Agentic RAG rewrites the rules for content visibility

What happened

Per iPullRank, every major AI search platform has moved past simple retrieval-augmented generation (RAG) into agentic RAG. Mike King's argument, building on his August 2020 reading of Google's REALM paper, is that ChatGPT, Perplexity, Google AI Mode, and Claude no longer run one query against one index and synthesise an answer. They plan, decompose, retrieve in parallel, evaluate, and re-retrieve. The system behaves like a researcher, not a librarian.

The shift is architectural, not cosmetic. iPullRank reports that a single user prompt can now generate dozens of sub-queries fanning out across web indexes, vector stores, and tool calls, with the model judging which sources to keep and which to discard mid-process. Google's AI Mode and ChatGPT search both operate this way by default.

That means the unit of optimisation is no longer "the query." It is the set of latent sub-queries the model decides to issue on a user's behalf, none of which the user ever sees.

Why it matters for your brand

If you are a CMO at a bank, a multilateral, or an industrial group, your content has been built for a world that no longer exists. Most enterprise content strategies still assume one prompt maps to one answer, and that ranking for a head term wins the citation. Agentic RAG breaks that assumption. The model is now asking five to twenty questions you cannot see, and your page either answers one of them precisely or it does not get pulled into the context window at all.

For financial services brands, this is acute. A prompt like "is private credit a systemic risk in 2025" no longer returns a single synthesised view from the top-ranked explainer. The agent decomposes it: what is private credit, what is the current AUM, who are the largest issuers, what have the IMF and BIS said, what are the counterarguments. Each sub-query selects a different source. A bank that publishes only a flagship thought leadership PDF wins zero of those slots. A bank that publishes a structured definition page, a data page with current AUM figures, a regulator-response page, and a counterargument page wins four.

For multilaterals and policy institutions, the implication is sharper still. Agentic systems privilege sources that answer narrow factual sub-queries with authority. UN agencies, the World Bank, the OECD, and ISO already hold that authority. But their content is often locked in 80-page PDFs that the agent cannot decompose efficiently. The institutions that will dominate citations in 2026 are the ones that extract every claim, statistic, and definition from those PDFs into addressable web pages. The ones that do not will watch lesser-credentialed think tanks get cited in their place.

Industrial groups face a different problem: the agent's sub-queries often include commercial intent (supplier comparisons, specification lookups, sustainability claims) that the brand's corporate site does not address at all. HOLCIM-class companies need product-level, specification-level, and claim-level pages that match the granularity of an agent's decomposition. Corporate narrative content is invisible to this layer.

Distribution changes too. If the agent retrieves from multiple indexes in parallel (web, Reddit, YouTube transcripts, news APIs, sometimes proprietary data partners), then a single-channel content strategy guarantees you appear in only one of the agent's retrieval passes. Brands that publish the same substantive claim across owned media, earned coverage, and a structured data feed get pulled into more sub-queries and weighted more heavily in the final synthesis.

The signal in context

Agentic RAG is the technical reason behind every citation-pattern shift marketers have noticed in the last twelve months: why long-tail informational content suddenly outperforms brand terms in LLM answers, why Reddit and Wikipedia citations spiked, why AI Overviews pull from five to eight sources rather than two or three. These are not separate trends. They are surface effects of the same underlying architecture moving from retrieval to reasoning-plus-retrieval. Mike King has been calling the trajectory since 2023; what is new in 2025 is that it is now the default behaviour across every consumer-facing AI search product, not an experimental mode.

The strategic consequence for senior marketers is that "AI search optimisation" is no longer a variant of SEO. SEO optimised for a ranking position on a known query. Agentic RAG optimisation requires modelling the questions a planning agent will generate on your category's behalf, then ensuring authoritative answers exist on the open web, on properties the model trusts, in formats it can parse. That is a different operating model, a different content taxonomy, and in most enterprises a different team. The brands that restructure for it in the next two quarters will set the citation defaults that everyone else inherits.