Topical Authority: A Practical Website Architecture Guide

The phrase "topical authority" gets used as if Google maintained a per-site, per-subject number that you could raise. It does not publish one, and no engine exposes one. What the phrase usefully describes is a set of measurable properties of your own content: how much of a subject you cover, how coherently that coverage is grouped, how well the pieces link to each other, whether they are deep or thin, and whether they duplicate each other. Those are things you can compute locally, explain to a client, and act on. This guide is about doing that.

Stop treating it as a score

There are two reasons the score framing is worth abandoning, and only one of them is pedantry.

The pedantic reason: no first-party documentation from any search engine describes a topical authority metric. Anyone reporting your "topical authority: 62" is reporting a vendor's own calculation, which is fine as long as they say so and explain it, and misleading otherwise.

The practical reason is better. A single number tells you nothing about what to do next. "Your coverage of employee benefits is thin — four documents, all under 600 words, none linking to each other" tells you exactly what to do. This is why SiteRank reports bands and counts rather than a 0–100 score for any domain: a percentage a reader cannot decompose into observations is not explainable, and under our methodology an unexplainable number does not ship.

So the working definition: topical authority is the strength, completeness and consistency of a site's information about a subject area — a property of your corpus, computed and explained, not a judgement handed down.

The five properties that matter

Coverage

Does the corpus address the subject's real subtopics, or only its headline? Coverage failures come in two shapes: absent subtopics (nothing on the site addresses a question your audience actually asks) and shallow subtopics (something exists, but it is a 300-word stub).

The honest caveat is about where the reference scope comes from. A "gap" defined by comparing your corpus against your own content is one thing; against competitor content is another; against questions real users asked is the strongest; against a language model's imagination is a HYPOTHESIS and must be labelled as one. Say which you used. A gap list that does not name its reference scope is unfalsifiable.

Clustering

Do the documents about a subject actually group together, or is the subject scattered across three unrelated sections of the site? Clustering is what turns a pile of posts into a topic you can reason about, and it is computed from the content itself rather than from your category taxonomy — because a site's taxonomy frequently reflects publishing history rather than subject structure.

Cluster membership is genuinely fuzzy. A document can belong to two clusters. Forcing a single assignment produces confident nonsense, which is why documents that resemble nothing else are better left unclustered than jammed into the nearest group.

Connectivity

Do the documents in a cluster link to each other, and does the cluster have an entry point? Internal linking is how discovery works, how a reader moves between related pieces, and how a crawler finds the deep parts of a site. A cluster of twelve excellent posts with no links between them is twelve isolated documents.

The two findings worth acting on are orphan pages (indexable, in no navigation, zero internal in-links from indexable content) and weak internal linking (a document with far fewer in-links than its cluster peers). Both are defensible because they are counts, not judgements — as long as you say whether you counted menus, breadcrumbs and archives as links.

Depth

Does the coverage go past the definitional level? Depth is not word count. Word count is a readability measure and a terrible quality proxy: a 2,000-word restatement of a Wikipedia summary is shallower than a 700-word piece containing original measurement. Depth shows up as distinct subtopics addressed, specific examples, primary sources cited, and questions answered that the obvious articles do not answer.

Duplication

Do two documents target the same intent? Near-duplicate content splits internal links and reader attention between pages that should be one page. But similarity is not proof: a series, a set of localisations, and a genuine update all look like duplication to a similarity metric. Treat it as a candidate with evidence attached and let a human decide. A tool that automatically declares cannibalisation from a cosine score will be wrong often enough to lose your trust.

How SiteRank computes it

The clustering runs entirely on the site — no content leaves the WordPress install — which constrains the method to something lexical and deterministic. That is a deliberate trade: TF-IDF does not capture synonymy the way embeddings do, but it is cheap, repeatable, and every result can be traced to specific terms in specific documents. The full parameters are published; the shape of it is:

1. Strip boilerplate first. nav, header, footer, aside, script, style, form and similar elements are removed before any text analysis. Untreated boilerplate dominates lexical similarity and produces "clusters" that are really the theme's footer. This is the single largest source of bogus results in any clustering pipeline, and skipping it invalidates everything downstream.
2. Vectorise with TF-IDF, with sub-linear term frequency and smoothed IDF. Title tokens are weighted ×3.0 and taxonomy tokens ×2.5, on the reasoning that both are human-authored subject signals rather than incidental prose.
3. Drop site-wide terms — any term appearing in more than half the documents is treated as boilerplate and discarded. This adapts to the site rather than assuming a language or shipping a stopword list.
4. Keep the top 40 terms per document and L2-normalise, which bounds the pairwise comparison.
5. Cluster by single-link agglomeration over cosine similarity, threshold 0.18 — a threshold rather than a fixed number of clusters, because the number of topics on a site is not known in advance and forcing a k produces arbitrary groupings.
6. Label from a taxonomy term shared by at least 60% of the cluster where one exists — a human-authored label beats a derived one — otherwise from the top three distinctive terms. Generated labels can be wrong, and the interface says so and lets you rename, merge or split.
7. Rank pillar candidates as 0.5 × normalised inbound links + 0.3 × centroid relevance + 0.2 × normalised word count. Word count is present but capped at a fifth of the weight, because length is not authority.

