Original Source Asset Development for AI SEO

Original source asset development is the top technique in the MERIT Framework, and the asset you pick decides most of what follows: the right one builds AI visibility for years, while the wrong one (built with the same effort) produces a polished page nobody cites. This chapter covers the 5 viable asset types, walks through the decision framework, maps the platform and industry variants, and lays out the selection workshop teams use. Most teams pick their asset type in a single short meeting; the teams that earn category-defining AI visibility treat that choice as strategy work with its own outputs.

Why This Technique Matters

The default is to pick whatever feels most rigorous (usually research) or whatever the team already knows how to make (usually blog posts dressed up as opinion). Both defaults are common and both are usually wrong, because they optimize for what looks defensible in a planning meeting instead of for what earns citations.

I'd replace both of those questions ("what feels most rigorous" and "what can we already produce") with one: "which asset earns the most sustained AI citations in our category?" Asked that way, the answer is rarely the default, and the asset that turns effort into citations most efficiently in a category is often one the team would never have picked on instinct.

The stakes are real, because original assets are the one durable citation surface a brand controls and the one place it can add the net-new information gain AI systems filter for. The rest of MERIT can't do that job: Mentions techniques depend on third parties for placement and timing, Relevance techniques only tune what you already have (and if the source content is derivative, structure tweaks hit a ceiling), and Inclusion techniques are technical baseline work, since entity optimization and crawler access make an asset findable without earning it a citation.

Original source asset development is the one technique where the brand controls the input, the timing, and the topic, so picking the wrong asset wastes the lift the rest of the Evidence pillar is built to amplify.

The cost of getting it wrong is opportunity cost. A rigorous-looking asset that earns few citations stalls on its own, and it also stands in for the other assets that never got made, because the operator time and editorial attention were already spent; one bad pick can cost you the framework, the calculators, and the opinion pieces that could have run instead.

The Five Viable Asset Types

There are 5 asset types that reliably earn AI citations when they're built well and promoted right, and AI systems don't prefer one over another; what looks like a format preference usually comes down to promotion quality and co-citation depth. So I'd pick the type that fits your situation over the one that feels most defensible.

Expert Frameworks and Methodologies

A framework is a named structure that organizes existing ideas so people can think about a problem in a new way. The MERIT Framework, Jobs-to-be-Done, Growth Loops, RACI, AIDA, the messy middle, and similar widely-cited frameworks share 3 traits (a memorable name, a set of components, and a path for using it), and they earn citations because they give writers, analysts, podcasters, and operators a shared vocabulary for a complex topic. Once a framework is named and adopted, people cite it by its name and stop quoting the text, so it travels by reference.

When frameworks win. Frameworks fit categories where the current vocabulary is borrowed, vague, or in conflict (a clear, well-built framework replaces the borrowed words because it's easier to use), categories where the same problem comes up across many surfaces such as analyst reports, industry publications, podcasts, and conferences, and categories where operator experience repeats across many cases, since a framework is a synthesis of that experience and needs no new data. They also fit when the strategic intent is thought leadership and category positioning rather than transactional capture.

When frameworks lose. A framework struggles when the category already has 1 or 2 named frameworks with strong analyst adoption, when the team can't sum it up in a single sentence or a single slide, when the target audience is buyers in late-stage decision instead of the people shaping the category, or when the brand has no named author (founder, executive, or named SME) to attribute it to.

What it takes: operator time for synthesis, an editorial pass, diagram design (the highest-yield design work a framework gets), a canonical page build, and promotional copy. The work runs in order, with the foundation first, then distribution, then depth and reinforcement.

Citation profile: a slow ramp at the start that compounds once the first analyst or industry publication uses the framework's name. Frameworks have the most durable citation curves of any asset type, because each future piece written about the topic may cite the framework by name.

