The Phantom Wing

This site’s floor plan says it has a wing that the building does not contain. Walk every hallway of its link graph, and the hallways are good: real sections, working cross-links, sensible depth. You will still pass through fewer than three fifths of the pages the site publishes. The other two fifths, 140 pages, are listed in the sitemap, live, and rendering, with zero inbound editorial links: no in-content path pointing into them. This episode is about the gap between what a site publishes and what its structure admits to.

The Subject: Cloister

We’ll call this site Cloister, a corporate property of 343 pages joined by roughly 1,300 internal editorial links, organized into 10 communities with an average path length of 2.5 clicks. What makes its shape new for this series is the contrast at its center: the connected part of Cloister is one of the healthier structures we have measured, and it sits beside the largest proportional void we have seen on a site this well-built. Like its namesake, a cloister is a beautiful enclosed garden; the architecture defines it by what it walls out.

The default map looks almost fine. You see colored neighborhoods, a layered core, bridges between sections: eight readable clusters plus two small fragments (seven pages and two) that are easy to miss at the edge. The phantom wing is not a separate gray cloud outside those colors. Most of the 140 orphan pages still have outbound links, so Louvain paints them into the same neighborhoods as the pages they point at. They look like ordinary members of the graph until you ask the right question: who has zero inbound editorial links? Click Orphans under the map to light those 140 pages in red. That is the wing the floor plan declares and the hallways never reach.

Interactive: zoom, pan, and explore the anonymized topology

OrphansHubsBridgesSilosIslandsReset

The Insight

We call this pattern The Phantom Wing. The sitemap (the site’s own published floor plan) declares a wing the link structure does not contain: 41% of the whole building. A visitor arriving through the front door and following hallways would never know the wing exists; there is no corridor to it. Only a reader holding the floor plan (a crawler parsing the sitemap) can find those rooms, by teleporting straight to them.

Organizationally, this is what happens when publishing and linking are owned by different processes. The pages in Cloister’s phantom wing are real content (program pages, resource pages, location and service detail) produced by teams and templates that push to the CMS and register in the sitemap automatically. Linking, meanwhile, is editorial: someone has to decide that an existing page should point at the new one, and no process makes them do it. Publish is automated; weave is manual. Run that asymmetry for a few years and you get exactly this graph: a curated core that editors actively maintain, and a growing outer estate that the sitemap dutifully records and the structure never acknowledges.

The trap is that every conventional signal says the site is fine. The core is well-linked, so navigation feels complete. The orphaned pages are in the sitemap, so search consoles can report them “Submitted and indexed” with no errors, masking that they receive zero internal link equity and little reason for sustained crawl attention. Index entry without editorial wiring is not discoverability: those pages cannot compete for rankings the way content inside the graph can, and the organization has already paid the full engineering and editorial cost to produce inventory its own structure treats as nonexistent. The failure deepens when the reader is a link-following AI agent or multi-hop research crawler: those systems map a company through relationship graphs, not flat sitemap feeds, and so build their model from only the 59% of content the structure acknowledges.

Structural Signal: When a site’s link graph is healthy but its sitemap lists two fifths more pages than any editorial link can reach, the property has a phantom wing: content that is published, indexed, and structurally nonexistent.

The Five Lenses

Skeleton: Size & Connectivity

343 nodes, 1,281 edges, density 0.0109 (roughly 1% of possible page-to-page link pairings realized), average path length 2.5, diameter 5. For the connected core, these are solid numbers: pages sit a comfortable two-to-three clicks apart, and the diameter of 5 shows genuine layering rather than the flat everything-touches-the-hub shape of Episode 10 (Spire). The caveat is what the skeleton cannot see: these path lengths are computed over the pages that have paths. The 140 phantom pages have no path length to measure at all.

Circulation: Authority Flow

PageRank Gini of 0.79 (severe authority inequality; 1.0 would be total monopoly), a shade past even Episode 9 (Rotunda, 0.77), with the top 1% of pages holding 45% of authority and the single most central page holding 21%. This is the lens where Cloister genuinely fails inside its walls, not just outside them. One page, a section index at the heart of the dominant community, carries a fifth of the site’s entire link equity, with the next hub at 15% and a steep drop after the top five. The concentration is in Spire’s league (top 1% at 50%), but on a structure that otherwise works, which makes it more fixable: there are real neighborhoods here for authority to flow into, if the linking spread beyond the same few indexes. Click Hubs to see the drop-off.

Organs: Community Structure

10 communities at modularity 0.52 (crisp topical clustering, well above the ~0.4 mark that usually separates real sections from noise), with sizes of 84, 60, 55, 40, 30, 30, 22, 13, 7, and 2. That matches what the map encodes: eight clear color blocks, then a seven-page fragment and a two-page fragment that read as noise unless you look for them. This is the most balanced community structure since Episode 9 (Rotunda): a largest cluster holding just 24.5% of pages, four mid-sized sections, and only one two-page fragment. The site’s topical anatomy is real; these are recognizable functional sections, not artifacts. The amber is for the imbalance at the top and the fragment at the bottom, but the honest reading is that Cloister’s organs work. Which sharpens the question the Health lens answers: if the structure is this competent, why is 41% of the site unreachable by inbound editorial links?

