Topical Authority Without Backlinks: 32 Posts

What happens in the first month of a cluster build?

Faster than the guides on this subject say, and only for one kind of page. RankX AI published 32 articles across one subject between 20 August and 15 September 2026, and Google Search Console recorded the whole thing.

The domain needs describing first, because the answer depends on it. This was not a fresh registration. A small WordPress site sat here before, with eight or so indexed pages and about eight weeks of Search Console history, earning between zero and 21 impressions a day. It was replaced wholesale on 21 August 2026. Call it a near-standing start.

Worth stating before the first table: this is RankX AI measuring RankX AI with RankX AI. Every figure below comes from our own account, and the closing section says which surfaces are paid and which are free.

What followed the cutover was not gradual.

Two things in that table run in opposite directions, and the second one is the interesting one. Impressions spiked within six days, fell by two thirds over the following week, and recovered. Average position improved almost monotonically throughout, from roughly 70 in late August to the low 30s by mid-September.

The spike and dip is the shape most new sites see and it says little. The position trend is the one that tracks the cluster filling in. Read the caveats in the final section before taking either number to a stakeholder, because one of them is partly an artefact, and Search Console impressions with no clicks sets out how to tell an artefact from a result.

How much of the cluster did Google actually index?

Most of it, and the long articles did markedly better than the short supporting pages. A RankX AI index sweep running from 6 to 15 September 2026 inspected 198 of the 207 pages Google reports for the site and found 178 indexed, 19 not indexed and one unknown.

One denominator needs care, because it is easy to misread as the cluster. The 34 articles and guides are the pages this site classes that way, a group that also holds documentation pages and the blog index. It is not the same 32 as the 32 published articles, and three of those articles were too new to have been inspected at all.

The 146 supporting pages include the 60 glossary entries, and they contributed 16 of the 19 unindexed pages. Long articles reached 94 per cent indexed against 89 per cent for supporting pages. Length and substance predicted indexing better than anything structural did.

That gap is worth planning around. A cluster built from many short definitional pages is a cluster with more of its surface area exposed to the crawled, currently not indexed verdict, and pages Google declines to store rank for nothing at all.

Which pages in the cluster ranked first?

The narrow, specific ones. Within the RankX AI cluster the pages that reached Google's first two pages fastest each answer one specific question somebody typed, and that is the most actionable thing in this measurement. It is a tendency, not a rule.

The table below is every blog post that earned at least 40 impressions in the 28 days to 16 September 2026 and averaged position 20 or better. Below 40 impressions a position average moves too much on a handful of queries to mean anything.

Each of the four answers a specific question somebody typed. Broad category terms are a different contest, where the results are Ahrefs, Semrush and the rest of the SEO press. Narrow only helps where narrow also means uncontested: a specific question that every publisher in a field also targets behaves like a head term.

Publication date is a confound, since these four were among the most recently published, so part of their speed may be a freshness effect that decays. Read the table as direction rather than dose.

The simplest explanation is competition, and no special mechanism is needed. A specific question has a handful of pages written for it, and a category term has thousands, most of them older and better linked. Query fan-out predicts a similar asymmetry on the AI surfaces, but this table is classic web results.

The site-wide position distribution points the same way. Over the same 28 days, 13 queries sat in the top three and earned a 10.1 per cent click-through rate.

How long does topical authority take to show up?

Days for the long tail, on this site's evidence, which disagrees with almost every guide on this subject. It is worth being precise about what the disagreement is.

The published timelines cluster tightly. The most detailed one on the first page of results sets out a month-by-month schedule: nothing significant ranking for the first two months, long-tail queries reaching the second or third page around months three and four, and page one arriving at months five and six. Another page ranking for the same question puts first progress at eight to twelve weeks and full establishment at three to six months.

Neither cites a study. The more specific of the two attributes its numbers to the authors' own client work, which is a reasonable thing to do and is also the reason they cannot be checked by anyone else.

This site reached the long tail inside the first month. By day 26, seven pages held an average position of 20 or better and 13 queries sat in the top three. The fastest was an article on ChatGPT product feeds, which was averaging position 9.2 within four days of publication.

Competitive head terms are a longer contest, and one month of data cannot test the later stages of those schedules. What it does show is that the first stage arrives far sooner than the consensus says.

