The Authority Asset Framework

Introduction

The framework has six parts: the taxonomy of asset classes, the ratings layer, the selection logic, the original research methodology, the build standards, and the measurement and lifecycle layer. It connects directly to the Five Signal Model: Identity Clarity (the entity layer), Subject Authority (assets prove you know the subject), Meaning Architecture (assets structure the subject), Ecosystem Validation (the external layer) and Signal Consistency (dates, sources and named authors that match everywhere the brand appears).

What changed in v1.1, v1.2 and v1.3

Version 1.0 was reviewed against the current source-selection behaviour of Google organic, AI Overviews and AI Mode, Gemini, ChatGPT, Perplexity, Bing and Copilot, and Claude. Five independent assessments converged on the same finding: the original framework built citable pages but not a believable source. This version adds an eighth asset class (external validation and entity assets), an information-gain filter, query fan-out mapping, static companions for interactive tools, a conversion bridge, corrected schema and llms.txt guidance, anti-scale safeguards, and a measurement lifecycle.

v1.2 rebuilt the ratings layer after a second seven-engine stress test: retrieval and citation are now scored separately per engine, extraction and rendering risk are explicit dimensions, decay is paired with refresh cost per channel, a mechanical reference table records what each engine actually sees per asset type and what it cites the page as, and Classes 7 and 8 moved from enablers to sequencing prerequisites. The ratings are framed openly as working hypotheses: nobody outside the engine teams knows the real weights, so these are best guesses held to account by the Part 6 measurement loop.

v1.3 corrects the citation-lens error the v1.2 restructure introduced: rating everything by citability made "poor citation surface" read as "poor asset", and nearly cancelled legitimate link, ranking and conversion assets. Every asset type now carries a primary job label (RANK, EARN LINKS, RETAIN, CONVERT, GET CITED, SUPPORT TRUST, INFRASTRUCTURE) and is judged only against that job; behind the label sit six payoff columns and four viability modifiers (distribution dependency, uniqueness threshold, maintenance burden, harm exposure). Verdicts corrected on re-audit: spreadsheets and models, embeddable widgets, quizzes, hand-built segment pages, explainers and calendar-driven expert commentary rescued for their non-citation jobs; llms.txt, un-adopted APIs, literal data mirrors, consumer-less feeds and generic curated libraries confirmed weak on every payoff.

Part 1 — The Taxonomy: Eight Asset Classes

Every Authority Asset falls into one of eight classes. A site does not need every type; it needs the right types for its niche, built to standard, plus the eighth class in all cases, because without it the other seven under-perform.

Class 1 — Reference assets

The workhorses. Evergreen, high citation value, and the hardest class for competitors to replicate once established.

• Indexes — structured collections of criteria, providers, rules or prices, searchable and filterable. The strongest single asset type, because nobody else holds the same cut. Examples: a lender criteria index, a card-processing fee index, a price-per-effective-dose index.
• Comparison tables — provider-vs-provider and product-vs-product, dated and versioned, with inclusion criteria and commercial relationships disclosed on the asset itself. Undisclosed pay-to-play comparison content is avoided by answer engines.
• Glossaries and definition libraries — term-by-term plain-English definitions. Cheap to build and slow to decay, but engines prefer dictionaries and institutional sources for generic terms; a glossary earns citations only where its definitions carry niche-specific information gain.
• Directories — structured provider, regulator or scheme listings with consistent fields per entry, published inclusion and exclusion rules, and verifiable current status.
• Timelines and chronologies — regulatory change histories, rate histories, scheme deadlines presented as structured chronology. Rare in most niches, which makes them easy citation wins.
• Curated libraries — books, podcasts and resource collections on the subject, annotated. A soft signal of subject immersion; only worth building with genuine annotation.

Class 2 — Data assets

Numbers you own, or present better than the body that publishes them.

