Citation Reinforcement and Topical Clusters for AI SEO

Chapter 4 made the asset and Chapter 5 explained why it gets cited; this chapter is about making those citations compound. A single high-information-gain asset earns citations in a single topical area, while a group of 5 to 10 related assets in the same area earns citations at 3 to 5 times the per-asset rate, because each new asset builds on the topic signal the earlier ones set. I break the work into 4 layers (cluster design, citation pattern review, refresh cycles, and attribution networks), and together they turn one-off Evidence wins into citations that keep compounding.

Why This Technique Matters

The default Evidence motion is shaped like a project: a team orders a research study, publishes it, promotes it for a short window, and moves on to the next one. The asset earns citations for a while, then slows down and drops out of active retrieval, and the next study repeats the cycle, so the brand piles up decaying citation curves and never builds an asset base that compounds.

Citation reinforcement treats Evidence as ongoing work. Each asset is built to back up the earlier assets in its area and to set up the ones that will join the cluster next, which makes the cluster as a whole the unit you measure (the single asset stops being the unit). What the brand earns is entity-level authority on a defined set of topics, where the project approach spreads the same effort across unrelated content and gets scattered visibility for it.

We see the compounding math in our own work. Across 40-plus Searchbloom partner engagements, the 3rd and 4th assets a brand publishes on the same topic earn citations at much higher rates than the 1st, because by then the model has enough signal to retrieve the brand for queries in that area; the first asset fights a cold-start problem, and the later ones benefit from the topic association it started to build.

The competitive impact is large, and it's easiest to see with 2 brands that each publish 1 solid asset per quarter on a sustained run. The first ends up with 8 scattered assets; the second puts the same 8 into a single topic area and has a cluster a late mover can't easily displace (and the cost of displacing it keeps rising as the cluster matures, because a new entrant has to build a comparable cluster plus the third-party citation density the leader spent a long time gathering).

The Depth-Before-Breadth Principle

Topic depth earns AI citations at multiples of topic breadth from the same total content effort. The mechanism behind that is entity-aware retrieval: AI systems keep track of which domains cover which topics, so 1 asset on a topic sets a thin signal (the model retrieves it on its own merit when it happens to be relevant), while 5 assets on the same topic set a brand-as-source signal and the model reaches for any of those 5 first, since the brand has shown depth there. It's the same reason topical authority has long outperformed scattered keyword coverage in organic search.

This leads to a call that feels backwards at first. If a team has the capacity for 10 Evidence assets, it can spread them across 10 topic areas (option (a)) or group them into 2 clusters of 5 (option (b)), and I'd usually take (b), because (a) makes 10 weak signals that each fall below the cutoff for topic recognition while (b) makes 2 strong signals that each clear it. Same total content volume; the citation outcome differs by a factor of 3 to 5.

Multi-product brands that serve distinct buyer groups are the exception. Think of a platform company with 3 product lines sold to 3 buyer groups: collapsing to 1 topic cluster would leave 2 of those product lines invisible, so what I'd build there is parallel clusters, 1 mature cluster per product line, ordered by which product needs AI visibility most. 3 parallel clusters of 5 assets each takes several cycles to build. 3 scattered portfolios of 5 assets each is an engagement that stays below the cutoff.

Topical Cluster Architecture

A topical cluster is 1 hub asset plus 4 to 9 spoke assets, all linked and each covering a different part of the same topic: the hub is the main reference page, and the spokes go deep on specific sub-areas. It's the same hub-and-spoke pattern a national SEO engagement has used to build topic authority since the early 2010s, now applied to AI retrieval instead of PageRank flow.

The Hub Asset

The hub is the strongest standalone asset in the cluster; it sets the vocabulary, defines the scope, and links to the spokes that go deeper. It covers each part at survey depth (full depth on every part is the spokes' job) and routes readers and the model to the spoke that owns that part in detail, because a hub that tries to be complete gets too long to retrieve cleanly, and a hub that's too thin fails to set the topic claim.

