RAG Signal
Back to Insights
Research August 5, 2026 Bora Kurum

Prompt Discovery #1: The Real AI Prompts That Cite Filmfolk — and Why

First edition of the Prompt Discovery series: five real buyer prompts from the Filmfolk tracking set, analyzed signal by signal — what triggered retrieval, what won the citation, and what to replicate.

Prompt DiscoveryCase StudyAI CitationsRetrieval

This is the first edition of the Prompt Discovery series — real AI prompts from our tracking sets, analyzed signal by signal: what triggered retrieval, what won the citation, and what you can replicate. All data comes from the same 63-prompt framework we publish openly.

In the Filmfolk engagement, we tracked 63 buyer prompts across ChatGPT, Claude, Perplexity, and Gemini — from 0% to 81% citation rate in 90 days. This edition dissects five of those prompts at the moment retrieval fired, and explains which signal won each citation.

Prompt 1: "Best corporate video production agency in London"

Platform: ChatGPT · Day 1: not cited · Day 90: cited

What won it: Entity Linkage (15%) + Competitive Diff (6%). The restructured homepage stated the category ("corporate video production"), the geography ("London"), and the proof ("35+ enterprise clients, 97% retention") in the first two sentences — as standalone, retrievable claims. Before the build, the same facts were scattered across three pages with no explicit category-geography pairing. The model had nothing to attach "London video agency" to.

Signal breakdown: the winning chunk satisfied two requirements at once: it answered the query's category constraint and its geography constraint. Chunks that answered only one of the two never made the citation.

Prompt 2: "Video production company for internal communications UK"

Platform: Claude · Day 1: not cited · Day 90: cited

What won it: Factual Consistency (20%). Filmfolk's internal-comms service line was stated identically across the homepage, a dedicated service block, and the case study page — three independent surfaces agreeing on the same claim. Claude's retrieval weighted the corroboration heavily. Before the build, the internal-comms service existed only as one line in an about-page paragraph.

Signal breakdown: this is the corroboration effect: the same assertion in multiple retrievable locations raises factual-consistency scores. One location = a claim. Three locations = a fact.

Prompt 3: "Event coverage agency for enterprise conferences"

Platform: Gemini · Day 1: not cited · Day 90: cited

What won it: Temporal Freshness (12%). Filmfolk had published a dated case study from a recent enterprise event coverage project. Gemini — the most freshness-sensitive of the four models in our data — retrieved that dated chunk over older, undated page content. The fresh chunk carried a date stamp, the older content carried none.

Signal breakdown: temporal queries ("for enterprise conferences" implies current capability) reward date-stamped evidence. Undated claims score near-zero on freshness regardless of quality.

Prompt 4: "London video agency with enterprise client experience"

Platform: Perplexity · Day 1: not cited · Day 90: cited

What won it: Citation Frequency (10%) + Source Authority (25%) — the slowest signal to build. Perplexity cited Filmfolk only after third-party citation anchors accumulated (industry directory entries and a client-facing reference page). This was the last model to move, and it moved on anchors, not content.

Signal breakdown: Perplexity weights external corroboration harder than the other models in our set. If your brand is invisible on Perplexity but visible elsewhere, check your third-party references before rewriting content.

Prompt 5: "Corporate video company specializing in employer branding"

Platform: all · Day 1: not cited · Day 90: still not cited

Why it never fired: a genuine competitive-diff gap (6%). Filmfolk's employer-branding work was real but never stated explicitly as a standalone service line. Competitors who named the specialty explicitly won the citation every time. The content existed; the retrievable claim did not.

Signal breakdown: this is the prompt that didn't convert — and the most instructive one. The fix was straightforward (add an explicit employer-branding service definition), which is why Competitive Diff carries weight: it is the dimension most brands ignore and easiest to win.

What This Series Is For

Every prompt above is a data point, not a story. We log the prompt, the platform, the outcome, and the signal that decided it. Over time, the series becomes a map of how retrieval actually behaves — which models weight what, which signals compound, and where the cheapest wins are.

Series: Prompt Discovery · Next edition: a B2B SaaS prompt set.

Ready to get your brand cited?

Start with a baseline AI Visibility Audit. Know where you stand across every major AI model.

Get Your Citation Baseline →