• Stat pages and data roundups — curated statistics on one topic with sources, updated on a schedule so the URL becomes the canonical reference.
• Original research — primary surveys, derivative narrowing of public data, or recombination of datasets. Covered fully in Part 4.
• Trackers — live or regularly updated figures: rate trackers, arrears trackers, insolvency trackers, countdowns. The gold standard is a self-updating tracker fed by an official data pipeline. A stale tracker is worse than no tracker.
• Benchmarks — "average X for businesses like yours". Positions the site as the measuring stick for its niche.
• Public data mirrors — cleaned, humanised versions of official datasets. Warning: reformatting alone is not added value. A mirror earns citation only when it adds resolution, interpretation, history or usability the source lacks, and it must state what it adds. Where the official source is directly available to engines, an unimproved mirror is redundant.
• Annual-refresh assets — assets built around a predictable official calendar: April statutory rate changes, annual data releases, yearly date lists. The asset is the refresh cycle itself. Carry the year in the title, meta description and H1; update these every January; version editions explicitly ("2026 Index") so prior years remain citable as history. Most AI queries now carry the current year; undated assets lose to dated ones.
• Evidence assets — first-hand proof: test logs, screenshots, photographs, correspondence, FOI responses, collection logs, code and reproducibility packages. Engines increasingly reward demonstrable first-hand experience; evidence assets are how experience is proven rather than claimed.

Class 3 — Interactive assets

Tools that answer a personal question. High engagement, strong snippet performance, natural link targets. One rule governs the whole class: the engine cannot operate the tool. No AI crawler executes JavaScript, moves sliders or reads a computed output. The page carrying the tool gets cited as "a tool exists here"; the page that also shows the tool's output in static HTML gets cited as the answer. Every interactive asset therefore ships with a static companion: the formula, worked examples at realistic values, and a results table across common scenarios, in plain server-rendered HTML beneath the tool. This is the machine-readable form of "show the working".

• Calculators — affordability, stress tests, cost comparisons, repayment modelling.
• Checkers and eligibility tools — "can I get X if Y" as an interactive rather than an article, including solvency and wrongful-trading style diagnostics.
• Quizzes and diagnostics — structured self-assessment producing a result worth linking to, including "do you actually need this" honesty tools.
• Configurators and estimators — build-your-scenario tools, including combination tools that join an official dataset with data the user holds themselves.
• Tools that accept personal or sensitive data carry a stated privacy and data-minimisation position. Where useful, tool results are shareable and resumable via stable URLs, which also gives engines crawlable example outputs.

Class 4 — Downloadable assets

Things people take away and keep. Consistently the thinnest class on most sites, which makes it a reliable quick win.

• Templates — letters, invoices, dispute documents, scripts, contract clauses.
• Checklists — printable, task-tied process checklists, published as real documents.
• Guides as PDFs — long-form pieces packaged as downloads. The full content also exists as HTML; a PDF-only asset is invisible to most answer engines.
• Spreadsheets and models — working calculators people keep: cashflow forecasts, fee comparison models. Office-friendly formats travel well in business contexts and are legitimate capture gates because they extend the free tool.
• Sample documents — worked examples of completed forms, applications and statements.
• Datasets as CSV/JSON — downloadable copies of your data and research, each with a stated licence, version identifier and provenance note. Journalists and researchers cite datasets; engines cite the explanatory HTML around them. Both must be first-class.

Class 5 — Editorial structures

Content shaped as an asset rather than an article.