Hub content typically lands at 3,000 to 5,000 words and covers 6 to 8 major sub-areas with 2 to 4 paragraphs each. The hub carries the main schema for the topic, the top stats and frameworks, and the cleanest summary of the brand's view; spokes inherit context from it, so I'd route readers who land on a spoke first back to the hub for the fuller treatment.

The Spoke Assets

Spokes go deep on specific sub-areas of the topic. Each one is a high-information-gain asset on its own merit (Chapter 5 has the build pattern) that also sits inside a topic area the brand has committed to. Spokes typically land at 2,500 to 7,000 words depending on the size of the sub-area, and the depth target is whatever the spoke needs to become the main reference on its sub-area, so I wouldn't hold it to a fixed word count.

Spokes link to the hub and to each other where the sub-areas overlap. Inside a cluster, the link graph matters less for AI retrieval than it does for organic search, but it matters more for human readers and for the cross-citation patterns analysts follow when they reference the brand; a human auditor should come away seeing one body of work, and a weaker build reads as isolated pages that happen to share a topic.

The 5-to-10 Asset Rule

Reinforcement starts at 5 total assets, which is 1 hub plus 4 spokes. Under 5, the topic association is there but too thin to beat the model's bias toward more established sources; over 10, each new spoke adds less because the cluster is already set, and I'd move that effort into a second cluster or into attribution-network depth.

I'd build in order, one asset after another, instead of all at once. The hub goes first (the strongest standalone asset the brand can produce), then I'd watch citation patterns for a while, see which sub-areas are actually drawing queries, and order the spokes against those, since spokes built against real query traffic compound faster than spokes built against guessed traffic.

The Spoke Build Order Algorithm

Building against the query branches you see is the right rule, but it's vague. The Spoke Build Order Algorithm turns it into something you can calculate: every candidate spoke gets a 0 to 100 score across 4 criteria, and spokes get built from the highest score down. In the engagements we've measured, that ordering has compounded clusters much faster than picking spokes ad hoc.

• Traffic Potential (TP, 0 to 30 points). Current citation traffic from AI systems to the sub-area, plus search volume for the query family behind it. A high-volume sub-area with an active AI citation rate gets 25 to 30; a low-volume one with no citation activity yet gets 0 to 5.
• Substance Availability (SA, 0 to 25 points). Can the brand promptly publish an asset that clears a strong Information Gain Score here? Strong proprietary data plus operator depth on the sub-area scores 20 to 25, and weak substance with no clear IG technique stack scores 0 to 5 (Chapter 5 has the IGD Technique Stack patterns).
• Competitive Whitespace (CW, 0 to 25 points). How thin competitor coverage is. If no competitor cluster has more than 1 asset on the sub-area, it scores 20 to 25; if a competitor has already built 5+ assets at A-grade IGS, it scores 0 to 5.
• Hub Linkage Strength (HL, 0 to 20 points). How cleanly the sub-area maps to the framework the hub already has. Sub-areas the hub references explicitly get 15 to 20, and anything that would mean extending the hub's scope gets 0 to 5.

Add them up (TP + SA + CW + HL) and each candidate spoke lands somewhere from 0 to 100. I'd build the top one first and re-score the rest after every publication, because hub linkage shifts as new spokes go live, and running the algorithm on a regular cadence also catches sub-areas whose traffic potential has climbed since the last round.

Experienced operators do this ordering by instinct, and the algorithm writes that instinct down for the many engagements that don't have an experienced operator (it also keeps experienced operators honest). Pet-topic bias is the most common reason spokes get built out of order; the score makes it obvious when the spoke slated next isn't actually the highest-scoring move.

Worked Example: Searchbloom's Information Gain Cluster

The Cluster Citation Density Score

The 5-to-10 asset rule tells you when a cluster reaches reinforcement, but it doesn't tell you how well a given cluster is actually compounding. With the same asset count, 2 clusters can produce very different citation outcomes depending on how well the assets back each other up (in our engagements that spread has been wide), so we coined the Cluster Citation Density Score, a Searchbloom diagnostic that puts the compounding effect into a single number.