Common framework mistakes. The forgettable name is the first one; terms like "Strategic Alignment Methodology" aren't names, and a framework needs one a sharp non-expert can recall after hearing it once. Over-factoring is the second (12 parts in nested matrices, when the limit is 3 to 7 parts total). The third is no application path, since definition-only frameworks without diagrams, templates, or worked examples don't travel, and the fourth is a borrowed framework with new words; relabeling existing methods fools nobody, so I'd either build new work or cite the original.

Worked example: the MERIT Framework itself. I drafted it from patterns that held across Searchbloom partner engagements, iterated on the name and the parts over multiple cycles, and designed a single diagram for it. The canonical whitepaper went up at searchbloom.com/merit-framework-whitepaper/, and distribution ran through 3 contributed pieces in industry publications, 12 LinkedIn long-form posts, and 4 podcast appearances. The work expanded into a per-pillar playbook in April 2026, with citation density tracked on the cadence in Chapter 13, and the outcome is that the framework name now gets retrieved consistently for queries about AI SEO, AEO, and GEO.

Expert Opinion and Analysis

Opinion is defensible operator reasoning under a named expert byline, on a topic where the operator's experience produces a non-obvious view. Most B2B teams underuse it because they treat it as risky, but that risk (being wrong in public) is what makes opinion citable. Our counter-claim at Searchbloom, drawn from partner engagements, is that AI systems cite opinion at rates close to empirical research when the source has verified expertise and the claim is specific.

When opinion wins. Opinion fits when the operator has deep domain experience or runs a brand with visible category authority, when the category has new questions that data doesn't cover yet, and when the team needs fast citation traction with little production capacity beyond operator time. It also fits categories with reactive moments (platform changes, competitor moves, rule shifts), where speed and sharpness produce outsized citation density.

When opinion loses. Opinion struggles when the brand has no recognizable expert byline, when the category rewards rigorous data above everything else (hedge fund research, regulatory analysis, deep technical fields), or when the team isn't comfortable making sharp claims; hedged "it depends" pieces don't earn citations no matter how credible the operator is.

The 3 traits of citable opinion. The first is that it's specific, meaning the claim is concrete enough to disagree with: "Content marketing matters" is invisible, while "Most B2B blog posts published in 2026 have zero AI citation potential because they restate widely available consensus" is citable. The second is that it's defensible, backed by reasoning the reader can follow and judge (even though opinion isn't a research finding). The third is that it runs counter to consensus in at least one specific way, because an opinion that restates the consensus adds no information gain and doesn't get cited.

What it takes: operator drafting time, an editorial pass, light design, and distribution through the operator network (plus optional PR coordination). Long-form canonical pieces move from idea to publication fast, reactive opinion publishes soon after the trigger event, and citation traction begins reasonably quickly for sharp pieces with a credible byline.

Citation profile: spiky early with a long tail. Sharp opinions compound, since each new query on the topic may surface the original piece, and operators on a steady monthly cadence build entity-topic links that AI systems learn to weight; after a sustained run of steady publishing, the operator becomes a citable source whose opinion pieces earn citations at multiples of the average rate.

Common opinion mistakes. Hedged opinion comes first: "Five things to consider" listicles and "it depends" pieces are analysis without a position, so they don't count as citable opinion. Restated consensus is next, because opinion that matches what everyone already says adds no information gain (the piece can be correct and still earn no citations). A brand byline is the third, since opinion under the brand name with no attributed person earns lower citation rates, so I'd put it under the founder, the CEO, or a named SME. The last is wide-topic publishing; operators who publish across too many unrelated topics weaken the entity-topic link, and I'd keep them inside the topical cluster that matches their expertise.

Data-Driven Research

Original research earns AI citations when the method is credible and the findings come in formats AI systems can extract. The visibility upside is real, but the time to citation is long and the failure modes are harsh, so under the wrong conditions research is the most expensive asset type to get wrong, and under the right ones it's the hardest competitive advantage for anyone to copy.