• Pillar hubs and spoke clusters — the definitive resource with sub-pages and dense internal linking. The default architecture, with a discipline: clusters are built to cover real sub-queries, not to maximise page count. Fewer, stronger pages beat bloated repetitive clusters.
• Q&A libraries — structured question sets, one capsule answer per question, built from real query and conversational demand. Kept genuinely useful; a Q&A library built to harvest search demand reads as scaled content and is treated accordingly.
• How-to and process guides — numbered stages, prerequisites, documents needed, timings.
• Explainers — single-concept pages that define and resolve one thing.
• Case studies and worked examples — real or anonymised scenarios with numbers, attributed to a named individual. Rarely built as a distinct type, which is why they stand out.
• Expert commentary streams — a running, author-attributed take on each period's data or news, including rapid-response commentary on regulatory and market events. Turns a named expert into a recurring source and feeds news-shaped demand the evergreen classes miss.
• Troubleshooting and emergency clusters — "something has gone wrong" content: frozen funds, terminated accounts, missed deadlines. High anxiety, high loyalty, high citation. The one context where an immediate, non-obstructive contact route above the fold is correct.
• Segment and vertical pages — geographic, sector and matrix combinations. The highest-risk editorial type: built only for combinations real searchers use, each carrying genuinely distinct criteria, data and documentation. Template-generated matrix libraries are the asset class most likely to trip scaled-content enforcement.
• Multimedia assets — original video with transcripts, annotated diagrams and charts, audio and podcast episodes with transcripts. These participate in separate surfaces (video search, YouTube-weighted engines, image search, multimodal answers) that text assets never reach, and the transcript is the citable text layer.

Class 6 — Technical and machine-facing assets

Assets whose primary audience is a machine: crawler, engine or integrator.

• APIs — public endpoints serving your data. Moves the site from content to infrastructure; integrators compound.
• Embeddable widgets — your calculator or tracker served on other sites, linking back. Must render reliably; a fragile widget is ignored as a source.
• Structured feeds — RSS or JSON feeds of updated figures, so changes propagate without a crawl.
• Code repositories — public repos holding tool logic, calculation code and analysis notebooks. Heavily cited in technical contexts and strong reproducibility proof for research.
• llms.txt and AI-instructions pages — experimental discovery aids, nothing more. No major engine has committed to them as a ranking or citation input, and instruction-like content addressed to models can be filtered as suspected prompt injection. Standard: factual site description only, no instructions to models, no claims absent from public pages, verified on the same schedule as public copy. Never treat as a trust mechanism.
• DOI-anchored publications — research minted with a DOI through an open repository. A DOI makes work persistently identifiable and formally citable; it does not by itself make work authoritative. Value is realised only alongside the distribution and corroboration work in Class 8.

Class 7 — Proof and trust assets

Strength signals rather than traffic assets. These make every other class more citable.

• Methodology pages — how numbers, scores, rankings and inclusion decisions are produced, including conflict and remuneration disclosure for anything with a commercial dimension.
• Data Sources & Updates page — every source used on the site, date last checked, next scheduled update. The citation anchor that turns a set of tools into a maintained reference.
• Editorial, update, correction and funding policies — the standard trust template including AI disclosure, plus a corrections log and right-of-reply route. A visible correction record is a trust signal, not an admission.
• Publisher identity — the legal entity behind the site stated plainly: registered company name and number, regulator registrations where they exist, ICO registration, contact routes, and who receives any enquiry a form generates. Machine trust anchors to legal identity, not to an about page.
• Author hubs — credentials, qualifications and publication history for named experts, linked from every asset they stand behind, with Person schema and sameAs links to external profiles. An author hub documents authority; Class 8 manufactures it.
• Versioned knowledge products — changelogs, release notes, archived editions and correction records on data assets, so engines and users can see what changed and why.
• How-to-cite blocks — copy-paste citation text on data pages. A courtesy for researchers and journalists; no evidence engines read them as instructions, so never rely on them for selection.

Class 8 — External validation and entity assets

The class v1.0 lacked, and the one the engines weigh most. Everything in classes 1 to 7 is self-asserted. Engines resolve the brand and author as entities and test whether the wider web corroborates them before citing. This class is mandatory for every flagship asset and every YMYL site.