CCDS = (total citations across all cluster assets) / (estimated citations the hub asset alone would earn standalone)

I'd calculate the denominator from the hub's citation count for the period when it was the only asset in the cluster, or estimate it from comparable single-asset benchmarks in the category. Here's how I read the bands:

• CCDS above 4. Strong reinforcement: the cluster is operating as one entity, and the spokes are lifting the hub's standalone citation rate on top of adding their own citations, which tells you the model recognizes the brand as the source on the topic.
• CCDS 2 to 4. Moderate reinforcement, where the cluster produces additive lift but the reinforcement effect is only partial. I'd look for weak linking between assets, mismatched named-author bylines, or spokes with low information gain.
• CCDS 1 to 2. Weak reinforcement: the cluster is mostly a collection of standalone assets, and the reinforcement effect hasn't engaged yet. The usual causes are spokes that drift off-topic from the hub, a hub too thin to anchor the cluster, or gaps in the attribution network.
• CCDS below 1. Degenerative. The cluster is doing worse than the hub would on its own, either because weak spokes drag the hub down through cross-linking or because the spokes compete with the hub for the same citation slot.

I'd track CCDS every quarter, per cluster, and let the number set the next move: at 4+ the engagement keeps building inside the cluster (more spokes, deeper attribution), at 2 to 4 it fixes the reinforcement gaps before adding anything, and below 2 it audits the cluster architecture first, because new assets added to a cluster under 2 pick up the same per-asset drag.

At year-boundary reviews I'd pair the CCDS with the Cluster Audit Workflow. The workflow says what to merge or build; the CCDS says whether the existing cluster is compounding at all before any new investment lands on it.

Citation Pattern Identification

Building a cluster without citation pattern data is guesswork (publish, link, promote, and hope it compounds). I'd lead with the data instead: find out which assets are getting cited and for which queries, then order the next assets so they back up what's already working. 3 signal types do most of that work.

Hub-Candidate Signals

When an asset gathers citations across many distinct queries, including queries the brand may not have targeted, it's a hub candidate; that breadth of retrieval is the model signaling that this asset is the right main reference for the topic. When the pattern shows up, I'd treat the asset as the cluster hub and build spokes to match the query branches appearing around it.

I'd run detection monthly: pull the top-cited assets across every tracked AI platform, pull the queries behind each one's citations, and group them by topic. An asset cited across 4 or more distinct queries is a hub candidate, and it becomes the anchor for a planned cluster build over the next few cycles.

Spoke-Opportunity Signals

A spoke opportunity looks different: a topic that earns steady citation traffic without having its own asset. The brand gets cited next to the topic, probably through the hub or some related asset, so the model is grabbing a thinner anchor than it should, and a dedicated spoke built against that traffic often lifts citation share fairly quickly (it gives the model a stronger anchor to retrieve).

Detection here is monthly too. What I'd look for are queries where the brand earns citations but the cited URL is something generic, like the homepage, a marketing page, or a thin blog post; every one of those is a spoke opportunity, and the traffic each query produces sets the build order.

Attribution-Gap Signals

Attribution gaps show up when an asset earns citations only alongside a third-party reference. The brand gets named next to another source (an analyst report, an industry publication) and the asset gets cited in that pairing, but not on its own merit, and the fix is named-author bylines and third-party syndication; more original content won't close that gap.

You can spot these assets by their shape: real citation rates, low standalone retrieval, cited in compound queries and missing from topic-only ones. They need attribution-network work, which comes later in this chapter.

Refresh Cadences

Cited content decays if nobody refreshes it. AirOps's August 2025 analysis of 4,000+ ChatGPT-cited pages across 900 high-intent queries found that pages not updated within the past year are more than twice as likely to lose citations. The reason is recency weighting in AI retrieval (when 2 candidates carry similar information gain, the more recently updated one shows up first), which makes refresh the upkeep that holds citation share against the citation half-life wearing down every cited asset over time.

