Agentic RAG is real. The GEO industry isn’t ready.

Mike King published a long piece yesterday arguing that single-shot retrieval-augmented generation — the architecture every "GEO playbook" on LinkedIn is implicitly built on — is already obsolete. Every major AI search platform has moved to something he calls *agentic RAG*: systems that plan, route between tools, retrieve in multiple passes, evaluate their own first drafts, and only synthesise an answer once they've decided the evidence is good enough.

I think he's right. And I think the implications are worse for the GEO consulting industry than the piece spells out.

If a single user query now triggers five to twenty internal sub-retrievals — and the agent decides which ones to keep, ignore, or re-run — then almost every measurement and optimisation tactic currently being sold as "GEO" is looking at the wrong layer of the pipeline. The visible layer. The last one. The one where decisions have already been made.

That's the loop. And the industry built a whole consulting category on top of the wrong frame.

What "agentic" actually means here

Strip the jargon and the shift is simple. The old pipeline was a single embedding query against an index, top results stuffed into a prompt, answer generated. If your page made it into the top-k, you had a chance. If not, you didn't. Either way, the retrieval was the citation set, and the citation set was visible.

Agentic RAG doesn't work like that. The model plans a series of sub-queries based on what the prompt actually asks. It pulls passages, reads them, decides whether the evidence is sufficient, and if it isn't, it goes back for more — sometimes reformulating its own queries, sometimes routing to a different tool entirely. By the time you see the final answer, the system has already made dozens of include-or-discard decisions you have no visibility into.

The traditional reverse-engineering playbook — rank checking, citation counting, prompt-by-prompt sampling — only sees the final stage. Everything upstream is a black box. You can rerun the same prompt and get a different citation set because the agent took a different path through the retrieval graph this time.

The measurement problem just got an order of magnitude worse.

Why this kills most of the GEO consulting playbook

I wrote last week that Google's new generative AI optimisation guidance basically declares the honest GEO playbook to be the SEO playbook with structured data and earned media on top. King's piece is the technical companion to that argument from the other direction — the systems themselves are now too complex for prompt-level tactics to work.

Here's what stops making sense once you accept agentic retrieval as the default:

Citation counting as a primary KPI. If a query triggers fifteen sub-retrievals and your page is in three of them but doesn't make the final answer, you scored zero in any tracker. Were you visible? Influential? The data can't tell you.

Prompt-by-prompt audits. Running the same prompt twice now produces different retrieval paths. Sampling a few prompts and extrapolating to "we appear in 14% of relevant queries" is closer to astrology than analytics.

Optimising for the final synthesis. The synthesis step is the easiest part of the pipeline to influence and the least valuable. The upstream retrieval and grading steps — which you can't see — are what actually determine whether you're considered at all.

Every "AI visibility tool" priced at four figures a month is, at best, monitoring the visible 10% of the pipeline. That's not nothing. But it's being sold as if it's the whole picture, and it isn't.

The honest position is uncomfortable

I'm going to say something the rest of this industry won't: most of the people selling GEO services right now are operating with a worse mental model of the underlying systems than the people they're selling to assume they have.

That's not an insult. The systems moved. The vocabulary hasn't caught up. There are senior, well-credentialled SEOs publishing GEO frameworks built on a 2023 understanding of how retrieval works, because that understanding was correct in 2023 and nothing about the consulting language has updated.

But if King's read is right — and the patent evidence and platform behaviour he cites suggest it is — then the gap between what GEO services claim to do and what's technically possible is widening, not closing. You cannot reliably optimise for a stage of a pipeline you cannot observe, in a system where every query takes a different path.

What you *can* do is make your content genuinely useful, structurally clean, citably authoritative, and present in the places agents are likely to retrieve from. Which is — and I'm sorry to keep saying this — the SEO playbook.

What this actually means for the reader

If you're a business owner being pitched GEO services right now, three questions cut through most of the noise.

First: when the agency talks about measurement, do they distinguish between citation, retrieval, and final synthesis? If they conflate the three, they don't understand the pipeline. If they only measure the final answer, they're measuring the easiest 10% and pricing it like the whole thing.

Second: what's the underlying work? If the playbook is "structured data, useful content, earned mentions, technical hygiene, brand-building" — that's honest. That's the work. It's also indistinguishable from competent SEO. Pay the right amount for it.

Third: are they selling certainty? Anyone telling you they can guarantee AI Overview placement, or citation share in ChatGPT, is selling a product the underlying architecture doesn't support. The agent decides. The agent's decisions are not reverse-engineerable from the outside. Anyone claiming otherwise is either confused or lying.

The industry is going to take a while to catch up to this. The vocabulary will lag the architecture for at least another year. In the meantime, the best protection against being sold something useless is understanding that the system has more moving parts than the tools watching it can see — and pricing your scepticism accordingly.

The bit nobody wants to say out loud

King ends his piece with what he calls his strongest opinion of the year: that the only honest way forward is model distillation — training your own smaller models on your own data and operating outside the visibility-of-someone-else's-pipeline problem entirely.

I'm not sure he's right about that as a tactic for most businesses. It's expensive, capability-heavy, and presupposes a strategic posture most SMEs don't have.

But the *underlying* point is the right one. If you can't see the system, can't measure the system, and the system keeps changing, then the only durable position is to be the kind of brand and the kind of source that any reasonable retrieval system would reach for. Brand. Authority. Genuine usefulness. Technical cleanliness. Wide presence across surfaces, not just Google.

Everything else is theatre performed for an audience you can't see.