• The entity package — Organization and Person schema with disciplined sameAs links connecting the site to Companies House, regulator registers (FCA and equivalents), ICO, ORCID, Wikidata where eligible, and major professional profiles. One unambiguous identity everywhere, with identifiers matching to the character.
• Register and body evidence — regulator register entries, professional body memberships, accreditations with real selection criteria, deep-linked from the site so the credential is verifiable rather than claimed.
• Third-party corroboration — coverage, citations and mentions from trade press, national media, institutions and recognised specialist publishers. Unlinked brand mentions and co-occurrence with the subject matter count; engines read consensus, not just links.
• Adoption of your work — third parties using, citing or republishing your data, indexes and methodology. The strongest authority evidence there is: the market treating you as a source.
• Distribution assets — newsroom and press page, journalist data packs (quotes, bios, charts, embargo-ready summaries), newsletter, syndication feeds and a media contact route. Research without a distribution plan is a page, not a publication.
• Review and reputation presence — the review platforms and directories relevant to the niche, actively maintained, with complaints handling that shows. For lead-generating sites, engines and users both check.
• Author external footprint — the named expert's publications, quotes, talks and podcast appearances off-site. The author entity must exist beyond the brand's own domains, or it reads as manufactured.
• Community presence — genuine participation where the niche is actually discussed (sector forums, Reddit, LinkedIn, Q&A platforms). Engines lean on community sources for experience and opinion queries; a brand absent from the conversation cedes those citations entirely. Participation, not spam.

The on-page layer

One structural pattern sits underneath the taxonomy rather than inside it: answer capsules with fan-out coverage. Every asset page opens with a short, self-contained answer block, and the page then covers the natural sub-questions engines generate from the head query, each as its own extractable, heading-matched passage with its supporting evidence close by. Engines assemble answers from multiple sub-query passages, not from one heroic page; passage-level self-containment across the cluster is where citation is won. Capsules are written per page, not templated across pages, because repeated capsule text reads as duplication.

Interactive lead funnels (stepped quote and eligibility wizards) remain commercial assets rather than citation assets, but the wall between the two is a bridge, not a moat: see the conversion bridge standard in Part 5.

Part 2 — The Ratings Layer

v1.2 restructured the ratings around the seven-engine stress test: retrieval and citation are separate events, scored per engine, and the citable surface of every asset is its static HTML text layer. v1.3 corrects the error that restructure created. Rating everything through the citation lens made "poor citation surface" read as "poor asset", and it nearly cancelled legitimate link, ranking and conversion assets, an embeddable widget among them. Citation is one payoff of six, not the definition of value.

The primary job label

Every asset type now carries a primary job, and the governing rule is: an asset is judged, and cancelled, only against its primary job. An asset that fails at its primary job is cut; an asset that succeeds at it is never cut for performing poorly at a job it was not built for. The labels:

• RANK — built to win organic positions and pull search traffic in its own right.
• EARN LINKS — built to give other sites an editorial reason to link, embed or quote; links feed authority, and authority feeds position.
• RETAIN — built for repeat visits, bookmarks, subscriptions and brand.
• CONVERT — built to capture, qualify and route leads, even if no engine ever quotes it.
• GET CITED — built to be the extracted, attributed answer in engines and press.
• SUPPORT TRUST — built to raise the credibility and evidence base of everything else; earns nothing directly.
• INFRASTRUCTURE — retrieval and distribution plumbing with no direct payoff of its own; justified only by what it carries.

Behind the label, each flagship is scored on six payoff columns: organic acquisition, citation acquisition, link acquisition, retention and distribution, lead contribution, and authority support. Four viability modifiers then adjust the verdict: distribution dependency (some assets only work with outreach or a named consumer attached — a widget without a distribution campaign is an expensive calculator, an API without a user is maintenance), uniqueness threshold (how much original substance the type needs before it stops being interchangeable), maintenance burden, and harm exposure (scaled-content, link-scheme, legal and stale-data risk).