Most assets fit 1 of 3 refresh cycles.

Quarterly Refresh: Benchmarks, Statistics, Time-Sensitive Data

I'd refresh quarterly anything whose main value is a number that ages: benchmark studies, pricing comparisons, performance metrics, market share figures, and rule updates. The refresh updates the numbers, stamps a new date, and republishes, so the asset picks up a recency boost and its figures stay credible to the analyst tier whose co-citations drive AI visibility.

To run it, I'd keep a quarterly editorial calendar listing every asset in that band with a named owner. Per asset it's light work (pull the data, update the narrative and the schema, send the IndexNow notification from Chapter 12), and that effort is small next to the citation lift.

Annual Refresh: Frameworks, Opinion, Evergreen Reference

Frameworks, opinion pieces, methods, and evergreen reference content age more slowly than time-sensitive data, so a yearly pass is enough: new examples, newer references, a section on what has changed since the first version, and a fresh datestamp. The substance holds while the framing gets current, and it's a moderate effort per asset.

We run the MERIT whitepaper on this same cycle. The October 2025 version got an April 2026 update (a fresh round of new research, more case studies, refined positioning), and the 2027 update will follow the same pattern, because a framework that skips its annual refresh starts to look dated even when the core logic is still sound.

Reactive Refresh: Triggered by External Events

Some refreshes are triggered by outside events instead of the calendar. A platform change counts (say, a new ChatGPT release that changes citation behavior), so does a rule shift like an FTC ruling that affects category claims, and so does big news, such as a rival announcement that resets the category. I'd update the affected assets promptly after the event, and of the 3 patterns this one gets the strongest recency boost, because the model is actively retrieving against that event.

The mechanics are a watch list of outside triggers that the team tracks, each one mapped to the assets it affects and to the editor who owns the refresh. Reactive refresh pays off most during platform release cycles, news on rule changes, and announcements that set the category.

Attribution Networks

Citations compound through attribution networks at the entity level. When many assets share the same named author, AI systems learn to link the topic with that author entity, and when the author also shows up in third-party sources, the network compounds further; the best Evidence engagements build a planned attribution network alongside the content cluster itself.

Named-Author Discipline

I'd put a named-author byline on every asset in a cluster, pointing to a real person with proven expertise, because brand-byline assets ("by The Searchbloom Team") earn fewer citations than named-author assets ("by Cody C. Jensen, CEO & Founder of Searchbloom"). AI systems weight author-entity authority alongside organization-entity authority, and the compounding effect is large: an attribution network with 1 expert author makes a citation curve that a second expert author can copy (though the second author can't share the first one's curve), so multiple expert authors in the same cluster grow the network without diluting it.

The Person schema (covered in Chapter 10) attaches the author entity to the asset in code, and the author's standalone main page (a bio on the brand domain, a LinkedIn profile, or both) acts as the entity anchor. Author bios that match across every asset in the cluster back up the entity signal.

Third-Party Attribution

The best attribution networks include third-party references, meaning the brand's expert gets quoted by outside sources, writes for outside publications, and gets cited in outside research. Industry contributed pieces, podcast appearances, analyst report mentions, and academic citations all count, and each outside reference backs up the topic association on surfaces the model retrieves from beyond the brand's own domain.

This ties back to the Mentions pillar work in Chapters 1 through 3. Third-party support is part of cluster building (the layer that turns citations of owned assets into broader entity authority), and engagements that skip it hit the ceiling of owned-domain citation share before long, while engagements that fold it in keep compounding.

Co-Citation Velocity

Co-citation velocity is the metric I'd use to track attribution-network health. It measures how fast third-party sources start referencing the brand's work after publication, and it explains why a high-IGS asset with weak co-citation velocity earns spotty citations while the same asset with strong velocity earns steady citation share: the model retrieves the asset alongside the third-party references that back it up.

