Funnel in. Refine. Funnel out.

Why any of this is necessary

Advertising has always been built for one reader. A brochure, a billboard, a landing page — every

craft decision downstream assumes a person is deciding, so the work optimises for emotion, aspiration,

and the image doing what the copy cannot.

That assumption is now incomplete rather than wrong. A second reader sits between the brand and the

buyer. It does not respond to aspiration. It parses, retrieves, weighs sources, and produces a

recommendation. The human still decides — increasingly from a shortlist the machine composed.

The working position: content should carry human appeal and inform machine decision-making, in the

same artifact. These are not competing briefs to be traded off. The machine layer — structured data, a

parseable answer, a transcript — is a property of a well-made asset, not a separate genre of dry

content sitting beside the good one.

The model in one line

Funnel in. Every major AI answer engine is asked the questions your buyers actually ask, and tells

us what it recommends today and who it cites instead of you.

Refine. The engine measures, diagnoses, prescribes and routes — turning raw answers into a specific

brief with the content, format, medium and warrant already decided.

Funnel out. That brief drives the creation tools that build the asset, chosen by the agency. Then

the same frozen questions are asked again, which is the only step that can tell you the work caused the

change.

Stage 01 — Measure

We ask a frozen question set: the same buyer questions, in the same words, every cycle.

This sounds like a detail and it is the whole basis of honesty in the method. If prompts are

regenerated per run, every comparison is a re-roll and every before-and-after is a story. Freezing the

set at a subject's first measurement is what makes a later claim of movement checkable.

Every answer and every citation is recorded — not just whether you were mentioned, but which sources

answered in your place. That second half is what makes the diagnosis actionable.

Stage 02 — Diagnose

A gap is not "improve your content". It is a specific, named absence: the page that does not exist,

the question that goes unanswered, the claim a competitor makes that you do not.

Each gap carries the engines that surfaced it, how many agreed, an impact estimate, an effort estimate,

and a priority derived from them. Cross-engine agreement matters more than any single reading, because

engines are stochastic — one engine saying something once is a watchlist item, not a commissioning

brief.

Stage 03 — Prescribe

The gap is routed to a format — the informational structure that answers it. The catalogue holds

36 of these across eight families:

• ●Informational — article, FAQ, how-to, buying guide, ultimate guide, glossary definition,

checklist, spec sheet, coverage roadmap, intent gap analysis

• ●Comparative — comparison, listicle, alternatives page, vs matrix
• ●Proof — customer proof, certifications page, case study, test results
• ●Corrective — claim clarification, misconception correction, myth vs fact, review response
• ●Earned — press release, media pitch, contributed article, creator brief
• ●Transactional — page enhancement, pricing transparency, where to buy, product release
• ●Social — social post, newsletter section, YouTube description, video script
• ●Technical — schema meta set, markdown variant

Each format declares the components it requires and a governance mode. Five are request-only,

because they assert evidence we cannot invent — a case study, a test result, and all four corrective

formats need a human to ask for them.

Stage 04 — Route

The brief is matched to a medium, and the providers who could build it are ranked.

Ranked, not chosen. Output provider is always the agency's decision — there are several ways to

produce anything, and an agency usually has preferences, existing subscriptions, or a seat it already

pays for. We hand over a brief and a shortlist; we do not pick the vendor.

Where a provider supports acting on a user's behalf, the work can run inside the agency's own account,

on its own credits, under its own brand and themes. That is an integration, not a resale.

Format is not medium

Every one of the 36 formats is assembled from just four components:

Three of the four are data structures, not prose. An FAQ block is a list of question-and-answer

pairs; nothing about it requires a paragraph. That is the mechanical reason a format can change medium.

The rule this yields: a format may be expressed in an alternate medium if, and only if, its

non-markup components carry the meaning — and the markup must still be published as text somewhere the

engines can fetch, whatever medium the human-facing version takes. A video FAQ does not replace

FAQPage schema. It sits beside it.