The mechanical reality: what the engine sees, and what it cites the page as

AI crawlers execute no JavaScript, click nothing, fill nothing in, and see aggregate assets only in their default state. The consequences per asset type:

Freshness and decay by channel

Decay is channel-specific, and the layer now pairs it with refresh cost, because when an asset rots matters less than what it costs to stop it. A pipeline-fed tracker (fast decay, near-zero refresh cost) is a build; a manually-maintained regulatory guide (fast decay, high refresh cost) is a consolidation candidate before it is a build.

Ratings by class

Each type carries its primary job, effort, an honest payoff description (citation and non-citation), and decay paired with refresh cost. Verdict corrections from the non-citation re-audit are reflected throughout.

Reference assets

Data assets

Interactive assets (companion mandatory)

Downloadable assets

Editorial structures

Technical and machine-facing

Proof, trust and external validation

The channel scorecard

Before a flagship asset is commissioned: first, name its primary job and the success metric that job implies — the asset will be judged against that and nothing else. Then apply two hard gates. Gate one, rendering: if the asset's answer only exists behind interaction, JavaScript, a gate or a non-HTML format, it fails for every AI channel until the static layer is designed — unless GET CITED is not its job. Gate two, authority anchor: on YMYL query classes, if the publisher entity is not machine-verifiably linked to its anchors (Companies House, regulator register, ICO), citation likelihood is capped regardless of content quality. Then score the six payoff columns (organic, citation, links, retention, leads, authority support), scoring retrieval and citation separately per engine (Google organic, AI Overviews and AI Mode, Gemini, ChatGPT live and cached, Perplexity, Bing/Copilot, Claude); then apply the four viability modifiers: distribution dependency, uniqueness threshold, maintenance burden, harm exposure. An asset scoring low on its primary job is redesigned or dropped before build; an asset scoring low elsewhere is not.

Reading the ratings: indexes, trackers, calculators-with-companions and original research remain the four disproportionate-strength types for citation, each earning it only through its static, extractable layer. The v1.3 re-audit rescued a second tier whose jobs are not citation: spreadsheets and models (CONVERT), pipeline-fed embeddable widgets (EARN LINKS), quizzes and triage diagnostics (CONVERT), hand-built segment pages (RANK) and calendar-driven expert commentary (EARN LINKS). Classes 7 and 8 remain prerequisites that gate citation eligibility itself, and they sequence first.

Part 3 — Selection Logic: What Should This Site Build?

Never build the full taxonomy on one site. Run every candidate asset through seven filters, in order.

Filter 1 — The authority ceiling test, with the corroboration question

Can this entity plausibly be cited for this query class? For YMYL queries, engines select government sources, regulators, institutions and externally corroborated experts, and deeper reasoning modes weight institutional sources more heavily still. Credentials answer half the question; the other half is corroboration: who else says this brand and author are credible, and does the wider web reflect it? If the ceiling test fails, change the query class, close the corroboration gap through Class 8 work, or do not build the asset. Self-asserted credibility does not pass this filter.

Filter 2 — The honest-hook test

Does the claimed source or angle genuinely exist for this vertical? "Built on official data" works where an official dataset exists and collapses where it does not. A positioning hook that only fits part of a portfolio must not be stretched across all of it.

Filter 3 — The information-gain test

What does this asset reveal that the current top sources and AI answers do not? A new observation, cut, series, calculation, comparison dimension or decision rule counts; a prettier restatement does not. The unique contribution must be identifiable in one or two extractable passages, and provably derived rather than rewritten. If the gain cannot be stated in a sentence, the asset is redundant before it is built.

Filter 4 — Demand and demand shape

Is there real demand, and what shape is it? Evergreen head terms suit indexes and hubs; situational long-tail suits Q&A and checkers; annual spikes suit refresh assets; seasonal and visual demand adds Discover and multimedia angles; news events suit commentary streams. Measured keyword volume is not the whole picture: some of the best citation opportunities recur across thousands of varied conversational prompts with no conventional volume, and mid-funnel comparison intent often converts best of all. Map the query journey (discovery, comparison, validation, action), not just the keyword.