To measure it, I'd track the time between an asset's publication and its first 3 independent third-party references. A fast pickup is strong and a moderate lag is typical, while a slow pickup points to a distribution problem more than a content problem; it usually means the third-party outreach layer is under-funded, and the content underneath may still be good.

Building the Attribution Network: A Phased Plan

The build pattern that holds up runs in sequenced phases and assumes 1 named expert committed to publishing on a steady cycle (often the founder, the CEO, or a senior SME). Brands with several possible expert voices can run parallel tracks against the same plan.

Phase 1: Foundation. I'd build the expert's main bio page on the brand domain, rich in schema and linking out to LinkedIn and any speaker profiles, then set a monthly opinion publishing cycle on the owned domain and build out the expert's LinkedIn presence (a connection to the company page, regular long-form posts, and planned cross-links to owned content). Outcome by the end of the phase: 3 owned-domain published pieces, 6 to 12 LinkedIn long-form posts, and Person schema across the whole content cluster.

Phase 2: First external references. I'd pitch the expert for contributed pieces in 4 to 6 industry publications that match the topic cluster, pitch for 3 to 5 podcast appearances, and submit the expert's opinion to 2 to 3 industry roundup articles, aiming for a first wave of third-party references that link back to the cluster. Outcome by the end of the phase: 2 to 3 contributed pieces published, 1 to 2 podcasts recorded, and the named expert appearing in third-party sources for the first time.

Phase 3: Densification. This phase reinvests in the channels that show traction: contributed pieces in publications that drove strong inbound interest get follow-ups, podcast appearances that turned into citation traffic get repeated, and new publications get added based on tier-up patterns. Outcome by the end of the phase: total external references in the 8 to 12 range, with citation velocity for cluster assets speeding up in a way you can measure.

Phase 4: Compounding signals. The network reaches the point where third-party sources reference the expert without being asked, so industry roundups include the expert by default, analyst reports cite the expert's framework or research without outreach, and new cluster assets earn co-citation soon after they publish. Outcome at the end of the build: a mature attribution network that the next clusters can be built on top of, with no fresh foundation needed.

The pattern repeats for each added expert voice. A second expert added later follows the same phased build and compounds faster, because the brand entity is already in place and the second expert inherits some of its authority; as the engagement matures, the brand often runs with 3 to 4 publishing experts whose attribution networks overlap and back each other up.

The Cluster Maturity Curve

Clusters move through 4 maturity stages you can tell apart, each with measurable signals and its own build pattern. Engagements that recognize their stage match the work to it, and engagements that don't tend to run a default pattern that fits one stage and fails the others.

• Stage 1: Foundation. The hub is published and 1 or 2 spokes are live, citation share for cluster queries sits at baseline or slightly above, and co-citation velocity is slow (third-party references take a while to arrive). The work pattern is to publish the hub at full strength, publish 2 spokes against the highest-confidence sub-areas, and start the attribution-network outreach; I wouldn't measure citation lift yet, because the baseline is still forming.
• Stage 2: Reinforcement. The cluster crosses the 5-asset threshold and the first measurable citation lift appears, with CCDS climbing from 1.0 toward 2.0 and co-citation velocity picking up to a moderate pace. The work pattern is to build spokes 3 to 5 against the query patterns observed in the foundation stage, run the first attribution-network outreach milestones from Chapter 3, and publish the first quarterly refresh on the hub.
• Stage 3: Compounding. The cluster reaches 7 to 10 assets, CCDS climbs into the 2.5 to 4 range, co-citation velocity is fast for new spokes, and citation share sits at multiples of baseline, with the model retrieving the brand for queries the cluster didn't directly target. The work pattern is to build the final spokes against late-emerging query patterns, deepen the attribution network with second-tier outlets, and start planning the cluster of clusters if a second topic area is in scope.
• Stage 4: Mature Defense. Full size, typically 8 to 12 assets including the hub, with CCDS stable above 3 and citation share leading the category. From here the job is defense and upkeep more than building (refresh cycles on schedule, competitive monitoring that catches rival cluster builds early), and a new spoke gets added only if a new sub-area has emerged with a strong signal.