Above 600 documents the engine returns no clusters at all rather than a partial result you would read as complete, and records why in the activity log. That ceiling is checked before the corpus is loaded, since clustering needs every document's text in memory — and memory, not comparison count, is what actually binds: the term-frequency map alone runs to about 97 MB at 1,500 documents, which is past a typical PHP limit before the vectors even exist. It is an honest limit, stated up front.

The bands

These summarise counts. They are not quality judgements, and they should not be read as one — a three-document cluster of exceptional pages is "thin" in coverage and may be exactly right for that subject. The band tells you where to look; you decide whether anything is wrong.

Turning it into architecture

Pillar and supporting structure

A pillar is the document that represents a cluster comprehensively and serves as its entry point. Supporting documents each cover one facet in depth. The pillar links down to every supporting document; each supporting document links back up and sideways to two or three genuinely related siblings.

Pick pillars from the data, not from word count: breadth of subtopic coverage, internal in-links, depth, and the site's own hierarchy. Then check the choice by hand — if the identified pillar is not the page you would send a new client to first, either the page is wrong or your architecture is. Both are findings. Pillar pages and topic clusters covers the structure in detail.

Internal linking, done properly

The useful recommendation is not "add a link". It is: this passage in document A discusses a facet covered by document B, and there is no link; here is the sentence and here is the anchor text. A recommendation without a source passage is busywork.

Two limits worth holding. Do not propose hundreds of links — that is link stuffing, and it degrades both the reader experience and the signal. Do not link the same target from every page. Internal linking for SEO, retrieval and AI goes further.

Consolidate before you create

The reflex on finding a gap is to commission a new article. Often the better move is to expand or merge existing pages. A careful editor consolidates; a content mill publishes. Google's spam policies define scaled content abuse as generating many pages for the primary purpose of manipulating rankings rather than helping users, and explicitly include using generative AI tools to generate many pages without adding value. A gap list turned into a fifty-article brief is precisely the failure mode that policy describes. OFFICIAL PROVIDER GUIDANCE.

Exclude what should be excluded

Analysing noindex pages, canonicalised duplicates, tag archives and paginated listings distorts every metric above. Build the corpus from canonical, indexable content, record what you excluded and why, and show that list — the exclusions explain half of the surprising results.

Evidence tiers

The last row is the one that gets overstated. Broad, genuine coverage of a subject is defensible on its own merits and plausibly helps a retrieval system find something relevant to more prompts. It is not a documented mechanism, and it must not be sold as one.

What not to assume

• That Google publishes a topical authority metric. It does not.
• That similarity implies duplication. It implies a candidate for human review.
• That age implies staleness. An evergreen page is not defective for being old, and bumping a date to look fresh is deceptive. Use real signals — a superseded reference, a dead outbound link, a stated version. See content freshness.
• That your category taxonomy reflects your topic structure. Compute it and compare; the disagreement is often the most useful output.
• That a cluster label produced by an algorithm is a category. It is a suggestion you should be able to rename.
• That word count measures depth.

Key takeaways

• Topical authority is a property of your corpus — coverage, clustering, connectivity, depth, duplication — not a score any engine publishes.
• Report bands and counts a reader can decompose, not a composite number.
• Strip boilerplate before any text analysis, or every downstream result is contaminated.
• Choose pillars from in-links, breadth and centrality, with word count capped at a minor weight.
• Recommend specific links from specific passages, and consolidate before commissioning new pages.
• Deep coverage helping AI citation is a hypothesis. Coverage helping readers is not.

Official sources & further reading

• Creating helpful, reliable, people-first content — Google Search Central
• Spam policies for Google web search — Google Search Central
• AI features and your website — Google Search Central
• Intro to structured data markup on Google Search — Google Search Central

Related reading

• Pillar pages and topic clusters
• Internal linking for SEO, retrieval and AI
• Content freshness
• Structured organization information
• Topical authority analysis