Filter 5 — Query fan-out mapping

Engines split the user's question into sub-queries and cite pages that surface consistently across the set; ranking first for the headline term no longer guarantees selection. For every asset, document the natural sub-questions (typically five to ten) that fan out from the target query, and confirm the asset or its cluster answers each one as a self-contained, heading-matched passage. Fan-out coverage is a build requirement, not an editorial nicety.

Filter 6 — Competitor weakness

Where are incumbents thin? The recurring pattern: thin provider lists, gated PDFs, undated data. The counter: expert-updated depth, un-gated tools, shown working, dated sources. Whole-class white space (timelines, models, case studies, widgets are commonly absent everywhere) is the cheapest differentiation available.

Filter 7 — Channel fit

Name the target channel before build. Utility tables surface in Bing, which feeds Copilot; capsules and companions win snippets and AI answers; datasets and DOIs earn journalist and academic citation; video and transcripts reach YouTube-weighted surfaces; trackers earn repeat direct visits; community presence earns experience-query citations. Register the site on the channels the assets need: Search Console, Bing Webmaster with IndexNow, and explicit crawler access per engine (OAI-SearchBot, GPTBot, PerplexityBot, Claude, Google-Extended) as deliberate policy decisions, not defaults.

Part 4 — Original Research Methodology

Original research is the highest-value asset class and the least understood. The bar is not "new data". The bar is a question with real demand that nobody currently owns the answer to, answered with a method you can defend, and distributed so that third parties corroborate it. There are three routes.

Route 1 — Primary observation

Generate data that does not exist: your own operational data, a commissioned or self-run survey (with stated sampling and consent standards), freedom of information requests, or systematic collection from public records. Highest effort, highest defensibility, and the only route that makes you the primary source outright. The question comes first, the collection method second; never collect data and then hunt for a question.

Route 2 — Derivative narrowing

Take existing public data and cut it by a dimension nobody else has published: region, sector, company size, time of year. A national statistic becomes a regional index; a UK-wide rate becomes a sector league table. The insight is new resolution, and the specific cut exists only on your page. The fastest route to ownable research and the default for most sites. The information-gain test applies in full: the cut must reveal something, not merely subdivide.

Route 3 — Recombination

Join two or more public datasets that have never been put together: insolvency rates against payment terms by sector, adviser density against regional demand, price against effective dose. The joined view is the original work. Recombination also powers combination tools, where the user supplies one dataset themselves.

Worked example — invoice finance

Topic given: invoice finance. The commodity content is saturated and earns nothing. The research process:

1. Find the unowned question. Candidate: "what day of the week do UK businesses actually pay invoices?" Nobody publishes this. Concrete, quotable, and every finance journalist writing about late payment wants the line.
2. Choose the route. Primary observation if payment-timing data is available operationally or by survey. Otherwise derivative narrowing: take published UK late-payment statistics and cut to a region or sector, so the national story becomes "late payment in the East of England" or "which UK sector pays invoices fastest".
3. Verify nobody owns it. Search the exact question and its fan-out sub-questions across Google and the AI engines. If an answer exists, narrow further until the cut is genuinely unclaimed, and record the information gain in a sentence.
4. Source and document. Named datasets, retrieval dates, sample sizes, exclusions, the exact calculation, licences on inputs, and the evidence trail (collection logs, code, notebooks) captured during the work, not reconstructed afterwards.
5. Publish in citation format. Capsule answer first, then the tables, then the method, with each claim's source and date beside it. Dataset ships as CSV with licence and version. How-to-cite block on the page. DOI minted for the formal version. The named, qualified author stands on the work; it links to their hub and external profiles.
6. Distribute. The step v1.0 lacked. Journalist data pack prepared; the finding pitched to trade press and relevant nationals; the dataset deposited where researchers look; the author entity updated everywhere (ORCID, professional profiles); sector bodies and commentators sent the cut relevant to them. Research without distribution is a page. Third-party pickup is the point, because corroboration is what engines select on.
7. Build the refresh cycle. The finding becomes an annual, year-tagged index: same question, new period, versioned editions with prior years archived and citable. Compounding begins at the second release: the trend is a second story and prior citations carry forward.