Getting the stage right matters because each stage's work pattern is wrong for the others. Put Stage 1 work (foundation focus, no citation measurement) on a Stage 3 cluster and you waste the compounding window; put Stage 4 work (defense focus, no new spokes) on a Stage 2 cluster and the build stalls before it crosses reinforcement. Engagements that match the work to the stage reach Stage 3 quickly, and the ones that don't often stall in Stage 1 or Stage 2.

The curve also helps with executive sponsors, who need a reason to be patient. Stage 1 shows no measurable citation lift, and stakeholders can easily read that as a failed engagement, so I'd name the stage and the expected timeline upfront. The visible payoff arrives in Stage 3 (stakeholders who pulled support at Stage 1 never get to see it).

Cluster Maintenance: The Permanent Commitment

A mature cluster needs ongoing upkeep, the kind a one-off project can't supply: refresh work, defense against rival clusters, and new spokes as the topic shifts. Engagements that build once and walk away watch citation share decay after the cluster has sat for a while in a post-build steady state.

Quarterly upkeep review. Each quarter I'd pull citation share per cluster, broken out by hub and by spoke, and flag any asset that dropped more than 20% quarter-over-quarter. Every flagged asset gets an audit for the cause (out-of-date stats, rival displacement, a shift in query intent, structural decay) and a fix on the schedule soon after.

Competitive cluster monitoring. I'd track when rivals publish new clusters in the same topic area. New rival clusters take time to compound, and that build period is the defense window, when I'd back up the established cluster with 1 or 2 new high-IGS spokes covering the parts the rival cluster is staking out; that takes far less work than letting the rival displace you and then rebuilding share.

Topic-evolution response. Topics shift: new sub-areas emerge, old ones fade, and the vocabulary changes. Bi-annual topic reviews catch those changes and trigger a new spoke, a merge of a retired spoke, or edits to the hub, and topics that move fast (any AI-adjacent topic in 2026) need quarterly evolution reviews instead.

Attribution-network defense. Named experts hold their network value only while they stay visible, and an expert who stopped publishing a while ago adds less to the network than at their peak. Upkeep includes keeping each expert publishing on a steady cycle after the first build is done, which is light, ongoing effort per expert.

Mature-cluster upkeep costs a fraction of the original build effort each year. Engagements that plan for it keep steady citation curves, engagements that don't get decaying curves, and once decay sets in, the first build stops mattering.

The Cluster Audit Workflow

An engagement that inherits an existing content library needs an audit before the cluster build can start. The audit produces 3 outputs (a cluster map of existing assets, a merge plan for thin or duplicated content, and an ordered build plan for new spokes), and it's a contained effort for libraries under 100 pages that takes longer for bigger ones.

Step 1: Inventory and bucket. I'd pull every published asset and note its main topic, word count, last refresh date, and citation data if you track it, then bucket the assets into candidate topic areas (3 to 8 buckets often emerge on their own). Any assets that resist bucketing are worth noting, since they're often the scattered tail that drags the engagement down.

Step 2: Identify hub candidates. In each bucket, the strongest standalone asset becomes the cluster hub candidate, with strength coming from information gain density, citation traction, page depth, structure quality, and a named-author byline. A bucket without a strong asset needs a hub built from scratch, because spokes can't compound on a missing hub.

Step 3: Flag merge chances. Within each bucket I'd look for 3 kinds of asset: duplicates (several posts on the same sub-area), thin pages (under 1,500 words on topics that need depth), and pages that have gone a long time without a refresh. The substance worth keeping gets merged into the main asset, and the merged URLs get a 301-redirect.

