Appendix F — Case Studies and Teardowns

Four studies, and one warning before the first of them: these are reconstructions from the research behind this book, not interviews. No operator sat for a recorded interview, and no account below is identified or identifiable. Where a figure comes from a named first-hand report in the manuscript, it is reproduced as it appears there and cited by chapter. Where a study is a composite assembled to show a mechanism, it carries [UNVERIFIED]. Every current-state claim carries [AS OF 2026-10]. Treat the arithmetic as a demonstration of a shape, not as a forecast for your account, and replace the numbers with your own as soon as you have them.

Teardown A — the brand-safe persona

[UNVERIFIED] This study is a composite. The mechanism is drawn from the sources cited; the account is not.

The setup. A home-and-interior persona — one point of view about domestic space, no travel, no movement, no food preparation. Model A in 0.3: a persona with an aesthetic, built for brand work, with the highest ceiling and the longest runway in the book.

Why this niche. On 2.1's three-axis matrix it scores 4 for competitive headroom, 4 for buyer density, and 4 for production feasibility, a total of 12, second among the fifteen niches scored. The scores are a judgment rubric rather than measured data [UNVERIFIED], and the deciding axis is feasibility: interiors live in frames a generated persona can hold, unlike travel, fitness, or food, which fail Axis C for the reasons 2.1 gives. Hands are the most reliable tell in synthetic imagery (1.4), so the category that requires the fewest hands on camera is the category that survives QA.

The positioning. Archetype: the artisan (2.2) — someone with one specific obsession, which is what makes a narrow slice credible. Relationship: aspiration rather than identification, and that is a revenue decision rather than a taste decision. 2.2 ties aspiration to product, brand, and licensing revenue and identification to subscriptions, so a brand-first model takes the aspirational shape knowingly. The persona is designed as a channel for the brand's product, not as the product.

The format mix. Built from the cost ladder in 6.1 rather than from ambition. Roughly half the output is stills and carousels, the cheapest format and the one where identity QA is strongest. Roughly a third is stills carried by voiceover, which is text-to-speech over a validated keyframe and avoids the identity-in-motion problem entirely. The remainder is generated motion, kept small on purpose because native text-to-video with a consistent face across shots is not a solved problem [AS OF 2026-10] and shot-to-shot identity is the open problem of the field.

The production system. A forty-slot prompt file (5.2, and Appendix E, Template 2). A four-slot week (7.6): production on Monday and Wednesday at 90 minutes, a bank-building slot on Friday at 60, and a 45-minute review on Sunday. Three to four content hours a week to hold the rhythm (9.3), with a bank of approved-but-unpublished frames in reserve (5.6) as the buffer that survives an illness or a bad batch. This matches the working-account band in 9.3, and the chat queue that dominates model B is absent by design.

The rate card. The market bands in 9.5 [AS OF 2026-10]: $100 per post at the low end for micro accounts; $500–2,000 per post between 10k and 50k followers; above $2,000 per campaign for an established persona; $10,000 and above at the top of the distribution. A fifth line exists that needs no audience at all — deliverable work priced per asset, sold to a performance team rather than an influencer team. The lower band is the market's estimate of the risk of an unknown account rather than a judgment of the work.

The disclosure posture. Three placements, three jobs (11.3): an account-level bio line naming the persona as AI-generated, asset-level disclosure whenever the content is paid, photorealistic, advice-shaped, or likely to be reuploaded, and interaction-level disclosure in every new conversation. The standard applied is "unavoidable, not available," which is the stricter US advertising standard and survives being tested by someone who is not on your side. The platform label is treated as satisfying the platform, never as satisfying the law (11.2), and the FTC penalty anchor of $53,088 per violation, per-violation rather than per-campaign, is the reason the disclosure sits in the first visible lines [AS OF 2026-10].

What worked. Four things, in order of effect. The category had visible existing spend, which is Axis B and the axis most operators skip. The media kit led with the two fields a brand-safety reviewer signs off on — production capacity and compliance documentation (9.5) — and a documented pipeline beat larger accounts carrying no paperwork. The deliverable model removed platform risk from the transaction, because files the brand publishes under its own accounts cannot be interrupted by a suspension. And the persona stayed inside a category where the audience forgives a render, which is not true of food, fitness, or travel.

What capped it. The channel position requires a broad, legible persona, and breadth caps intimacy, which caps subscription revenue before it starts. Reach was working against a falling baseline: the Instagram reach figure in 10.1 fell roughly 35 percent year over year [AS OF 2026-10], and 48 percent of brand content is now AI-generated in some form [AS OF 2026-10], so the volume competing for a fixed attention budget keeps rising. Buyer density scored 4 rather than 5: the clearest open field is financial brands, where 22 percent have never used a virtual persona [UNVERIFIED], but the ones that have are concentrated among the largest advertisers and procurement is slow. The hardest ceiling is arithmetic. Deliverable work is priced cost-plus, so the ceiling is capacity (9.5), and capacity is hours.