Publication standards for research

• Official data and your own analysis labelled separately, never blended. A projection is labelled as a projection, always.
• Every figure states the period covered and the date retrieved or calculated, with the source beside the claim, not in a footer.
• Method transparency complete enough that a competent third party could reproduce the number; code and collection logs published where they exist.
• The dataset ships with the page as CSV or JSON with licence, version and provenance. The explanatory HTML and the dataset are both first-class citation surfaces.
• A DOI is minted for research the brand wants formally cited, the publication added to the author's hub and external profiles, and the distribution plan executed. All three, or the research is not finished.

Part 5 — Build Standards

Every Authority Asset, in every class, meets these standards before it ships. Most were learned from observed failures; each names the failure it prevents.

Answer and evidence standards

• Answer first, fan-out covered. The capsule answer sits above the tool; the tool sits above the explanation; the page's sub-questions are each answered in a self-contained, heading-matched passage. Never bury the high-demand answer beneath a calculator or an introduction.
• Show the working, and make it the real working. Every calculated figure displays its method, and the displayed method is the method actually used. Displaying one formula while calculating with another is a documented trust failure and, on health or money queries, a disqualifying one.
• Static companions on every interactive tool. The formula, worked examples at realistic values, and a results table across common scenarios, in server-rendered HTML. Engines cannot execute tools; a tool without a companion is invisible to every AI channel. Key content is verified readable with JavaScript disabled.
• Honest data labelling. Data carries its release and period, never a current-year title it does not cover. Old data dressed as new is deception, and engines detect it.
• Figure currency and year discipline. Annually-changing figures state the year applied, the date checked, and link the official source. Annual assets carry the year in title and H1, updated each January, with versioned editions archived.
• Claim-level sourcing. Every material claim has its source, date and qualifying context adjacent to it, using the best available primary or authoritative source, with source limitations stated. Original observation and practitioner evidence are valid sources when presented as such. If no source exists and it is not your own finding, the claim does not ship.

Trust and identity standards

• Named, verifiable authorship. Every asset carries a named author linked to their hub, and for YMYL content the credential is externally verifiable (register entry, professional body, academic profile) and deep-linked. Self-reported credentials are insufficient; reviewer attribution means the reviewer genuinely reviewed. Engines now surface voices, not just pages; one publisher name spread thin across many sites is a scaled-content footprint, not authority.
• Publisher identity on every site. Legal entity, company number, regulator and ICO registrations, contact and complaints routes, and disclosure of who receives enquiries. For introducer models in regulated niches, regulated status and the introduction path are stated plainly.
• Commercial conflict disclosure on the asset. Rankings, comparisons and directories state inclusion rules, remuneration and conflicts on the page itself, not only in a site-wide policy. Undisclosed pay-to-play is a selection killer.
• Versioning and corrections. Data assets carry a visible change log; corrections are logged, not silently overwritten. Freshness stamps reflect substantive change only: artificial date refreshing without real update is maintenance theatre and engines treat it as such.
• YMYL care. Health and money tools state estimate status and the professional to consult, and carry named reviewers with review dates. Disclaimers manage risk and clarity; they do not create citation eligibility, so they supplement verification rather than substitute for it.

