Topical authority in AI search describes how useful a retrieval system finds a source across the full set of questions around a subject, along with how consistently the rest of the web connects that brand to the subject. It overlaps with the version of topical authority SEO teams have worked toward for years, where thorough coverage and outside references help a site rank across a cluster of queries, but different systems do the evaluating.
Google’s AI Overviews, ChatGPT, Perplexity, and Claude each build answers from individual passages and sources, which is a different task from ordering a list of pages. At Curious Minds Media, we see the gap between those two processes in the client data we review every month, and it has changed how we plan content and technical work for the sites we support.
No AI platform has published a topical authority signal. Google has described a topic authority system only for news results, and it has never documented a general one for traditional search.
What follows draws on the largest third-party studies available, chiefly from Ahrefs and Semrush, on the platforms’ own documentation, and on what we observe across client sites. Most of the study evidence is correlational, and we’ve tried to say so where it matters.
How ranking authority is assigned
Ranking authority is evaluated one query at a time. A search engine compares the candidate pages for a keyword using relevance, links, technical health, and historical performance, and the page that comes out ahead holds the position.
Topical depth feeds into that comparison, since a site with thorough coverage of a subject tends to build the internal links and outside references that strengthen its individual pages. The result is an ordered list of URLs, and the page is the unit being judged.
How an AI system assembles an answer
An AI-generated answer starts from a different unit of work. Google describes AI Mode as using a query fan-out technique, which issues multiple related searches across subtopics and data sources at the same time and brings the results together into one response.
The system then looks for passages that resolve those related searches, synthesizes them into a single response, and attaches a handful of citations. A source earns its citation because a specific passage answered a specific part of the question, regardless of how the full page performs for the original keyword.
A Google patent granted in 2023, US11663201B2, describes a version of this process in detail. It covers a trained generative model that produces variants of a submitted query, including equivalent, follow-up, generalization, specification, and clarification queries, then submits those variants to a search system and builds its output from the responses.
The patent also describes using the responses to those variants to corroborate an answer to the original query and flagging answers the variants don’t support. A patent documents what Google has built and protected, and it can’t tell us which products use the method today. The mechanism it describes does line up closely with how Google characterizes fan-out.
Ahrefs has tracked what this does to the overlap between rankings and citations. In its July 2025 study of 1.9 million citations from one million AI Overviews, 76.1% of cited pages ranked in Google’s top 10.
When Ahrefs re-ran the analysis across 863,000 keywords and 4 million AI Overview URLs, the share of cited pages that also ranked in the top 10 for the same query had fallen to 38%. Ahrefs connects the shift to fan-out: the closer a document’s passages match the sub-queries, the more likely it is to earn a mention or citation in the generated answer.
We read that data as a change in what ranking accomplishes. A strong position still places a page among the sources a retrieval system considers, and top-ranking pages remain a large share of what gets cited. Ranking alone no longer determines which of those sources appear in the answer. That fits Google’s own description of AI Mode as rooted in its core quality and ranking systems.
Where ranking strength stops predicting AI visibility
Semrush and Kevin Indig of Growth Memo measured a similar divergence in ChatGPT. Their study tracked 1,094 US categories monthly from January through June 2026, with five representative prompts per category, and found that only 15.2% of categories had a clear owner.
When the researchers compared each category owner against the runner-up, the owner had higher branded search volume in 55.7% of pairs, higher organic traffic in 48.4%, and a higher Authority Score in 52.5%. Those results sit close to even odds.
Only branded search volume was a statistically significant differentiator, and its edge was modest. Indig is careful to note that the data shows traditional SEO metrics aren’t enough to explain topic ownership, which is different from showing they don’t matter.
The same study found that topic ownership, once established with a clear margin, tends to last. Owners with a clear lead stayed on top in 90.4% of month-over-month comparisons, while narrow leads changed hands often. That durability is the strongest practical argument for building depth on a focused set of subjects before expanding into new ones.
A related effect shows up in traffic. On a travel marketplace account we support, commercial pages held their positions over recent months while informational blog content lost clicks, even with impressions holding steady, as AI Overviews covered more of the results page for those queries.
Pew Research Center documented the same effect from the user side in an analysis of March 2025 browsing data. Google users who saw an AI summary clicked a traditional search result link in 8% of visits, compared with 15% of visits where no summary appeared.
Cited and named are separate outcomes
AI visibility shows up in two forms. A citation means a system retrieved a passage from a page and linked to it. A mention means the model named the brand in its answer, with or without a link.
Semrush found that the two diverge sharply: only 21% of the most-cited domains in a category were also the most-mentioned brand, and the two signals showed a slightly negative correlation of -0.229.
That distinction explains why off-site presence matters alongside on-site content. Citations respond most directly to what a site publishes and how well its passages answer the sub-queries. Mentions depend more on how often, and how accurately, other sources connect the brand to the subject.
In an Ahrefs study of 75,000 brands, web mentions of a brand showed the strongest correlation with that brand’s visibility in AI Overviews, at 0.664, and the strongest correlations overall came from off-site factors, while link metrics correlated only weakly. It’s worth tracking citations and brand mentions separately, because a site can gain citations while its share of mentions stays flat.
Retrieval starts with crawler access
None of these signals count if the crawler that feeds an AI system can’t fetch and read the page. The crawlers that feed AI answers are often different from the crawlers teams have in mind when they write blocking rules.
OpenAI documents its robots.txt settings as independent of each other: a site can allow OAI-SearchBot so it appears in ChatGPT search while disallowing GPTBot so its content isn’t used for model training. Sites that opt out of OAI-SearchBot won’t be shown in ChatGPT search answers, though they can still appear as navigational links.
