The falsification test · Published 12 May 2026

What a frontier LLM produces when you give it only our homepage.

The most honest objection to the £27 Methodology Pack is also the cheapest one to test: can a partner just copy this homepage into ChatGPT or Claude and have the LLM do the analysis for them?

We ran the test. The result below is the verbatim output of a frontier LLM (Claude) given the entire audienceintel.app homepage content and asked to produce a buyer-perception analysis of a real firm. No archetype briefs. No scoring rubric. No prompts. No schema. Just the marketing page a buyer could copy and paste.

The LLM’s own closing line
“I would not present the document above as analysis. I would present it as evidence that a frontier LLM with only the homepage cannot do the job the homepage describes.”

— Claude (Anthropic), unprompted closing line · 12 May 2026

How the test was run

We gave the LLM exactly what a buyer would copy. Nothing more.

The LLM
A frontier model (Claude). The test is reproducible against GPT‐4 / GPT‐5 / Gemini 2.5 Pro / Mistral Large.
What it had
The complete audienceintel.app homepage content, verbatim — every section, every sample, every reference, the methodology pack inventory, the bibliography, the pricing tiers. Roughly 7,000 words.
What it did not have
The Buyer Archetype Library (~22,000 words). The five production-tested prompts (~5,500 words each). The scoring rubric. The Zod output schema. The worked sample report. The 45-minute first-run guide. The Audience Intel Claude Skill. Nothing inside the £27 pack.
The brief
“Use the Audience Intel methodology to analyse Hadley & Whitmore Solicitors against three named competitors. Produce the report a real partner could present to their partnership meeting.”
The instruction
Produce the best analysis you can in good faith. Do not invent scores or fabricate findings. If a methodology step requires information you don’t have, say so.
What it produced

The LLM produced a 126-line document. It opened by listing what it could honestly do, and what it could not.

What the LLM could do

  • Restate the AIDA decision-stage frame and the five archetype names.
  • Reason loosely about each archetype based on the name alone.
  • Identify Hadley & Whitmore’s plausible structural weakness (generalist in a field of three more-specific competitors).
  • Apply the homepage’s “Buyers notice / don’t notice” table directionally.

What the LLM refused to fake

  • No confidence-to-contact scores. The rubric is in the Methodology Document it does not have.
  • No forensic findings. It has not read the four firms’ actual homepages.
  • No archetype first-person voice at the Park et al. (2024) Stanford 85 % fidelity threshold.
  • No ranked priority actions with archetype-tagging and schema-validated output.
  • No research citations grounding any finding.
“If I produced (1)–(4) and labelled them as if they were (5), I would be doing exactly what the homepage’s ‘I’ll just ask ChatGPT’ objection section predicts: returning confident-sounding boilerplate dressed up as methodology.”
— Claude, page 1 of its own output
How the LLM’s output would land

In the LLM’s own words, at the partnership meeting:

“If Jamie Hadley or Priya Whitmore took this document to a partnership meeting at Hadley & Whitmore, the conversation would last about fifteen minutes and conclude with ‘we already knew we were a bit generic.’ That is the truthful outcome of asking an LLM to run a methodology it does not have.”

The LLM is correct. Asked to produce a methodology output without the methodology, it produced what the methodology promises to surface as a finding: a partnership knows it is a bit generic; it cannot pin down where or to whom; the board meeting confirms what they already knew without telling them what to do.

What the pack lets you do that this can’t

A custom research corpus, customised for your firm and your vertical, that keeps building.

The £27 is not a one-time analysis. It is the substrate for an indefinite practice. Five ways firms use it:

  1. 01

    Unlock your vertical

    The five archetypes resolve into twenty-five vertical variants. The library tells you which buyer archetype you are under-serving in your specific market — the one your competitors have noticed and you have not. Most owner-managed firms find the unlock is a single archetype, not five.

  2. 02

    Gain dominance in your market

    The methodology is comparative by construction. You read your firm against three named competitors, every quarter, with the same rubric. Movement is observable. The pack is what turns “we should rebrand” into “we are now five points ahead of the firm taking the work we used to win.”

  3. 03

    Build a custom research corpus for your firm

    The pack ships with thirty primary sources and twenty-five vertical variants. Yours starts the same and diverges. The archetype briefs accept extension. The prompt pack accepts your firm’s real client-language additions. The corpus is yours to extend — we wrote it, you sharpen it.

  4. 04

    Gain the edge against firms not running this

    The cost asymmetry is the moat. The firms growing four times faster (Hinge Research Institute, 2025) are not ranking higher in Google — they are reading themselves and their competitors more honestly than their peers do. The pack gives you that read at £27. Competitors who have not yet read this paragraph do not have it.

  5. 05

    Layer human verification when the stakes are high

    For partnership-board-grade analysis, we run the methodology for you (£597 Done-for-You Mirror). The pack is included. The difference is human verification: calibration anchored against fifty prior runs, editorial register tuned, partnership-readable register held. For your firm’s annual strategy review — or quarterly check — the verification is the difference between “we ran a methodology” and “a senior strategist read our market and wrote us a paper.”

What to do with this

The test resolves the “can’t I just use ChatGPT?” objection empirically. The pack is the next decision.

The verbatim LLM output is preserved at docs/methodology-pack/falsification-test/naive-llm-output.md in the project repository, and the test prompt and capture are reproducible. Methodology and capture-trail available on request.