Step 4: Identify spoke gaps. A spoke gap is a sub-area the hub references without covering it in depth. I'd check each gap against citation patterns (is the cluster earning citations there with no dedicated asset?) and against rival coverage (does the strongest rival cluster cover it when yours doesn't?); gaps that show both a citation signal and a rival presence go first.

Step 5: Order the build. What comes out is a sequenced build calendar: the cluster hubs to merge, the spokes in build order, the refresh cycles to set up, and the named-author bylines to develop. Each cluster's spoke build runs in stages, starting with the hub merge and the first 3 spokes.

Step 6: Set measurement. Before the build starts, I'd set a citation share baseline for each cluster and track it monthly. Share should lift early in the build; if it hasn't, the spokes are too thin, and I'd revisit the build pattern before stacking more spokes on a foundation that isn't working.

Skipping the audit and building fresh clusters on top of a scattered library gets mixed results. The new clusters do compound, but the old scattered content keeps diluting the topic signals (the model still retrieves those old pages), so in the cases we've measured, audit-and-merge has usually been the higher-impact pattern.

Cluster-of-Clusters: The Macro Pattern

Mature Evidence engagements end up building clusters of clusters. A single cluster covers 1 topic area with 5 to 10 assets; a cluster of clusters covers a broader category with 3 to 5 clusters, each anchored in a related topic area, and earns entity-level authority on the category as a whole.

The Searchbloom example: the Information Gain cluster (5 assets) sits inside the Corpus Engineering super-cluster (6 components, 1 of which is information gain), and the Corpus Engineering super-cluster sits inside the broader AI SEO category, where how Corpus Engineering relates to Relevance Engineering sets where retrieval optimization fits inside that broader discipline. Searchbloom is building authority around that category across 3 nesting levels (topic, super-topic, category) that each back up the others, and AI systems retrieve the brand for category-level queries because the lower-level clusters have shown the brand covers the parts in depth.

The cluster-of-clusters pattern is the long-range strategic view for serious Evidence engagements. I'd build it in sequence (1 cluster to maturity, then the second, then the third along with the explicit super-topic hub that links them), and from the later stages on, the compounding adds up to entity-level dominance that small rivals can't match without the same multi-year investment.

Platform-Specific Considerations

Cluster reinforcement compounds across platforms, though each platform weights the signals in its own way.

Industry Variants

Ben Wills's March 2026 research showed that the compounding mechanism reacts to industry-specific signal patterns, so cluster work pays off most in categories where entity-topic association is the main retrieval signal and less in categories where other signal types drive retrieval.

• Wikidata-dominant categories (accounting software, CRM software). Clusters compound through entity-level signals that flow into Wikidata, and each spoke that produces a verifiable, attributable claim adds to the entity record.
• Wikipedia-citation-dominant categories (CRM software). Clusters compound through Wikipedia inclusion, so the cluster needs at least 1 asset with enough analyst pickup to clear Wikipedia's inclusion cutoffs; once 1 cluster asset is cited in Wikipedia, the cluster's overall retrieval rate lifts a lot.
• Harmonic-centrality-dominant categories (affiliate marketing networks, auto insurance). Here clusters compound through embeddable assets and tools; a spoke that produces embeddable result URLs (calculators, dashboards) feeds the harmonic centrality signal.
• SE-outbound-link-dominant categories (beauty and cosmetics retail, beer brands). Templates, listicle placement, and widely referenced assets are what compound in these categories, so I'd lean the cluster toward shareable items over single deep references.
• Backlink-count-dominant categories (car rental brands). Earned backlinks drive compounding here, which means every spoke needs a hook that earns backlinks (data, a tool, a framework) to feed that main signal.

Common Mistakes That Defeat Citation Reinforcement

1. Scattering assets across unrelated topics. The most common failure mode. Every asset is well made, none of them back each other up, and together they produce no compounding. Counter-test: can the team name the 2 or 3 topic clusters the brand is building, with each spoke mapped to one?