One caveat carries real weight here, and it is the reason the section above describes the domain before the numbers. This was a near-standing start rather than a fresh registration, and a site with no index presence at all would plausibly be slower. Anyone quoting the four-day figure without that sentence is quoting it wrong.

Can a topic cluster rank without backlinks?

Yes, for the long tail. RankX AI ran no link building during this period: no outreach, no digital PR, no paid placements and no guest posts. Every position in the table above was earned by publishing and internal linking alone.

What that bought was the long tail. The cluster picked up rankings across a few hundred long-tail keyword variations inside a month, and put seven pages at an average position of 20 or better with no backlink profile behind them.

Bing is stricter, because it names link signals as an input to whether a page enters its index at all rather than only to where it ranks. Bing indexing and the inbound links warning sets out what Bing says about that and what it means for a cluster built without links.

The mechanism that did the work costs nothing. Every supporting article links up to its pillar and across to its siblings, so each new page hands inbound internal links to the pages already published.

Zyppy's 2022 analysis of 23 million internal links (opens in a new tab) across 1,800 sites carries the counter-intuitive part. More inbound internal links tracked more Google clicks up to about 40 to 44 per page. Past roughly 45 to 50 the effect reversed and traffic fell again. More is not better past that point.

Treat that as direction rather than dose. It is a correlation, not a controlled test, and the data predates AI Overviews. SearchPilot's split tests are controlled and moved organic sessions the same way, though the largest of those ran in 2020. The internal-linking evidence base is old across the board, and a precise number off it should always carry its date.

One month is far too short to conclude that backlinks do not matter. It is long enough to show that a small domain can reach the long tail without them, which is a smaller and more useful claim.

How do you measure topical authority?

You cannot measure topical authority directly. Google has never published a specification for it, and no Google API or report exposes a score. The 2024 Content Warehouse leak (opens in a new tab) is the nearest thing to evidence, documenting an attribute that measures a site's topical coherence and another that measures how far individual pages deviate from its main topic.

That leak proves less than it is usually made to. Its own record states that the documents did not reveal how attributes are weighted or whether all of them are in use. Ahrefs' guide to topical authority concedes that Google has published no formal specification, then describes how Google measures it from those same attributes, which is the shape of the entire subject.

What you can measure is a proxy set of four readings, and every number in this article comes from it. Each one answers a different question, and each one is routinely quoted as though it answered a different question from the one it does.

None of those four is topical authority. Together they are what anybody outside Google actually has, and the discipline is reading each one for the question it answers rather than the one you wanted answered. Whether the thing itself is a mechanism or a description of what happens when a site covers a subject properly is the topical authority glossary entry, which argues it properly.

How did we choose 32 subtopics to write about?

By reading the search results rather than the tool scores, and that is the most transferable lesson here. Keyword research for a cluster has to produce subtopics somebody searches for, and the tools will mislead you about which of those you can win.

This article nearly made the mistake. Its own head term returned a competition rating of LOW from two separate keyword tools. Both were reporting Google Ads auction competition, which measures what advertisers bid rather than how hard a term is to rank for, and the two are unrelated numbers that share a word.

The search results settled it in one query. Page one held Ahrefs, Semrush, Search Engine Journal, Search Engine Land, HubSpot and MarketMuse. So the article was retargeted to the question underneath it, where a new site can win a place, and the head term was left to the established publishers.

The working rule: read the search results before committing to a term, treat every tool difficulty score as a hypothesis, and sort planned subtopics by whether a reader's search intent is genuinely unmet rather than by volume. A cluster targeting terms you cannot reach is a cluster that never ranks.

Does a topic cluster cannibalise itself?

Far less often than the advice implies, and the fear costs more than the problem. Thirty-two articles on one subject is exactly the shape that is supposed to produce keyword cannibalisation, so this cluster is a reasonable test of it.

The standard advice says to audit for pages competing on the same term and consolidate them. The best evidence available says most of what those audits flag is fine. Ahrefs went through 9,700 cases (opens in a new tab) of multiple pages ranking for one keyword on their own site, hand-reviewed a sample of 80, and found exactly one that needed action.

Treat that sample carefully. Eighty hand-reviews on a single site, published in February 2024, is directional, and a long way from settled, and it is still more evidence than the consolidate-everything advice has ever produced.

