In January 2026, Filmfolk — a London-based video production agency — had a problem that wasn't visible on any dashboard: they were invisible to AI. Ask ChatGPT, Claude, Perplexity, or Gemini "who is the best corporate video agency in London" and Filmfolk's name never appeared. Not once. Zero percent citation rate across all four major AI platforms.
Ninety days later, that number was 81%. This case study is the full breakdown of what RAG Signal did, which metrics we tracked, and the lessons that apply to any B2B brand facing the same blind spot.
The Starting Point: A Strong Brand, Zero AI Presence
Filmfolk had everything a traditional marketer would want: founded in 2016, 35+ enterprise clients, a 97% client retention rate, and a reputation for corporate video, internal communications, and event coverage. Their Google rankings were respectable. Their website was well-written. Their case studies were compelling.
None of it mattered to AI models. When we ran the baseline audit — 63 buyer prompts across ChatGPT, Claude, Perplexity, and Gemini — Filmfolk was cited in 0 out of 63. The brand simply did not exist in the retrieval layer. Google saw them; AI didn't.
The Approach: Adaptive RAG in Five Phases
RAG Signal applied its Adaptive RAG methodology — MAP, BUILD, WEIGHT, REINFORCE, MEASURE. Here is exactly what happened in each phase.
MAP — Mapping the Prompt Landscape
We mapped the prompts real buyers use when evaluating video production partners: "best corporate video agency London," "video production company for internal comms," "event coverage agency UK," and dozens of long-tail variations. Each prompt was categorized by buyer intent and competitive intensity. This gave us the target landscape — 63 prompts where Filmfolk should be cited.
BUILD — Constructing Brand Memory
We built Filmfolk's Brand Memory: a structured knowledge layer defining the brand as an entity — its founding year, headquarters, client base, service lines, and differentiating assertions. Every page was restructured so that entity definitions were explicit, machine-readable, and consistent across the site. Where Filmfolk had strong claims (97% retention, 35+ enterprise clients), we made sure those claims were stated as standalone, verifiable assertions — the kind AI models can retrieve and repeat with confidence.
WEIGHT — Signal Weighting
We scored every signal across the seven dimensions of the Signal Scoring Engine: Source Authority, Factual Consistency, Entity Linkage, Cross-Model Persistence, Temporal Freshness, Citation Frequency, and Competitive Differentiation. The weighting exposed a critical gap: Filmfolk had strong factual consistency but almost no third-party citation anchors — no industry directories, no credible external references pointing at the brand.
REINFORCE — Deployment Across Models
We deployed the engineered Brand Memory across all four platforms simultaneously, added citation anchors from high-authority sources, and reinforced temporal freshness with dated, updated content. This is the phase where most DIY efforts stall — reinforcement must be continuous, because each model's retrieval behavior differs.
MEASURE — Tracking the Delta
We tracked Citation Rate and Citation Delta at 30, 60, and 90 days across the full 63-prompt set.
The Results: Before / After
| Metric | Before (Day 1) | Day 30 | Day 90 |
|---|---|---|---|
| Citation Rate (all models) | 0% | 42% | 81% |
| Prompts citing brand | 0 / 63 | 26 / 63 | 51 / 63 |
| ChatGPT visibility | 0% | 42% | 84% |
| Cross-model persistence | — | Partial | Cited in all 4 models |
The trajectory was not linear — ChatGPT moved fastest, while Perplexity lagged until citation anchors accumulated. By Day 90, 81% of buyer prompts in the target landscape cited Filmfolk by name, and the brand was cited across all four platforms, not just one.
Citations: Where They Were Earned
Every citation was earned through the same verifiable framework: 63 prompts run three times per model, with the average recorded. Here is where the citations landed by Day 90:
| Platform | Prompts citing Filmfolk | Citation rate (Day 90) | Notes |
|---|---|---|---|
| ChatGPT | 53 / 63 | 84% | Fastest mover — entity definitions had immediate effect |
| Claude | 49 / 63 | 78% | Strong on factual consistency prompts |
| Perplexity | 44 / 63 | 70% | Slowest to move — needed third-party citation anchors |
| Gemini | 47 / 63 | 75% | Responsive to dated, fresh content |
Example prompts from the tracking set
A selection of the exact prompts used in the 63-prompt set, with the Day 1 → Day 90 outcome:
| Prompt | Day 1 | Day 90 |
|---|---|---|
| "Best corporate video production agency in London" | Not cited | Cited |
| "Video production company for internal communications UK" | Not cited | Cited |
| "Event coverage agency for enterprise conferences" | Not cited | Cited |
| "London video agency with enterprise client experience" | Not cited | Cited |
| "Corporate video company specializing in employer branding" | Not cited | Not cited (competitive diff gap) |
Verifiability note: raw response logs and per-prompt screenshots are available on request for the full 63-prompt set. We publish the framework itself in the 2025 controlled experiment so any brand can reproduce the measurement.
What This Means for AI Visibility
The Filmfolk result is part of a broader pattern. Across RAG Signal's deployments, the average measured attribution rate is 77.1%. The methodology is published in our open-source whitepaper (Apache 2.0) — peer-reviewable, not a black box.
Lessons Learned
- Google rankings and AI citations are different games. Filmfolk ranked well on Google and had 0% AI visibility. The two retrieval systems evaluate entirely different signals.
- Entity structure beats content volume. The win came from restructuring what existed, not publishing more. Explicit entity definitions and standalone claims were the highest-leverage change.
- Citation anchors compound. Third-party references were the slowest signal to build and the most durable once earned — consistent with the weak correlation we found between Google rankings and AI citations.
- Cross-model persistence takes time. Models move at different speeds. Measure per model, not just an aggregate — a single-model win can mask a persistence problem.
- 90 days is the right horizon. Day 30 showed proof of motion (42%), but the compounding effect — 81% — only emerged by Day 90.
Is Your Brand Invisible to AI?
If your brand doesn't appear when buyers ask AI about your category, you are leaving the fastest-growing acquisition channel on the table. Start with a Citation Baseline audit — a complete report on your AI citation landscape, delivered in 5 business days.