2. Skipping the hub. Teams publish spokes with no main hub, assuming the topic is obvious; the model still needs an explicit hub to anchor the topic association, and without one a cluster compounds at half the rate. Counter-test: for each cluster, can the team point to the main hub URL?

3. Treating clusters as projects rather than ongoing work. The planned 5 to 10 assets go out, the cluster gets called done, and citations decay because nothing follows them (no refresh, no new spokes, no attribution-network growth). Counter-test: is there a long-range upkeep plan for the cluster covering refresh cycles, new-spoke triggers, and attribution-network growth?

4. Brand-byline attribution. When every asset in a cluster says "the team," the attribution network stays thin, since the brand is the only entity the model has to attach the work to. Counter-test: does the cluster have at least 1 steady named-author byline tying the assets together?

5. Refresh failure. Cited content decays without a refresh, and engagements that publish without a refresh routine watch their cluster's citation share drop once the assets go stale. Counter-test: is every asset on a quarterly, annual, or reactive refresh cycle with a named owner?

6. Building the wrong cluster. Teams sometimes build a cluster around a topic the model has no demand for, and since nobody is querying it, the cluster compounds inside the site but earns no citations. Counter-test: before the build, can the team show real query volume and current citation activity in the topic area?

7. Ignoring third-party backup. Owned-domain clusters hit a ceiling without outside attribution, and engagements that ignore Chapters 1 to 3 (the Mentions pillar) see their cluster's citation share plateau before long, whatever the publishing volume. Counter-test: is the cluster paired with an active third-party distribution motion?

8. Premature breadth. Teams start a second cluster before the first has matured, and the second pulls effort and focus from the first until neither compounds and both stay below the cutoff. Counter-test: has the first cluster hit the 5-asset reinforcement cutoff and shown citation lift before the second cluster build starts?

Questions & Answers

Why does topical depth beat topical breadth for AI citations? Topical depth beats breadth because AI retrieval is entity-aware: a single asset gets cited on its own merit, and later assets on the same topic get cited at higher rates because the topic association is already in place. In Searchbloom's engagements, depth has compounded roughly 3 to 5x faster than scattered breadth.

What is the 5-to-10 asset rule? The 5-to-10 asset rule says a cluster reaches reinforcement at 5 total assets (1 hub plus 4 spokes). Below 5 the signal is too thin, and above 10 the next asset adds less, so 1 hub plus 4 to 9 spokes is the working range.

How do I find which assets are being cited so I can back them up? I'd track citations with Profound, Peec AI, or Semrush AI Toolkit and watch for 3 patterns: hub candidates (assets cited across many queries), spoke opportunities (topics with citation traffic but no dedicated asset), and attribution-gap signals (assets cited only when paired with third-party references).

What is the right refresh cycle for cited content? The right refresh cycle for cited content is quarterly for benchmarks and stats, annual for frameworks and evergreen reference, and reactive for content hit by outside events. AirOps's August 2025 analysis found that pages not updated within the past year are more than twice as likely to lose citations.

Should I delete or merge old content as part of a cluster build? Yes, I'd delete or merge old content when it has low information gain or sits outside the cluster, merging the substance worth keeping into the main asset and 301-redirecting the old URL. The merged page builds up the share that was spread across weaker pages.

How does an attribution network work and why does it matter? An attribution network is the set of named-author bylines and entity references compounding around a brand, and it matters because that compounding multiplies instead of simply adding up. Brands with many expert authors build networks several times denser than brands with 1.

How long does it take a cluster to compound? A cluster's first reinforcement effects come after sustained work, and full compounding takes longer still. Co-citation velocity (how fast third parties start referencing the work) drives that curve more than the publishing cycle does.

Is there a point at which a cluster is finished? No, a cluster is never finished, though most reach a steady state at 10 to 15 assets, after which the work shifts from building to upkeep. Treating clusters as ongoing work instead of projects is what separates steady citation share from decaying curves.