When research wins. Research fits a category with missing or stale data, a brand with real research infrastructure (panel access, statistics capacity, or partnerships with trade groups or schools), and a category whose analyst coverage rewards proprietary data.

When research loses. Without research infrastructure, underpowered work fills the gap, and low sample sizes get filtered out by the analyst tier whose co-citations create AI visibility. Research also loses when the category lacks analyst coverage or the team lacks method rigor or partnerships.

What it takes. The options span a range of research infrastructure. DIY survey tools are the lightest, though their sample sizes are usually too small for citable research; managed survey platforms sit in the middle (Wynter and Centiment offer B2B-validated panels, SurveyMonkey and Typeform are general-purpose with weaker B2B panel quality, and Qualtrics is enterprise-grade); and full research firms are the option for analyst-tier categories. I'd put research in the portfolio only when the brand has real research infrastructure, whether that's a managed platform, a research firm, or in-house capacity, because without it the brand is better served by another asset type than by underpowered work the analyst tier will filter out.

Sample size decides whether journalists, analysts, and AI systems will cite the numbers. The minimum viable range is 500-750 respondents (+/-3.5-4.5% margin of error at 95% confidence); the target, and the sweet spot, is 1,000-2,000 respondents (+/-2-3% margin of error); and the premium tier of 5,000+ respondents lets you cut the data by segment, which yields several citable claims from one survey. Stratifying for segments needs at least 200-300 respondents per segment.

Research moves slowly from design to publication and then needs more sustained time to accumulate citations, so I'd plan for the asset to reach steady-state value only after a long ramp.

Method rigor. That covers sample selection (document the recruitment method, screening rules, and response rates), statistics (descriptive stats, confidence intervals on key claims, and significance tests for comparisons), and a limitations section, which every credible study has. The analyst tier reads the method first and the findings second.

Citation profile: slow to start but durable. Statistical claims with proper attribution show up in AI citations for years if the method holds up to scrutiny, and the compounding is strongest for research that becomes the "category statistic" everyone cites.

Common research mistakes. Underpowered samples top the list: surveys under 500 respondents for broad claims fail analyst credibility, and my read is that low-power research is worse than no research. A missing method page is next, since research without a documented method gets dismissed by the analyst tier (Chapter 5 covers method docs in depth). Image-locked statistics are the third; numbers locked inside infographics are invisible to AI, so I'd always pair visuals with parsable HTML, also covered in Chapter 5. Gated research is the last, because gating research behind a form means AI systems can't index it.

Interactive Calculators and Tools

Calculators are the most underrated original source asset type. They earn citations because the URL returns a personal numeric answer, which is what generative AI systems need to surface for quantitative queries, and in categories full of gated PDFs and weak comparison pages, one well-built calculator (far less effort than original research) can displace those incumbents.

When calculators win. Calculators fit when buyer questions are quantitative ("how much will X cost," "what is the ROI of Y," "how big a Z do I need"), when the current market answers are gated, stale, or fuzzy, and when the brand has domain expertise that maps to a clear method. They also need engineering capacity for the build, though no-code platforms like Convertful, Outgrow, or Wufoo can substitute.

When calculators lose. Calculators struggle when buyer questions are qualitative, when the method can't be documented, when rules limit what the calculator can claim, or when the team plans to gate it behind a form (gated calculators are invisible to AI).

High-value calculator types. ROI calculators work out the return on investment for a category-relevant decision, comparison calculators compare options across multiple dimensions, assessment tools score the current state against a benchmark, and cost estimators estimate cost or sizing for a service or build.

What it takes: operator time for method design, design and UX work, an engineering build (or a no-code platform in its place), benchmark data sourcing, and promotion. Calculators move from method design to launch fairly quickly, and citation traction follows.

Design principles for citation. The method should be transparent, with the calculator showing its logic inline or through a dedicated method link. Each calculation should produce a unique result URL that returns that specific result, which I'd call the most important design call for AI citation. Results should be downloadable (PDF or CSV with the method built in), and each one should carry benchmark context showing how the user's result compares.

