Appendix G — Sources and Methodology

This appendix says where the material came from, how sources were ranked, what deliberately is not here, and how to re-check the parts that expire. It is a method rather than a bibliography, because a link that resolves today and 404s in eighteen months is worse than a named category, and because most of the numbers in this book are checkable in five minutes by anyone willing to open the vendor's own page.

The source hierarchy

Six tiers, ranked. The working rule: when a claim can be sourced at a higher tier, use the higher tier; when only a lower tier exists, say so at the point of use with a marker.

Tier 1 — primary and first-hand. Operator dashboards and figures reported by the person who ran the account, read alongside the interest of the person reporting them. Platform documentation for terms you can verify in the signup flow: commission, identity verification, payout schedule, permitted content. Vendor pricing pages, read on the day you read them. Statute and regulator text. The strongest evidence for what one account earned is one operator's dashboard; the weakest possible evidence for what accounts earn is the same document (9.2).

Tier 2 — reference documentation and standards. Model cards, license terms, training papers, adapter documentation, and the transition timetables in the EU's AI rules. Stable, checkable, and dull, which is why it ranks above anything a market report says about the same subject.

Tier 3 — published measurement with a stated method. Spend analyses, surveys, and platform reports that describe a sample, a period, and a method. This tier is where the book's most-cited and least-trustworthy figures live, because a large share of it sits adjacent to companies selling avatar tooling. The audience-perception numbers in 1.4 — engagement of 5.67 percent against 1.89, trust of 42 against 73 — are tier 3 and carry [UNVERIFIED] for that reason.

Tier 4 — reporting. Useful for establishing that an event happened and when: the Sora 2 consumer application closing on 26 April 2026 and the API sunset on 24 September 2026 are events of this kind (0.2). Where a primary source exists for the same fact, the primary source wins.

Tier 5 — aggregated community knowledge. Forums, chat communities, and video walkthroughs. Their unique value is failure modes the vendors do not publish: bans, moderation sequences, refund disputes, model regressions. Their defects are survivorship and incentive, since the person describing a $300,000 quarter is frequently selling the method that produced it (9.2). A community number enters this book only with a marker, and never as the basis for a cost.

Tier 6 — vendor marketing. Not a source of fact but a source of claims, and a legitimate source of direction about what the vendor is building. Where a capability is described only by the company that sells it, the claim is marked [VENDOR CLAIM] at the point of use. Fanvue's scale figures are tier 6; its commission and verification requirements are tier 1, because you can verify them yourself.

One more rule sits above the tiers. A price belongs to tier 1 only if you read it on the vendor's own page today. Every other version of a price is a memory or an advertisement.

What was consulted, by type

Named by type rather than by link, because the links rot and the categories do not:

• Platform documentation and creator terms for revenue share, identity verification, permitted and prohibited content, and enforcement practice (8.1, 9.4, 11.2).
• Regulator and legal material for advertising disclosure, the AI transparency timetable, image rights, and data retention (11.3, 11.4, 11.6).
• Vendor pricing and model documentation across image, video, voice, and compute (5.2, 6.2, 6.5, 7.5).
• Model and training documentation on identity conditioning, adapters, and dataset preparation (4.3, 4.4, 4.5).
• Published measurement on marketing spend and audience perception, all rendered as [UNVERIFIED] where the report sits beside a vendor (0.3, 1.1, 1.4, 9.3).
• Operator-reported figures posted with a dashboard, a payout statement, or a month-by-month breakdown (9.2, 10.4).
• Practitioner communities and video walkthroughs, read for the failure modes and not for the numbers (8.5, 12.4).
• Price and policy pages, read once, dated, and marked [AS OF 2026-10].

Three things were excluded on principle: unnamed "studies show" with no sample, any revenue number lacking a denominator or a window, and material presented as leaked or confidential.

