Key findings#

Finding 1 — The instructions changed the observed source mix in some query groups, but the test did not establish why those shifts occurred or which specific pages would appear. An instruction sometimes shifted which source types were present; it did not provide control over which individual pages Perplexity showed.

Finding 2 — The Reddit instruction produced an observable shift across all five query groups in this sample. It was the most consistent change we recorded. Relevant Reddit discussions were available for the five tested query groups, and Reddit was less prominent in the baseline answers.

Finding 3 — A "review" source was not necessarily independent. Of the 116 written review articles classified in this test, 50 were published by software vendors rather than independent review publishers — including content about competing products.

Finding 4 — Removing listicles reduced the number of visible citations but did not establish that the remaining sources were better. Baseline answers contained about 13.5 visible citations on average, compared with 9.9 under the "avoid listicles" instruction.

Finding 5 — Source instructions were more useful as a diagnostic signal than as an optimization tactic. This is a commercial interpretation of the observed results, not a proven account of how Perplexity retrieves or ranks sources.

What this means for your company#

A missing source layer is not a prompting problem. It may mean the company lacks clear official information, credible customer proof, independent coverage, relevant community discussion, or defensible comparison content. Each gap requires a different investment.

If reviews are being written mainly by vendors or competitors, the company may need stronger independent proof. If official pages rarely enter relevant answers, the underlying product information may need to be clarified, strengthened, or distributed more effectively. If removing listicles leaves the answer with few alternatives, the wider market may lack credible source material.

The value of testing source instructions is therefore diagnostic: it can reveal where the company's market evidence appears thin, crowded, or shaped by competitors. The instruction itself does not fix the gap.

Detailed observed results#

Reddit — an observable shift, every query group. When we asked Perplexity to use Reddit or user discussions, the visible citations shifted in all five query groups. One possible explanation is that relevant Reddit discussions were available while Reddit was less prominent in the baseline answers. The test observed the change in visible citations; it did not isolate why Perplexity selected them.

Official pages — rarely present on broad questions. Official pages rarely appeared among the visible citations for the broad provider-selection query. The test cannot distinguish whether relevant pages were unavailable, not retrieved, or retrieved but not selected. On some brand-versus-brand questions, official comparison pages were already present in the baseline, cited whether or not we asked for them.

Reviews — a layer shaped by vendors. "Review source" sounds independent, but the visible review layer mixed review platforms, YouTube videos, blog posts, comparison pages, and vendors reviewing each other. Of the 116 written review articles classified in this test, 50 were published by software vendors. The most frequently cited vendor-authored review article in this sample was published by tl;dv, a competing meeting-notes provider, about Fathom. We classify it as vendor-authored rather than independent based on publisher ownership; this classification does not assess the accuracy or intent of the article. It is the same incentive flattening we found reviewing all 62 sources about one brand.

Avoid listicles — a smaller citation set, not a better one. On the broad "best tools" question, where listicles dominate, the instruction reduced them — but the visible citation set became smaller rather than demonstrably better. Baseline answers cited about 13.5 sources on average; the "avoid listicles" answers averaged 9.9.

A working interpretation: "something to pull, and room to add it." Across conditions, an instruction tended to change the visible citations when that source type both existed for the query and was not already saturating the answer. This is a working interpretation of the observed citation pattern, not a proven description of Perplexity's retrieval mechanism.

How we checked this#

We tested one category — AI meeting-note tools — using five buyer questions under five collection conditions: a baseline, "use official pages," "use reviews," "use Reddit," and "avoid listicles." Each query-condition combination was repeated three times, producing 75 searches in total.

We recorded the visible citations and classified the resulting records by domain, publisher ownership, and source type. Repeated runs helped distinguish recurring changes from the normal variation observed across the baseline answers.

We reviewed 926 visible citation records across 75 searches, representing 60 cited domains. Because there was no pre-registered statistical threshold, we describe an instruction as having "produced an observable shift" or "changed the visible source mix," not as having "worked."

Explore the source data#

You can explore every cited source yourself: filter, sort, and download the full source data — 926 visible citation records across 75 searches, representing 60 cited domains, each traceable to the exact search and answer it came from, plus rollups by URL and by domain.