The gating dilemma. The most common calculator mistake is gating results behind an email form for lead generation; the form captures a meaningful but modest share of users as leads, but AI systems can't fill out forms, so a gated calculator earns zero AI citations and the lead capture comes at the cost of the citation. The pattern I'd use gives the basic result ungated (so AI systems can index and cite it) and offers an enhanced report (extra breakdowns, exportable formats, custom benchmarks) behind an optional email gate.

Common calculator mistakes. An opaque method is the first: calculators that show one number with no reasoning earn lower citation rates, so I'd show the formulas inline or through a method link. Missing unique result URLs is the second, because a calculator that shows results only on the original page, with no shareable URL per calculation, is a UI feature that nobody can cite as an asset. Stale benchmarks are the third; benchmark data, pricing assumptions, and reference rates change, and quarterly refresh is the working cadence.

Templates and Downloadable Assets

Templates earn citations because they give the audience something to take away and use, and process checklists, planning templates, scoring rubrics, and decision matrices all serve that role by turning expertise into a transferable artifact. AI systems surface templates because the queries ("what is the right structure for X," "is there a template for Y") are common and the answers available are often weak.

When templates win. Templates fit an operational category with recurring work the audience needs to run well (project management, hiring, sales ops, marketing ops, finance close), a brand with tribal knowledge it can codify into a transferable artifact, and a team with little production capacity beyond operator time, since templates convert effort into citations as efficiently as any asset type. They also suit an operator-grade audience more than an executive one.

When templates lose. Templates struggle in strategic categories (those reward frameworks and opinion), when the template would look too generic to stand apart, and when the brand can't commit to refresh cycles, because year-stamped templates need a yearly refresh.

Template categories. Process templates cover workflows, SOPs, and quality control; planning templates cover project roadmaps, sprint planning, content calendars, and campaign briefs; analysis templates cover SWOT matrices, competitive scorecards, and decision frameworks; and assessment templates cover audit checklists, maturity models, and scoring rubrics.

Format hierarchy. Google Sheets has the highest citation value because it's editable, collaborative, and easy to share, and the audience can copy and customize it without breaking the original. Excel is next (full-featured, offline, and familiar to enterprise audiences), and PDF has the lowest citation value of the three; it's viewable but not editable, so PDFs earn fewer citations because users can't use them without retyping.

What it takes: operator design, format conversion, a branded design header, and distribution through the operator network. Templates move from idea to publication fast, and citation traction begins reasonably quickly when they're promoted on community surfaces.

Common template mistakes. PDF-only is the first, since PDF templates earn far fewer citations than editable formats and the audience needs to copy and customize them directly. Generic titles are the second; "Project planning template" is generic, while "Project planning template for SaaS marketing teams" earns the citation when the user's query matches the specifier. No refresh cycle is the third, because year-stamped templates decay without a yearly refresh (and that refresh is a citation event of its own). Bloated structure is the last; templates with too many fields or too much text don't travel, because the audience wants a working tool and won't wade through a manual to get one.

Citation Surface Yield by Asset Type

Asset selection works best when the operator can compare how efficiently each asset type converts effort into citations. The Citation Surface Yield is a framework we coined at Searchbloom for that comparison, measured across mid-market engagements at an early mark and a later mark, and it puts the asset-type comparison on numbers where it would otherwise run on intuition. Per-asset yield tracks closely with the count of distinct, sourced insights each asset carries, which Chapter 5 measures as Information Gain Density.

Two patterns stand out. Opinion and templates lead early yield, research leads later yield once the data starts compounding, and frameworks sit between them with the best balance of early lift and durable compounding, which is why mid-market brands chasing fast traction default to opinion and templates while enterprise brands with patience default to research.