The honesty markers, counted

Five markers carry the uncertainty, defined in How to Read This Book and used at the point of claim. Counted across the manuscript as delivered — 85 chapter and front-matter files, excluding the appendices and the generated contents page:

Seventy-two of those 85 files carry at least one marker. Seven occurrences define a marker rather than apply it — four in the front-matter table in How to Read This Book, and three in 1.1's explanation of the system, where the date marker is described in prose rather than written as a bracketed occurrence — which leaves 195 live uses and these corrected figures: 100 [AS OF 2026-10], 80 [UNVERIFIED], 9 [UNTESTED], 6 [VENDOR CLAIM], and zero [SNIPPET]. The last one is defined and never applied, because every figure that survived into the book was read on the page rather than in a search result, and a marker with no uses should be reported as unused rather than quietly dropped.

Concentration matters more than the total, because it maps what expires:

• [AS OF 2026-10] clusters in 9.2 numbers you can trust (7), 8.1 platform selection (5), 1.1 the 2026 market (5), 9.4 subscriptions and chat (3), 3.1 the model landscape (4), and 0.2 what changed in 2026 (4). Prices, platform policy, model availability, and market measurement: the four things that move fastest.
• [UNVERIFIED] clusters in 3.3 hardware and VRAM (6), 4.5 LoRA training (4), and several chapters at 3. Hardware throughput and training results are the least independently confirmable claims in the book, which is why they carry the most.
• A chapter with no markers is not more trustworthy than one with many. It is about something that does not expire, such as a definition, a checklist, or a works-every-time procedure.

What is not in this book

• No independent audit of platform figures. The subscription platform's revenue and creator counts are published by the platform and audited by nobody. Both readings of that number are worked through in 9.2 rather than one being asserted.
• No retention or earnings benchmark. The first-hand figures are two accounts reported by their operators. Two accounts are a lead, not a sample, and 9.2 says so before it uses them.
• No tax, corporate, or immigration advice, and no localization of either: those are jurisdiction-specific and belong to a professional who can see your facts.
• No income projection. Every table in this book that converts a rate into a monthly figure is arithmetic on assumptions you supply. Appendix D supplies the calculators; it does not supply the outcome.
• No model benchmark of our own. No controlled comparison of image, video, or voice models was run, so no ranking is offered with numbers attached. Where a model is named, it stands as an example of a category, not as a measured winner.
• No endorsements, affiliate links, or referral codes. The publisher sells books and a workshop, which is a commercial interest in this subject matter, and 9.7 is where that interest becomes visible.
• No interviews. The four studies in Appendix F are reconstructions from the research behind this book, and the composites are marked.
• No identifiable operators. Figures reported first-hand are reproduced without names, handles, or details that would identify the account.

How to verify a price yourself

The procedure takes five minutes and it is the only one that produces a number you can act on:

1. Open the vendor's own pricing page. Not a listicle, not a video description, not a screenshot inside a post. The page you can buy from.
2. Read the meter, not the headline. Per image, per second of generated video, per thousand characters, or per compute-hour. Then check whether a credit converts one-to-one into the unit, because a plan priced in credits is priced in a currency the vendor controls.
3. Find the terms that move the effective price: commercial-use rights, watermark removal, output resolution, storage, concurrency limits, and whether the commitment is monthly or annual.
4. Record the date, the tier, and the meter in the quarter's audit row (12.5). A price without a date is a rumor with a number attached.
5. Re-check before any decision that costs more than an hour to reverse, and re-check every quarter as part of the audit.

Two habits make the result honest. Your cost is not the list price but cost per usable asset — the meter times the failure rate of your own pipeline (7.5, 5.6). And a number spoken in a video is stale by construction, because production timelines mean the video was recorded weeks before it was published, which is the whole reason the community tier sits fifth.

Reference date and legal limits

Everything in this book that can move is dated [AS OF 2026-10]; the reference month is October 2026. Prices, model availability, platform permission, and regulator timetables after that date are yours to check, and the procedure above is how to check them.

None of it is legal, tax, or financial advice, and the disclosure, rights, and data material in Part XI is a checklist written to be handed to a qualified professional in your jurisdiction rather than a document to rely on as written. The EU's AI transparency duties phase in through a transition running to 2 December 2026, so a reading taken now and a reading taken after that date can differ on the same facts.

Two commitments govern corrections. Where a figure is wrong because it was read wrong, it gets corrected. Where it is wrong because the world moved, it gets a date. Anything less would be a book pretending that nothing in this business expires.