Does topical authority still move the needle in LLM answers?

Ahrefs has rebuilt its case for topical authority, arguing that the same depth-of-coverage signal Google has rewarded for years now decides whether ChatGPT, Perplexity and Google's AI Overviews quote you or your competitor. The blog frames topical authority as the site-level trust a search engine assigns when one publisher covers a subject more comprehensively, more consistently, and with better internal structure than the rest. The question worth asking: does that thesis survive contact with how large language models actually retrieve?

Mostly, yes. But the mechanism has shifted, and so has the payoff.

What the models are actually doing

Retrieval-augmented generation does not reward a domain for "owning a topic" in the abstract. It rewards specific URLs that match the embedding of a specific prompt, then leans on the model's prior trust in the domain to break ties. Topical authority, in LLM terms, is the tiebreaker. A bank with forty well-structured pages on trade finance does not out-rank a competitor because Google has crowned it; it out-ranks because forty pages produce forty embeddings that intersect with forty variants of a user's question, and because the model has seen the domain cited elsewhere often enough to treat it as safe.

Ahrefs is right that the underlying behaviour, build deep coverage of a defined subject, still works. The blog is less explicit about why it works differently now. Three shifts matter.

First, granularity has gone up. Ranking for "supply chain finance" mattered in 2019. In an LLM world, the unit of competition is the sub-question: "how does supply chain finance affect Basel III capital treatment for a mid-cap exporter". Brands that wrote one pillar page and ten thin satellites are invisible to that prompt. Brands that wrote forty sub-question pages are quoted verbatim.

Second, citation begets citation. Models that have seen a domain referenced in Reuters, the FT, an IMF working paper and three industry trade titles will retrieve and quote that domain more readily, even on pages those outlets never linked to. Earned media is now a topical-authority input, not a parallel channel. For multilaterals and policy institutions, this is the cheap win: the citations already exist in the training data, but the underlying pages are often gated, ungrammatical to a crawler, or hidden behind PDF.

Third, structural signals are doing more work. Clear H2s phrased as questions, definitional opening sentences, dated updates, and author bios with verifiable credentials are not cosmetic. They are what makes a page chunk cleanly into a retrieval index. Ahrefs covers internal linking and content hubs; it understates how much of the LLM uplift comes from chunkability rather than link equity.

Where the Ahrefs frame undersells the cost

The blog suggests topical authority is a matter of "covering a topic comprehensively." For a consumer SEO that may mean thirty posts. For a global reinsurer trying to be the cited source on parametric cover for climate risk, comprehensive means closer to two hundred pages, refreshed quarterly, written by named underwriters, and cross-referenced to regulatory texts the model already trusts. The investment is closer to a research desk than a content calendar. Most enterprise marketing teams are not staffed for it, which is precisely why the brands that do staff for it are pulling away in AI citations.

There is a second cost the post skips: defending the authority once you have it. LLM training cut-offs and retrieval indexes refresh on their own clocks. A bank that dominated ESG disclosure content in 2023 and then went quiet in 2024 is already being displaced in answers by competitors whose newer pages match newer prompts. Topical authority decays faster in LLMs than in classical search, because the models reweight on every retraining pass.

What this means for the brands we work with

For financial-services marketers, the implication is to stop measuring topical authority by share of voice on head terms and start measuring it by citation rate across the long tail of sub-questions a buyer actually types into ChatGPT. The tooling to do this is now available; the org charts to act on it largely are not.

For multilaterals and UN agencies, the prize is unusual. These institutions already hold the authority the models defer to. What they lack is web estate built for retrieval: too many PDFs, too few HTML pages with clean structure, too little dating and authorship. Fixing that is a six-month job, not a strategy reset, and it produces disproportionate gains because the underlying trust is already priced in.

For industrial groups and policy institutions, the question is narrower: which two or three topics do you intend to be the cited source on by the end of next year, and are you willing to publish at the cadence and depth that requires. Ahrefs has restated an old discipline. The brands that treat it as a new operating commitment will be the ones quoted back to buyers by a machine.