The CSY varies by category. Wikidata-dominant categories (per the Wills correlations) lift research's yield because Wikipedia inclusion drives outsized citation share, and SE-outbound-link-dominant categories lift template and calculator yield because broad listing surfaces reward utility-driven assets. I'd use the CSY as the starting comparison and adjust by category factor before locking in the asset choice.

The CSY also changes how much to make. Most teams plan one asset and put all their effort into it; I'd plan a mix of asset types across the year, because a spread covers more of the Citation Surface Yield curve than one large asset does (opinion and templates carry the early window, frameworks balance the middle, and research, where the brand can produce it, anchors the durable later yield), while a single asset can only sit at one point on that curve.

The Decision Framework

The asset-type choice rests on the Citation Surface Yield profile above plus 2 questions: what you can credibly produce (Expertise Fit) and what the category's cited content lacks (Competitive Landscape). I'd work both before picking, and the first one that rules an asset type out decides it. Most teams reverse this by starting with what feels rigorous or what the team already knows how to make, and working the 2 questions in the right order produces very different choices.

Expertise Fit

• Strong operator experience, weak data infrastructure. Frameworks and opinion.
• Strong domain expertise, no research operation. Calculators or templates, where the expertise becomes the method and the asset is the interface.
• Strong research capability, established analyst coverage in category. Data-driven research.
• Strong product or operational expertise, engineering capacity. Calculators that expose the logic of your domain.

Competitive Landscape

• Category dominated by one or two named frameworks. I'd skip building a competing framework and pick a different asset type.
• Category with no proprietary data sources. Data-driven research has high upside here, since the first credible study becomes the citation default.
• Category saturated with low-quality calculators. A calculator with a clear method and ungated result URLs displaces the incumbents.
• Category dominated by gated PDFs. Any ungated asset has unusual citation upside.

Conflict resolution: when expertise fit and competitive landscape pull in different directions, the question that rules an asset type out wins, so a team with no research capability rules out research even if the landscape favors it. I'd build what you can credibly produce and what the category lacks, because the 2 questions combine as an AND and an asset type has to clear both.

The Self-Published Entity Graph: An Emerging Asset

The 5 types above are things you write; this one is a thing you publish about yourself. Wikidata and Wikipedia put your entity into surfaces you influence but don't own (Chapter 10), and an emerging asset adds one you own outright, a machine-readable knowledge graph on your own domain that names your brand's entities, the relationships between them, and the facts, with every claim tied to a source URL on your site. It belongs in Evidence because it makes you the primary source of the structured facts AI systems repeat about you (so third parties aren't the ones setting them), and AI platforms that retrieve live can fetch the file directly when they build an answer about you.

The evidence is early, and I'd treat this as a bet until it's proven. Two independent groups arrived at the same idea across late 2025 and early 2026: a grounding-page standard (Hanns Kronenberg) and an open specification called EntityMap (Fred Laurent and Dixon Jones), which pairs a human-readable file with a machine-readable one and requires every entity claim to be a verbatim, source-attributed passage. EntityMap's own case studies report the published entity file drawing more retrieval than the same site's About page on Gemini and Perplexity; that's a vendor-run result on a single domain, and no AI provider has endorsed the format, so it sits in the same proposed-standard-without-provider-support category as llms.txt, with one difference (the early retrieval data here points up, where the llms.txt data is flat).

Where it's credible, the credibility comes from the same discipline as the rest of this chapter. The file only helps if every fact in it is true, canonically named, and sourced, and a machine-readable graph of vague or invented claims is worse than none, so I'd build it from verified facts, add the sameAs links to your Wikidata and profile URLs by hand (never guessed by a model), and keep it in sync with the canonical description your other surfaces carry.