Access and technical standards

• Crawlable, canonical, renderable. Every asset has one canonical URL, renders its content without user interaction, passes mobile parity, meets accessibility basics and keeps Core Web Vitals within thresholds. Extraction quality is downstream of page quality.
• Static state for every aggregate. Paginated indexes carry a view-all or per-page static URL; filterable directories expose their important filter states as real static URLs (a /directory/london, not a ?region=london parameter); tables are server-rendered with real cells. Engines see page one and the default state only; a row the crawler cannot reach is a row that cannot be cited.
• Deliberate crawler policy. Robots directives set per crawler as policy: OAI-SearchBot, GPTBot, PerplexityBot, Claude's crawler, Google-Extended, Bingbot. Snippet controls (nosnippet, max-snippet) reviewed so eligibility is not accidentally destroyed. IndexNow fired on substantive updates and on removals.
• Schema by eligibility, entity first. Organization, Person, Dataset and sameAs into Wikidata, Companies House and regulator registers are the priority markup, because entity resolution is what engines test. FAQ markup only where genuine visible FAQ content exists (rich-result display is restricted); HowTo rich results are deprecated. Schema is chosen by current eligibility and semantic accuracy, not habit, and applied at build.
• Machine-facing claims verified, never instructional. llms.txt, AI pages, APIs and feeds carry only facts present on public pages, verified on the same schedule, with no instructions addressed to models. Instruction-like content risks being filtered as prompt injection; unverified claims are false claims waiting to be repeated.
• HTML twin for every non-HTML asset. PDFs, spreadsheets and interactive outputs have full HTML equivalents. Non-HTML-only content is invisible to most answer engines.

Commercial and lifecycle standards

• No commercial interruption or distortion before the answer. Nothing between a tool and its result; nothing that biases the answer. This replaces a blanket above-the-fold ban: a non-obstructive urgent contact route is correct on distress queries, and intent-matched CTAs are legitimate after value is delivered.
• The conversion bridge. Citation assets and lead funnels connect: post-answer capture modules, contextual take-away gates (the raw model behind the free tool, saved or emailed results, alerts and benchmarks for early-stage users), and a defined lead path from every high-intent asset. Consent and data minimisation throughout. The asset answers first, then converts; it never charges for the answer itself.
• Un-gated by default. The asset itself is open; gating is reserved for take-away formats where capture is the explicit trade.
• Anti-scale discipline. Matrix libraries, templated capsules, fleet-replicated structures and demand-harvesting Q&A are the patterns most likely to trip scaled-content enforcement, and the risk multiplies across a portfolio sharing authors and templates. Each page justifies itself with unique intent, data and evidence, or it is not built. Fewer, stronger pages beat coverage theatre.
• Maintenance scheduled at build. Every asset ships with an owner, a refresh trigger (a date, an official release, or a pipeline), its Data Sources & Updates entry, and a decommissioning rule: stale assets are refreshed, consolidated or retired, never left contradicting the current data. An asset without a maintenance plan is a liability with a launch date.

Part 6 — Measurement and Lifecycle

v1.0 treated visibility as a one-way output of asset quality. It is not: retrieval pipelines differ per engine and change without notice, and no single metric exists. Every flagship asset therefore carries a measurement loop.

• Citation monitoring. Monthly tracking of where each flagship asset is cited across Google AI surfaces, ChatGPT, Perplexity, Copilot and Claude, using Bing Webmaster AI performance reporting, referral tagging, and prompt-based spot checks. Whether the site is merely listed among sources or actually cited for the key claim is the metric that matters.
• Search and index health. Indexation, rankings, snippet capture and crawl coverage per engine, including independent-index engines that do not inherit Google or Bing coverage.
• Commercial measurement. Source-specific conversion: organic versus AI-referred visitors, assisted conversions (AI-referred users often return later via brand search), and lead quality fed back into the asset roadmap, so citation volume is never mistaken for commercial value.
• The lifecycle decision. Each asset is periodically refreshed, consolidated or retired on the evidence. The framework is a loop, not a production line: build, prove, distribute, measure, decide.

Applying the framework

The framework is applied per brand as a gap map: the taxonomy is scored against what each site already holds, white space identified (fleet-wide white space first, since it differentiates every site at once), candidates run through the seven filters, flagship assets scored on the channel scorecard, and the build list sequenced by the ratings layer with Class 8 work attached to every flagship. The applied view for each brand is maintained as a live companion to this document.