Teardown B — the subscription persona

[UNVERIFIED] The account is a composite assembled from the two first-hand reports in 9.2 and 10.4. The figures are reproduced as the manuscript states them; the practices around them are the reconstruction.

The setup. Model B in 0.3: a persona that converts attention into a subscription on a platform built for it, where the revenue engine is private conversation rather than the subscription itself. Fastest cash flow, heaviest operational load.

The first-hand figures, reproduced as the manuscript states them. Account one, May 2026: $4,230 in the month and approximately $9,173 lifetime. Composition: roughly two-thirds from chat, with subscriptions about one-seventh of the total. [AS OF 2026-10] The same account in August 2026: best month above $8,000, lifetime approximately $25,000 [AS OF 2026-10]. Account two: $20,000 lifetime, more than 80 percent of it from chat, run by an operator who treats the subscription as a quality filter rather than as the product. These are two accounts reported by the operators themselves, not a sample (9.2).

The arithmetic, in chapter 9.2's order. In May, lifetime stood at $9,173 and a single month produced $4,230, so everything before that month summed to $4,943. Three months later, lifetime had reached about $25,000, a gain of roughly $15,800 across June, July, and August. With the best of those months above $8,000, the other two averaged a little under $4,000 each. The account was still ramping rather than plateaued, and the lifetime figure was dominated by recent months. Read only the August headline and you would conclude the account earned $8,000 every month.

The platform arithmetic, reproduced rather than replaced. Fanvue's published position: $200 million in annual recurring revenue, roughly 325,000 creators, a 20 percent commission, and identity verification for the operator behind the account [VENDOR CLAIM] [AS OF 2026-10]. The scale figure comes from the platform, which has an interest in it being large; the commission and the verification requirement are operational terms you can check in the signup flow. Two readings of the same headline are possible, and 9.2 draws out both:

Both readings put the average creator between roughly $41 and $205 a month, and neither supports a $4,000 month. The first-hand operator sits an order of magnitude above the platform average, which is not a contradiction. It is what a long tail looks like from the inside, and it is the best available confirmation that the $0–500-a-month band is not pessimistic. One further reading belongs beside it: annual recurring revenue is a platform metric, so a platform adding creators quickly can show revenue growth while median creator earnings fall.

The price structure. Four layers with different jobs (9.4): an entry subscription at 1× whose purpose is to filter intent rather than to be a bargain; a second tier at 2–3×; paid items at one to two months of the entry tier per item, batched as a catalog rather than released on request; and personalized work at 5–10× a paid item, which carries the best margin per unit and the worst scaling behavior because it consumes the operator's hours. The personalized layer is capped by slots per week — ten is a plan, "as many as come in" is an emergency with revenue attached. Item values run from about $8 for a single item to about $40 for something custom or bundled [UNVERIFIED].

The platform constraint. OnlyFans bans synthetic personas; Fanvue is the subscription platform that permits them [AS OF 2026-10]. Operator identity verification is an operational constraint rather than a formality: the payout name must match the verified name and the legal structure that receives money (9.4). Treat platform permission as a renewable agreement with a date on it, not as a property of the business.

What worked. The subscription as a filter rather than as the product, which is what account two does explicitly. A catalog built in one production day and released across a month, which costs less attention than items produced on request and shows which item converts. A slot cap on the personalized layer, set before demand existed. And disclosure in the chat surface, which is the placement most operators never build and the one a chat-first business cannot skip (11.3).

What capped it. The business is a service business with a synthetic face (9.3), and the service part scales with revenue. Two-thirds to more than four-fifths of the money arrives through conversation, so the queue grows with success. The working band in 9.3 is under ten hours a week for a hobby with a storefront and over twenty-five for a job, and the chat load is what moves an account between the two. Churn compounds the load: at 20 percent monthly churn the average subscriber lifetime is five months, and an eight-point churn improvement moves lifetime value by about two-thirds (10.4). There is no inventory that compounds while the operator sleeps, which is why 9.3 exists as a chapter.

Failure study — 2.35 million views, zero conversions

[UNVERIFIED] A composite of the failure reports in 1.5 and 10.1.

The report. 2.35 million views over ninety days, and conversions the operator describes as approximately zero. It is the most common failure report in this business, and almost everyone who tells it blames the content.

The five countable steps (10.1).