Google handles its AI features inside Search, so its controls work on a different principle. Google states that Google-Extended doesn’t affect a site’s inclusion in Google Search and isn’t used as a ranking signal.
AI Overviews and AI Mode are features within Search, so blocking Google-Extended doesn’t remove a site from them. Those surfaces draw on the standard index that Googlebot builds.
In practice, this means a robots.txt rule, a CDN bot-management setting, or a firewall configuration written with one crawler in mind can remove a site from an AI surface without anyone noticing. On one site we maintain, a third-party maintenance service’s requests started failing, and the client’s own security configuration was the likely cause.
Any rule that blocks unwanted bots can catch wanted ones too, so bot blocking added to fix a performance problem should be checked against the crawlers the site wants to keep. Managed hosts are adding their own bot-management controls at the edge, which gives teams one more layer where AI crawlers can be turned away by accident.
Content that only appears after client-side JavaScript runs carries a similar risk. On a recent headless build that generates location pages automatically, confirming that crawlers receive the rendered content was one of the first questions we raised.
Our engineers manage hosting, firewall, and CDN configurations for many of the sites we maintain. Crawler access and indexing are part of every technical audit we run, and AI agent readiness is part of our client health checks.
The signals we see matter most for AI visibility
With access in place, the evidence and our client work point to five signals that matter most.
The first is coverage across the full set of questions around a topic. A page written to win one head term often answers that term well and leaves the follow-up questions unaddressed, and those follow-up questions are what fan-out generates.
The variant types listed in Google’s patent, such as follow-up, generalization, and specification queries, make a practical planning list for those surrounding questions. Content that answers those questions in clearly separated, self-contained passages gives a retrieval system more places to find a usable answer.
Recent Google Research work points in the same direction. A September 2026 paper on fan-out retrieval trains models to produce sub-queries that are diverse, aligned with the original prompt, and grounded in what a database contains, and it names the failure it is built to avoid: near-synonymous sub-queries that return redundant results.
The experiments used fashion and music datasets, so they say nothing direct about Google Search. They do show researchers optimizing fan-out to cover the distinct facets of a topic, which rewards content built around those facets.
On client sites, we structure sections around the questions a reader asks next, and we write each section so it still makes sense if it is lifted out of the page.
The second is consistent presence across third-party sources. AI answers draw heavily on platforms outside a company’s own site. In Ahrefs’ updated citation analysis, YouTube alone accounted for 5.6% of all AI Overview citations.
The same patent describes checking an answer against the responses to related query variants, which may be one reason consistency across sources matters: a claim that appears in only one place gives a system built this way less to corroborate.
A company’s about page, leadership bios, directory listings, video presence, and independent coverage all contribute to whether a retrieval system can identify the organization and associate it with a subject. Where those sources disagree on names, services, or locations, the entity becomes harder to recognize.
The third is attributable expertise. No platform publishes how it weighs authorship, but named authors with documented experience are one of the few credibility signals a publisher fully controls, so we recommend author pages as a baseline on the sites we support.
On a retail security account, we recommended author pages built around the client’s own subject-matter experts for this reason, and the client’s team took on creating them. On that same travel marketplace account, adding expert insight sections to existing articles produced measurable growth on the pages that received them. That kind of first-party detail is what a generic summary can’t supply.
The fourth is structural clarity. Headings that state the question being answered, FAQ sections on in-depth content, and schema markup such as FAQ, Service, and Organization types all make it easier for a retrieval system to identify what a passage covers and who published it.
Semantic HTML does part of this work before any schema is added, since native elements give parsers the same structure a reader sees. We add FAQ sections selectively, on pages where the depth of the topic supports them, because adding them to every post dilutes their value.
The fifth is consolidation. When several pages on one site compete for the same subject, both ranking signals and retrieval signals get split between them.
Great News Life, a Midwest media organization, was publishing identical content across five WordPress domains, and the duplication kept any of those properties from building topical authority. We consolidated the content into a single WordPress installation and restructured the architecture so the organization could build visibility as one authoritative property.
How we approach this at Curious Minds Media
We treat SEO and AIO as one discipline, because the technical foundation that supports traditional rankings also supports retrieval: clean crawlability, structured data, Core Web Vitals, and well-organized content.
Curious Minds Media has been building and maintaining websites since 2004, our engineers are in-house and US-based, and our search visibility work runs alongside that development work. That combination lets us fix the access, rendering, and architecture problems that a content-only engagement can’t reach.
Most engagements start with a technical audit covering crawler access and indexing. From there, we identify which pages earn their place and which ones dilute the site’s signals.
For large content libraries, that often means redirecting or consolidating a significant share of low-performing articles, so the remaining inventory carries clearer authority on fewer topics. We then rewrite priority pages around the full set of questions a reader brings to the subject, add named authorship and first-party detail, and implement schema that describes both the content and the organization behind it.
Off the site, we review how the organization appears in the third-party sources AI platforms draw from, including directory listings, leadership profiles, video platforms, and review sites, and we flag the inconsistencies that make the entity harder to identify.
For measurement, we’ve made capturing baseline metrics part of how we start engagements, because AI visibility is hard to evaluate without a before picture. From there, we track the AI assistant traffic channel in GA4 alongside traditional rank tracking and monitor AI Overview appearances for priority queries.
We treat prompt-tracking tools as samples: they follow a fixed set of prompts, and AI answers can change from one run to the next. Search Console doesn’t separate AI Overview visibility from standard results, so we combine sources to give stakeholders an accurate picture of where visibility is moving.
If you want to understand how your content is currently being retrieved and cited in AI-generated answers, our SEO and AIO optimization services page outlines how we evaluate it.
