Appearance
Best practices — the evidence behind the recommendations
What this page is. Before you mark the review surfaces, this is the industry evidence in plain language — what the taxonomy world, the big platforms, and the standards bodies actually do, and what that means for each decision in front of you. Two research passes feed it: an eighteen-system pass on how content systems separate type, category, format, placement, and campaign (16 August), and a second pass on review process, identity and retail vocabularies, and documentation governance (17 August). Every claim in both passes was gathered live from a quoted source; the second pass additionally had an adversarial fact-checker re-fetch its most load-bearing claims — all of them held verbatim. This page carries no marks; it exists so the marks you make elsewhere are made with the evidence in view.
The seven rules the taxonomy world agrees on
These came out of the first pass, and they are remarkably consistent across standards bodies, asset platforms, ad platforms, and catalogs:
| # | The rule | Who proves it | For us, concretely |
|---|---|---|---|
| 1 | Only subject matter earns a deep hierarchy — type, format, purpose, audience each stay their own flat axis | the ad industry's taxonomy standard rebuilt itself in version 3 to fix exactly this | the type list stays one flat vocabulary; shelves are browse groupings above it, never a second stored fact |
| 2 | Belonging to a campaign is a relationship, never a category value | every asset platform files campaigns as collections; every ad platform puts the objective on the campaign object, not the creative | a piece is a member of a campaign; no piece stores a campaign-flavored type |
| 3 | Paid versus organic is a distribution state, not a kind of piece | the same social post runs as a paid ad and stays organic at once — nothing on the post changes | this is why Performance Ad came off the list; ownership and placement carry paid-ness |
| 4 | Type layers are small everywhere | the library world's core type list is 12 values; asset platforms ship 4–7 buckets; nobody ships a ten-bucket mixed layer | the ten MIXED families were the outlier the industry warns about — mixed-kind, not merely ten |
| 5 | A hierarchy is only valid between things of the same kind | the controlled-vocabulary standard states it outright — a shelf holding "a job" next to "a format" next to "a paid-ness" is malformed | the kind test on every candidate row: is this a recognizable kind of piece? |
| 6 | There is no magic number of top-level groups | the seven-plus-or-minus-two rule is a debunked misreading; observed real-world medians sit near seven with a range of 3 to 31 | the eleven shelves sit inside that observed range, and shelf count is a browse choice, not a schema constraint — exactly what the surface says |
| 7 | Compound labels people love should be computed views, never storage rows | a streaming service's seventy-six thousand micro-genres are generated queries over clean facets | the five worlds as computed lenses is the industry's strongest pattern, not a compromise |
What the evidence says about each thing you are marking
The mechanism (proposal 1 — taken provisionally). Rules 1, 2, and 7 together are the recommendation: type stored on the piece, family as a registry shelf, worlds as computed lenses. The three alternatives the research graded — the worlds stored per piece, the deliverable families stored, today's ten kept — each fail at least one of the seven; the fourth preview (the work-domain carve) was not graded by the research but fails on its own measured numbers, with most of the type set landing in a single bucket.
The Film naming picks. The label doctrine says a term is a noun phrase that names the KIND — which every name on the list is. The medium-word rule — a medium word is honest only where the kind is inherently that medium — is the review's own pressure-test logic, consistent with that doctrine but ours, stated so you know which is which. That is the whole test: bare "Launch" where the kind survives a medium change, "Brand Film" where the static sibling is already its own type. The singular labels ("Testimonial", not "Testimonials") are a deliberate deviation from the standard's plural-for-classes default, allowed because it matches how you talk about a piece.
The Retail & Commerce table. The retail-media standard splits the store into placement zones (exterior · entrance · checkout · in-aisle · connected) and keeps the piece kinds smaller — shelf talker, hanging sign, floor graphic, screen, interactive display. The table's pick between keeping the named displays (a) or one generic display kind plus a zone value (b) now shows that evidence in place. The standard leans (b); the kind-test logic leans (a); your mark decides.
The product-page pushback. The biggest commerce platform's own vocabulary agrees with the surface: its enhanced product-page content is a named content piece composed onto the product detail page — the page is the surface, the pieces keep their kinds.
The Brand & Identity element groups. The industry's canonical guideline elements are logo, color, typography, imagery, and voice — which the seven groups carry. Two honest notes now sit on the review page itself: motion as a top-level group is our deliberate extension (the platforms subordinate it), and the new-flagged gradient rows have standards support (the design-token standard makes gradients a first-class type).
Sponsored Content and Promotional Ad. Rule 3 pressure applies to both; the naming section now carries the boundary notes — advertorial and the designed ad unit are defensible kinds, and the alternative (real type + ownership fact) is stated beside each.
How big reviews are run well — and what the review pack does about it
The second pass looked at how serious organizations collect decisions at this scale. The consistent findings, each encoded first in the walk console (now a record) and carried into the review pack, the surface you answer on since 25 August:
| The evidence | What the console does |
|---|---|
| Ratification works on a small closed verdict vocabulary, with "everything marked" as the explicit gate | four marks — Confirmed · Amended · Open · Thread — and a computed N-of-M progress line |
| A blocking verdict without an explanation is not actionable | choosing Amended or Open prompts for a note |
| Reviews collect best as a pending batch, released as one submission | marks persist locally as you go and export as one block you paste back |
| Notes must anchor to the item they concern | every note lives on its card and exports with it |
| Judgment quality degrades past thirty-odd items in one sitting | the console is chunked into sections with per-section progress — walk it in sittings, it saves itself |
| A judgment should be made on a visible delta, not a description | each card shows what exists today beside what is proposed |
How the docs stay correct — what the best documentation teams do
The third lane checked our documentation practice against the strongest public ones. Most of it validated what this site already does: prose linting and structure checks as commit gates, generated content never hand-edited (fix the source, regenerate), decisions as versioned markdown published to a static site, plain language as the baseline — with the striking evidence point that expert readers prefer plain language more than novices, not less. One practice we lack is on the open-questions ledger for your mark: a per-page "last reviewed" date with an alarm for pages nobody has touched — the checks catch bad edits today, but nothing catches a page that quietly goes stale untouched.
Sources
The full cited records behind this page, with every verbatim quote and source link, are the two research documents in the repository's reference folder (16 August and 17 August, named content-taxonomy-best-practices-research and best-practices-research-extension). Nothing on this page asserts anything those records do not carry a quote for.