The asset-selection logic from this chapter still applies. On expertise fit, any brand that can state a set of verified, sourced facts about itself and publish a file qualifies, which is a lower production bar than research or a calculator. On competitive fit, it does the most in categories where AI already describes you thinly or wrong, since that's where a canonical source file has the most to correct. Its citation profile is unproven and early (no provider has committed to reading it), so I'd treat it as infrastructure that compounds the Wikidata and schema work in Chapter 10 and skip pitching it as a standalone play. Searchbloom runs a version of this as an AI-information page, a structured, sourced facts-about-the-brand page, and we measure it with the AI-crawler log audit and citation tracking from Chapter 13, so the decision to keep investing rests on whether it earns retrieval on your property and the format's promise doesn't enter into it.

The next step on the same idea is exposing your brand data to AI agents through clean feeds and Model Context Protocol endpoints, so an agent queries your facts directly without having to scrape them. Today that serves agentic tool-use more than organic citation, and the payoff is unproven, but it's the direction the retrieval layer is moving, and it rewards what this chapter rewards: structured, accurate, machine-legible facts about your brand.

Platform-Specific Considerations

AI systems behave differently from one another, so I'd let the asset-type choice reflect which platforms matter most to the brand's audience.

• ChatGPT. Favors structured listicles, comparison tables, FAQs, and step-by-step content; AirOps March 2026 data found lists and tables in nearly 80% of ChatGPT citations vs 29% in Google's top results, so asset types with list-heavy structure (templates, frameworks with component breakdowns, calculators with comparison views) over-index.
• Claude. Weights academic citations and source diversity, so research with a documented method, opinion with a credible author entity, and frameworks with named author attribution all over-index.
• Perplexity. Puts heavy weight on Reddit, community discussion, and recent news, which is why opinion shared through community surfaces, reactive opinion on recent events, and frameworks discussed in active Reddit threads over-index.
• Gemini. Pulls mostly from Google-indexed content with strong SEO signals. I see AI search as an evolution of SEO (I don't treat it as a separate discipline), so assets that earn organic Google rankings also earn Gemini citations, and the same crawlable, authoritative work that drives a national SEO footprint feeds both.
• Google AI Overviews. 97% of AI Overviews cite at least one source from the top 20 organic results (seoClarity, October 2025), so asset types that rank well organically (research with backlinks, calculators with utility, in-depth frameworks with topical authority) over-index.
• Microsoft Copilot. Pulls mostly from LinkedIn, Microsoft-indexed enterprise sources, and Bing-ranked content, so opinion shared through LinkedIn over-indexes.

Most teams optimize for a portfolio, since the citation spread across platforms is wide enough that single-platform optimization is rare.

Industry Variants

Asset-type winners vary by industry, and Ben Wills's March 2026 research (145 industries, 1,595 personas, 105,000+ LLM prompts) surfaced industry-specific signal patterns that help guide asset selection.

• Wikidata-dominant categories. Accounting software, baby care brands, budget hotel chains, and CRM software, which reward research and frameworks because both produce the entity-level claims that get coded into Wikidata and Wikipedia.
• SE-outbound-link-dominant categories. Agricultural equipment, B2B marketing data providers, beauty and cosmetics retail, beer brands, and bottled water, where templates and calculators that get listed across third-party sites earn the reward.
• Wikipedia-citation-dominant categories. CRM software (rho=0.577), which rewards research and frameworks with citation density high enough to merit Wikipedia inclusion.
• Harmonic-centrality-dominant categories. Affiliate marketing networks (rho=0.577), auto insurance, and brokerage and wealth management apps, which reward research with downloadable data and tools with embed-friendly result URLs.
• Backlink-count-dominant categories. Car rental brands, where calculators and templates that earn organic backlinks through utility come out ahead.
• Best-search-rank-dominant categories. Most industries fall here, at moderate correlation, and research and frameworks that produce ranking-friendly canonical pages are the defaults, which is why building topical authority on the core subject still does most of the work.

For asset-type selection, I'd check your category against the Wills correlations, find the dominant signal type, and bias the asset choice toward the types that produce that signal.