Low-end to high-end — the same gap, expressed five ways

Low-end is the cheapest artifact that honestly closes the gap. High-end is the richest expression of

the identical structure. They differ in production cost and human appeal, not in what the engine is

asked to find.

The invariant: the high-end version never removes the text layer. It adds a human-facing expression

on top of it. Skipping the text to "just make a video" makes the asset invisible to the engines that

do not surface video.

The engines do not read the same things

Text-first engines lean on community, encyclopaedic and institutional long-form. Multimedia-leaning

engines lean heavily on video. Commerce-leaning engines lean on major retail listings.

In our own measurement, 95% of video citations resolve to a single platform — so if video is worth

making, where it is hosted is not an open question.

The operating consequence: render text and you address one part of the field; render video and you

address another. One brief rendered twice addresses it whole, at roughly the cost of the second

rendering, because the questions, answers and markup were settled once.

Engine-preference characterisations above are externally reported and method-undisclosed. We cite

them as the third party's claim, never as our measurement. The 95% figure is ours.

Where answers are actually sourced

Classifying every citation we have measured by where the source sits:

Content published to a brand's own domain is aimed at roughly one citation in twenty. Video alone is

a larger citation surface than the subject's entire website. That does not make owned content

worthless — it is how a brand enters the 5.1% at all, and it carries the structured data everything

else references. It makes owned content insufficient, and it makes a programme that produces only

blog posts a programme working the smallest surface on the board.

The three largest categories are surfaces a brand does not own and cannot publish to directly.

Reaching them is different work: contributed articles, media pitches, creator briefs, retailer spec

and availability data. Those are formats, and they can be commissioned — but not by writing another

post.

Directional — measured across the categories we have run, not a published statistic.

What actually crosses to a creation tool

The step that must never be delegated is the content itself. A tool asked to "make an FAQ video about

cold-pressed juice" invents the questions, and that is generic content with no measured basis.

So the engine generates the actual questions and the actual answers — each sourced, or flagged for

brand confirmation — and only a rendering request leaves the building. It carries:

• ●the warrant: which engines agreed, how many, and the question that produced the gap
• ●the content: answer summary, the Q&A block, the JSON-LD, verbatim
• ●the rendering: medium, target surface, and the machine requirements for that medium — transcript

and captions required for video, schema types for a page

• ●the branding: resolved to the agency's brand, not ours
• ●the gate: never auto-publish; output is a draft for human review

The four gates every asset passes

• ●Evidence gate — five formats are request-only because they assert things we cannot verify. Any

unverifiable claim ships flagged for brand confirmation and blocks production until the client

confirms in writing.

• ●Never-publish gate — rehearsal and self-reporting subjects are excluded by standing rule. Their

data is never a finding.

• ●Provenance gate — every number carries its denominator and its documented limits. A metric our

own registry describes as uninterpretable from a single reading never appears as one.

• ●Review gate — a creation tool returns a draft for a human. It never returns a published asset.

What this does not yet establish

Stated plainly, because the sections above argue for a way of working and these are the reasons it is

not yet a proven one.

• ●No closed loop has been run. The re-measurement log is empty. Nothing here demonstrates that a

published asset moved a citation; it establishes that the mechanism to find out is built and has not

been fired.

• ●The distribution is ours, not the market's. Every figure comes from the subject mix we have run.

Directionally useful for planning; not a claim about any particular category.

• ●Execution is specified, not delivered. The outbound adapters do not exist yet. Drafting a brief

and producing a finished video are separated by an integration still to be built.

• ●The transcript question is open. Whether engines consume a video's transcript, its captions, or

only its metadata is unresolved — and the case for video as a citation surface rests on the text

layer being read.

What would change all of this: one subject, one gap, one asset produced and published, then the frozen

question set asked again. A single closed loop converts most of the above from architecture into

evidence.