1. Reach achieved. 2.35 million views in ninety days. The algorithm liked something — format, face, or hook. This link worked.
2. No routing. No link-in-bio destination that resolved, no story sticker sequence, no comment-to-DM trigger, no named destination in the caption. A viewer who wanted to go further had no next step that produced a measurable event.
3. No landing. No owned page carrying the click, the terms, or the disclosure, so nothing was recorded and the operator could not see the drop even after it happened.
4. No capture. No handle, no email, no consent, no channel membership. Ninety days of attention produced zero addresses that could be contacted again.
5. Therefore no revenue. Not a low conversion rate. No conversion surface at all.

The diagnosis. The content was not the failure. The funnel was. Content that produced 2.35 million impressions did its job; reach arrived and had nowhere to go. The instructive step is the third one, because reach without instrumentation does not fail loudly. It fails silently, and the operator concludes the niche is dead. 1.5 calls this the single most expensive failure in the book, because it wastes the hardest thing to obtain, which is attention.

The warning sign. You cannot state, in one sentence, what a viewer is supposed to do next after seeing your third post (1.5).

The cost, in the currency that matters. The measurement that would have caught it is nearly free: one landing page with one measurable action on it, one tracked link per channel, and an intermediate channel that survives a platform change (10.1). Neither of the two fixes costs money. The chain that produced nothing was views, then nothing — no clicks, no sessions, no captures, no paying users — and the operator's dashboard could not tell the difference between a broken funnel and a dead niche.

The general rule. Build the destination before the audience. A revenue figure without its denominator is a headline, and a view count without a measurable next step is a vanity metric wearing the same clothes.

Migration study — a video provider discontinued

[AS OF 2026-10] All current-state claims in this study are dated to the manuscript's reference month.

The event. Sora 2 was discontinued. The consumer application closed on 26 April 2026 and the API sunset followed on 24 September 2026 (0.2). Operators who had built a production pipeline and a content calendar on that API lost both, on a date announced by somebody else. There was no migration path, because the tool was not a category — it was a product.

Who survived it, and why. 12.4 states the four practices that decided the outcome, and none of them was decided after the announcement.

Still-first workflow. In the composition that works (6.1), identity is solved at still-image quality and the video model's job is movement. Validated keyframes are portable across every video model; a library of vendor-specific generated clips is not. Operators in the still-first group kept their assets and re-rendered motion. Operators holding only clips kept a library they could no longer extend.

Two known paths, not two subscriptions. The rule in 7.5 is working knowledge of a second provider, demonstrated with test clips generated at production settings. Migration is a week of testing when the second provider is already familiar and a month of lost output when it is not.

Provider-neutral prompts and references. A prompt written for one model's quirks needs rewriting. A prompt written as a scene description (6.2, and the filled variant in Appendix E) does not.

Assets stored locally. The library, adapters, and masters live on your own storage with three copies and a tested restore (7.4), not inside a vendor's project space.

The identity question, which is the larger one. A persona built on your own trained adapter is yours. A persona built inside a vendor's character feature is rented (12.4), and the export from a character feature is partial at best. That distinction decides whether a vendor sunset costs you a production week or a business.

The five-minute test, run now. If your primary video provider disappeared tomorrow, could you produce a post this week? If the answer is no, the exposure is not theoretical, and the fix is two test clips on the second provider (12.4).

What the study does not prove. It does not show that any particular vendor is at risk of the same outcome, and it does not predict the next sunset. It shows that the difference between losing a model and losing a business was decided months earlier, by four practices that cost almost nothing to maintain.

What the four studies have in common

The decisive move happened before the incident. The destination existed before the audience arrived, or it did not. The second provider was familiar before the sunset, or it was not. The compliance register was written at creation, or it was reconstructed. The slot cap was set before demand, or the queue set it. In all four, the moment that mattered was earlier than the moment that hurt, and it was cheap.

Ownership is the recurring variable. Keyframes, prompts, adapters, records, and an audience you can reach off-platform are assets. Vendor clip libraries, rented character features, platform permission, and a follower count without an address list are exposures. The line between them is not technical; it is a decision about where the work lives.

Two studies are about a missing destination and two about a missing second path. Teardowns A and the failure study both turn on whether a viewer had somewhere to go and whether the operator could see it. The subscription study and the migration study both turn on what happens when the single path closes — a platform rule, a vendor sunset, or an operator who cannot sustain the hours.

The figures are checkable, and yours will differ. One study is arithmetic on published first-hand numbers; three are composites marked [UNVERIFIED]. Read them for the shape of the decision, not for the number. Then run Appendix E's registers and Appendix D's calculators against your own account, and replace every figure here with one you measured.