Health: Content Isolation

41% orphan rate: 140 of 343 pages receive zero inbound editorial links. Dead-ends are a low 1.5% (five pages), islands 0.6%, and the largest connected component covers 99.4% of the pages that have links at all. The profile is the exact inverse of a decaying site. Decay looks like rising dead-ends and fraying islands. Cloister has neither; the wired part circulates fine. Its pathology is purer: a binary split between pages that receive editorial inbound links and pages that were never wired into from the rest of the site. Many orphans still link out, which is why they sit inside colored neighborhoods on the default map and only announce themselves under Orphans. This is the highest orphan share we have measured on a site whose other health metrics are green, and it echoes the pattern of Episode 5 (Nave), a healthy core with a frozen periphery, at nearly triple the proportional scale.

Nervous System: Depth & Bridges

Participation coefficient 0.36 (healthy cross-boundary linking), bridge rate 3.9% (50 bridge pages), average depth 2.3, with the depth distribution peaking at two clicks (206 pages) and three clicks (116 pages). For the connected core this is a working nervous system: fifty pages carry traffic between communities, cross-section flow is real, and nothing is buried deep. It is the strongest architecture score we have paired with a red health score in this series, and that pairing is the episode’s fingerprint. The wiring the phantom wing needs already exists everywhere else on the site. The site demonstrably knows how to link; it just stopped doing it for two fifths of what it publishes. Click Bridges to see the flow the phantom pages are missing.

What Would We Fix?

The Phantom Wing is the cheapest severe finding the series has produced, because the repair does not require inventing structure; the structure is already there and proven. Each of the 140 orphans has an obvious home: mapped by URL taxonomy and topic, their target communities already exist, already have hubs, and already have bridge wiring. A content repository is not an architecture; until a new page is wired into its parent hub and sibling cluster, it is hosted, not integrated. The work is to run the corridor to the wing: link each orphan from its community’s hub and relevant sibling pages, and fold the pattern into the publishing process so the wing stops growing back.

Running the optimizer on Cloister’s graph connected all 140 orphans into their community hubs, added 36 cross-silo bridges, and redistributed 40 links off the over-loaded central hub, lifting the edge count from 1,281 to 1,456. Orphan rate: 40.8% to zero. The residual task is editorial: the 21% single-page concentration only comes down when new links point somewhere other than the same section index.

OriginalOptimizedShow Changes

The original topology as crawled. All edges are from the live site.

Run a Corridor to the Phantom Wing

140 pages (41% of the site) are published and in the sitemap with zero inbound editorial links. Mapped by URL taxonomy and topic, each already has a community with an established hub. Linking every orphan from its community hub and sibling pages brings the entire published inventory into the graph.

+140 inbound links to orphan pages → Every published page becomes reachable by a link-following agent

Orphan Rate

40.8%→0.0%

Widen the Cross-Section Bridges

The core's ten communities are connected by 50 bridge pages (healthy), but the rescued orphans need lateral wiring too, or they hang off their hubs as leaves. Cross-silo links weave the recovered pages into the site's existing flow.

+36 cross-community bridge links → Recovered pages participate in cross-section circulation, not just hub spokes

Edge Count

1,281→1,456

Decompress the Central Hub

One page holds 21% of all authority and the top 1% holds 45%. Redistributing hub links toward section-level pages and deep-linking body content spreads equity across the neighborhoods the site already has.

redistribute 40 hub links + editorial deep-linking → Authority flows through sections instead of pooling on one index

Top-page authority share

21%→< 10%

Structure is Signal

Most episodes in this series found structures that failed at what they attempted: Episode 4’s islands never built ferries, Episode 7 never built a center, Episode 10 collapsed onto a single spike. Cloister is different in a way that should worry well-run sites more: the structure it attempted, it built well. Green skeleton, working neighborhoods, real bridges, sane depth. The failure is not in the graph. It is in the delta between the graph and the sitemap: between what the organization publishes and what its structure acknowledges.

That delta is invisible to almost every tool a web team uses. Crawl-based SEO audits start from links, so they mostly see the healthy core. Index coverage reports start from the sitemap, so they see the phantom pages as fine (“Submitted and indexed,” no errors) while those pages starve of internal equity and crawl priority. Indexed is not the same as discoverable or competitive. Only comparing the two inventories against each other exposes the wing, and that comparison is exactly what a link-following AI agent performs implicitly every time it reads the site: it takes the graph’s word for what exists, and the graph is silent about 41% of the answer.

The repair is one corridor at a time: 140 orphans linked into ten existing communities, 36 bridges, one overloaded hub decompressed, and a publishing process that adds the link when it adds the page. The Digital MRI is built to surface exactly this class of finding: not whether your pages are indexed, but whether your own structure admits they exist.