Our own data agrees. Across every query checked in the 28 days to 16 September 2026, Google returned exactly one RankX AI URL. That includes terms where this site publishes both a blog post and a glossary entry on the identical phrase, such as AI share of voice, where the blog post ranked and the glossary entry did not appear at all.

The method has a real limit worth stating. Search Console lists pages that earned impressions, so a second page with none would not show up. What this measures is that Google picked one URL per query. Whether the page it passed over was harmed is a different question and this data does not answer it.

How do you tell cannibalisation from diversification?

By intent, not by keyword overlap. Two pages sharing a term is normal. Two pages answering the same question at the same depth is the actual defect, and it has a different fix.

The practical check takes about ten minutes per head term and needs no specialist tool:

1. Pull the pages ranking for the term out of Search Console, not a third-party cannibalisation report. The pages-for-query view is the one that matters, because it shows what Google actually did, and a crawler only predicts.
2. Two URLs on one query is not a finding on its own. Look for them SWAPPING position week to week, which is the pattern that signals Google cannot decide, rather than sitting at stable different positions.
3. Apply the intent test. If one page could answer both searchers completely, you have one page written twice. If each answers a question the other would handle badly, you have coverage and nothing needs fixing.
4. Where it is genuinely one page written twice, merge and redirect the weaker URL. Where it is not, leave the URLs alone and fix the anchor text instead, so each page has a distinct phrase pointing at it.
5. Before publishing anything new, compare its sections against what is already live. That is where the defect is created, and it is far cheaper to catch than to merge later.

Step five is the one this site automates, because the judgement is the expensive part. Every draft is scored section by section against every section already published, and the score is a similarity measure, not a keyword match, which is what catches the same answer written twice in different words.

The lesson that travels is about scope: a corpus check only protects the corpus it can see. Compare a draft against every surface you publish, glossary entries and feature pages included, not only against other blog posts. The rest is reading.

What would we do differently on the next cluster?

Three things, in order of how much they cost to learn. Each is a decision the next RankX AI cluster will be built on, not general advice.

1. Publish the narrow pages first. The specific how-to articles reached page one or two within days. Front-loading the pillar delays the pages most likely to rank early.
2. Read the search results before committing to a term. Two tools rated this article's own head term LOW competition, and page one held six of the largest publishers in the industry. Both numbers were the ad auction rather than ranking difficulty.
3. Expect index churn on a young domain and measure it with two readings. Google reassesses pages in both directions, and a single coverage report read in isolation triggers pointless investigations.

One thing this measurement cannot settle, despite the temptation to claim it: whether to build one cluster at a time. This site published across seven sub-clusters at once, so it has no evidence either way, and topic clusters covers the argument without pretending otherwise.

What this measurement cannot tell you

Five limits, stated so every number above is read for what it is. Publishing a case study without its confounds is how most of the numbers in this field got into circulation.

• The impression counts include automated traffic. Part of them come from rank-tracker queries, which are tools rather than readers.
• The position trend is partly composition. Average position moving from about 70 to the low 30s reflects a changing query mix as well as pages ranking better. Treat it as directional.
• This was not one cluster. The 32 articles span seven sub-clusters of one subject, so nothing here supports advice about building clusters one at a time.
• The site is not a clean test. Tools, feature pages, comparison pages and a glossary all launched in the same weeks, so no claim of the form 32 blog posts caused this is available.
• It is one site. One domain, one niche, one month, no control. A case study with numbers is still a case study, and Search Console's own period comparison reports a 29,583 per cent impressions rise computed across windows with 21 and 59 days of data, which is an artefact and not a result.

What survives all five is the long-tail result, because it rests on per-page positions rather than on aggregate impressions. The index accounting survives too, since it comes from Google's own per-URL verdicts rather than from aggregate traffic.

How RankX AI tracks a cluster build

Every number in this article came out of RankX AI, which is our own product, and the measurement is the same one any account can run on its own site.

The index sweep, the position distribution and the per-page Search Console history are in the website audit surface, and brand mentions in AI answers are tracked in AI visibility. Both are paid. The AI readiness score is free and needs no account, and it covers the machine-readability half of the same question.

If you run the same measurement on your own cluster and it disagrees with ours, that is worth more to this field than another guide asserting that clusters work.