The Selection Workshop

Asset selection is most reliable when the team runs a structured workshop for it, where a planning meeting tends to settle on the default. The workshop is a single focused session with the executive sponsor, the operator who will produce the content, and 1 or 2 context stakeholders.

Pre-workshop data to gather ahead of the session.

• The Citation Surface Yield profile for the asset types under consideration.
• Current AI citation baseline for the category, measured via Profound, Peec AI, or Semrush AI Toolkit.
• Competitive map: the top 3 competitors and the assets they get cited for.
• Wills industry data for your category from the LLM Ranking Factors research.
• Operator-network audit: named experts with the credibility to carry content.
• An honest read on what the team can credibly produce, including whether real research infrastructure exists.

The workshop agenda, in sequence.

1. Review. Go through the pre-workshop data and confirm what the team can credibly produce and what the category's cited content lacks.
2. Work the framework. Run the decision framework and rule out asset types the team can't credibly produce or that the competitive landscape closes off.
3. Build the mix. Draft each asset (topic, format, author); multiple candidate mixes are fine at this stage.
4. Pressure-test. Test each mix against 3 questions: does it span the Citation Surface Yield curve instead of clustering at one point, can the team credibly produce every asset in it, and does it address what the category's cited content lacks?
5. Pick. Choose the mix and write down the decision and the reasoning.
6. Assign. Name the owners and confirm the near-term execution plan.

Post-workshop deliverables, produced promptly. A one-page summary with the asset list and owners, a decision document with the Expertise Fit and Competitive Landscape analysis, a calendar of asset publish dates, and a measurement plan per Chapter 13.

Common Selection Mistakes

The Asset Refresh Cadence Calendar

Published assets decay without refresh, and the decay rate varies by asset type. The Asset Refresh Cadence Calendar gives operators a per-type schedule for when to update each asset, because an engagement running on a single quarterly refresh cycle wastes effort on assets that don't need it and skips refreshes on the ones that decay fast.

• Frameworks: the longest refresh interval. Vocabulary drift is the trigger; categories evolve their language until framework terms go stale or misleading. The refresh updates examples, adds new sub-components if the framework's scope expanded, and replaces case studies that no longer reflect how the category works now.
• Opinion: a moderate refresh interval for evergreen pieces. Counter-claim erosion is the trigger, since the contrarian view that was sharp early may become consensus later, and the refresh either re-sharpens the claim against the new consensus or pivots to a new contrarian angle. Reactive opinion pieces (published soon after a trigger event) don't refresh; they retire.
• Research: an annual full refresh. Data staleness is the trigger, because survey-based research stays fresh only for a limited window before journalists, analysts, and AI systems weight it less, so annual refresh is the working cadence. Sample-sized panel research can refresh on a quarterly micro-release cycle (1 new cut per quarter) without redoing the full study.
• Calculators: the shortest refresh interval. Benchmark data, pricing assumptions, and reference rates change, and if you skip the refresh the citation curve flattens.
• Templates: an annual refresh for year-stamped templates. Year-stamp decay is the trigger; a "Q1 2026 Marketing Plan Template" stops being useful at the end of 2026. The refresh adds the new year, updates field labels, refreshes example data, and bumps the "last updated" date, while templates without year stamps refresh on a longer cycle, similar to frameworks.

The calendar drives quarterly planning. At the start of each quarter I'd audit which assets are due for refresh against the cadence, schedule the operator and coordinator time, and treat refresh work as part of the Evidence engagement so it doesn't get pushed as discretionary work. Refresh velocity (the share of assets refreshed on time per quarter) is a tracking metric in its own right; the public Chime case study (September 2025) reported refresh velocity rising from 16 to 27 posts per month and a 3x citation increase on priority questions within 4 weeks. Engagements at 80%+ refresh velocity see steady citation share, and engagements below 60% see citation share decline under consistent under-refresh.

The Asset Retirement Decision Framework

Not every asset deserves a refresh, and some should retire. The Asset Retirement Decision Framework uses 4 signals to separate a refresh-and-extend asset from a retire-and-replace one; I'd keep refreshing the assets that pass and retire the ones that fail, because refreshing a retire-grade asset spends operator time that should go to new asset development.

• Signal 1: Citation share trend. The asset's primary topic citation share dropped 30% or more over a sustained trailing window, a drop that reflects category evolution the asset can't recover from through refresh alone.
• Signal 2: Co-citation velocity decay. The Co-Citation Velocity Score per Chapter 3's framework has fallen below 0.5 of the asset's baseline, which means third parties no longer reference the asset at the rate that drives compounding.
• Signal 3: Competitive displacement. A competitor recently published an asset with materially better data, method, or format, the displacement shows in AI citation patterns, and the competing asset now holds the citation slot the original used to hold.
• Signal 4: Operator credibility gap. The operator can no longer defend the asset's claims because the category has evolved; common cases are a changed data assumption, framework components that no longer match how the category works, or an operator whose view has matured past the asset's original framing.

The decision logic:

• Three or four signals triggered. Retire the asset and replace it with a new one built on a refreshed approach, and if AI systems are still citing the original URL, 301 redirect it to the replacement.
• Two signals triggered. Run a deep refresh attempt that adds a new section addressing the strongest of the triggered signals, then re-measure once the refresh has had time to register; if at least 1 signal flipped to positive, keep maintaining the asset, and if none did, retire it.
• One signal triggered. Monitor it, and address that specific signal in the next scheduled refresh.
• Zero signals triggered. The asset is healthy, so it stays on the standard refresh cadence.

Most engagements over-maintain underperforming assets, and the sunk-cost effect is strongest with assets the operator built personally. The retirement framework turns that decision into a data-driven check, which frees operator time for new asset development (it has a higher expected citation yield than refreshing a stalled asset).

Questions & Answers

How do I choose between a framework and a calculator if the team can produce both? Build the framework first. A framework earns citations by becoming vocabulary that travels by reference, a calculator earns them by becoming a utility with its own URLs, and the framework sets the category vocabulary that the calculator then puts to use.

What if our category does not have established asset types? I'd default to expert opinion and frameworks. Categories without set asset types are categories where the vocabulary is missing, and opinion and frameworks both supply vocabulary, while research and calculators need category context that doesn't exist yet.

Can we mix asset types in one launch? Yes, and a mix is usually better than betting everything on one large asset. Opinion, frameworks, calculators, and templates together cover more of the Citation Surface Yield curve than a single research study does, so the mix produces more citation surface area for the same effort.

Does the right asset type change if we are optimizing for ChatGPT vs Perplexity vs AIO? Yes. ChatGPT favors structured listicles, Perplexity weights Reddit, AIO leans on the top-20 organic results, and Copilot pulls from LinkedIn; the citation spread across platforms is wide enough that most teams optimize for a portfolio and single-platform optimization is rare.

How long until I see citation traction after publishing my first original asset? Opinion earns traction fastest, followed by templates, then frameworks and calculators, with research the slowest, and most of that variance comes down to co-citation velocity.

Should we gate our research, calculator, or framework? No, with one exception. AI systems can't fill out forms, so gated content is invisible to AI Search; the exception is an enhanced version of an ungated base asset, where I'd give the basic results ungated and gate the downloadable enhanced report.

What is the smallest viable way to start an Evidence engagement? I'd start with the asset types that convert effort into citations most efficiently. A working starter cadence is 1 opinion piece per month plus 1 framework or template per quarter, all of which a credible operator can produce without research infrastructure, and consistency over several quarters matters more than the size of any single asset.

Can opinion pieces really compete with research for citations? Yes, when the operator's credibility is established and the claim is specific. Our counter-claim at Searchbloom, drawn from partner engagements, is that AI systems cite opinion-based content at rates close to empirical research when the source has verified expertise and the claim is specific enough to pull out as a discrete statement.