Agentic SEO is a workflow where autonomous AI agents — not humans — plan, execute, and verify SEO tasks like keyword research, technical audits, content drafting, internal linking, and citation tracking, then iterate based on how AI assistants and search engines respond. It's the discipline of putting a full SEO team into a loop of coordinated LLM agents that can run for hours without human keystrokes.
The moment I stopped calling this "AI-assisted SEO"
A series-B DevTools founder called me in October 2025. His team had wired up a slick agent chain — Claude for keyword expansion, GPT-4o for briefs, a custom Python tool that pushed 40 posts a week to WordPress. Traffic charts looked healthy for six weeks. Then his sales team pinged him: three enterprise buyers had all said the same thing on discovery calls — "we asked ChatGPT for the best tool in your category and yours never came up." Publishing had scaled. Presence in the systems buyers actually use hadn't. That gap is the reason we sat down and wrote our manifesto on agentic discovery.
Of 40 B2B sites we found running some form of "agentic SEO" (autonomous crawl → brief → draft → publish loops), only 9 had passed the more important second test — being machine-readable enough for AI agents on the other side (ChatGPT, Perplexity, Claude) to cite them. The other 31 were producing more content than ever and still invisible in AI answers. Source: internal WebFlur audit dataset, Q1 2026.
What "agentic" actually means (it's not just automation)
An automation runs a fixed script. An agent decides what to do next. That decision loop — plan a step, take it, observe what happened, revise the plan — is the entire difference. When someone says "agentic SEO", they mean the SEO workflow itself has been handed to an LLM (or several) with tools, memory, and the authority to change course without asking a human first.
The practical marker: if your setup asks a human to click "approve" between every step, that's AI-assisted SEO. If the system can chain "crawl the top 20 SERP results → cluster the intents → draft 12 briefs → assign them to writer agents → verify the drafts against the brand voice → publish → measure share-of-model in ChatGPT after 14 days → adjust the next batch" without a human in the middle, that's agentic SEO.
The academic root sits in the arXiv paper that first defined Generative Engine Optimization, published in 2023. GEO gave the field a vocabulary for optimising content to be cited by generative engines. Agentic SEO takes that vocabulary and hands the whole optimisation loop to agents instead of consultants.
How the agentic SEO stack actually works — planner, executor, verifier
A working agentic SEO system almost always looks like a three-layer loop, even when a vendor markets it as one product.
Layer 1 — the planner. One LLM instance reads the goal ("grow non-brand impressions on 'B2B procurement software' by 40% this quarter"), then decomposes it into a work tree: keyword clusters, target queries, page types, internal-link changes, schema updates. Anthropic's Claude and OpenAI's o-series models are the ones we most often see doing this job in production. The planner is where honest agentic SEO earns its name — a fixed content calendar isn't a plan an agent made.
Layer 2 — the executors. Specialised agents (with tools) do the actual work — an SERP-scraping agent, a brief-writing agent, one or more writer agents, an internal-linking agent that reads the sitemap and proposes anchors. This layer got a hard shove into the mainstream when OpenAI released ChatGPT Agent, which put a general-purpose browsing executor in every ChatGPT Plus account. Since that release, the marginal cost of an "SEO executor" that can actually visit a page and read it has dropped to pennies per run.
Layer 3 — the verifier. This is the layer most content-farm setups skip, and it's what separates a real agentic system from a scaled slop machine. A verifier is a second LLM instance whose only job is to grade the executor's output against a rubric — the "LLM as judge" pattern. Does the draft actually answer the query? Does it hit E-E-A-T signals? Does it match brand voice? Does the internal link go to a page that exists? If the verifier fails an output, the loop hands it back to the executor with the specific fault, not a vague "try again".
We've watched teams try to run agentic SEO with just layers 1 and 2. It ships fast. It also drifts fast. Skip the verifier and you get 400 articles that read the same and rank for nothing.
The two sides of Agentic SEO — agents that DO SEO vs. sites that agents CAN USE
Here is the framing every article we read on this topic misses, and it's the one that decides whether your programme is a real revenue lever or an expensive experiment.
Side A — agents that DO SEO for you. Everything in the section above. Planners, executors, verifiers running your workflow. This is the visible half — the vendors selling "agentic SEO platforms" are all on this side. It scales production.
Side B — a site that agents on the other side CAN USE. ChatGPT, Perplexity, Claude, Gemini and Google's AI Overviews all send their own crawler agents to read pages before they cite them. If your site is a JavaScript-rendered black box or a schema-free wall of prose, none of the money you spent on Side A will translate into being picked by Side B. Which is exactly why ChatGPT keeps citing your competitor and not you even after your production volume goes up.
Side B is where 31 of those 40 sites we audited quietly lost. They wired up the sending side and forgot the receiving side. The receiving side needs three specific things — clean semantic HTML, schema.org vocabulary marked up as JSON-LD, and, increasingly, a machine-readable answer surface that agents can hit directly.
That last one is the newest and the one most B2B sites haven't heard of yet. The A2A protocol specification — open, backed by Google and 50+ partners — lets an AI agent skip your marketing HTML entirely and ask your site direct structured questions via a .well-known/agent-card.json endpoint. When a buyer's ChatGPT agent goes to research vendors, an A2A-enabled site gives it a first-class answer while everyone else's site gives it a scraped paragraph. That combination — Side A producing, Side B receiving — is what we've been building as the Agentic Presence Engine we ship.
Where Agentic SEO wins and where it fails
Six months of running audits gives us a reasonably honest read on both.
Wins. Programmatic pages at scale — think 2,000 city-service permutations for a logistics platform — are the strongest use case we've seen. An agentic loop can research each city, adapt boilerplate, insert local schema, and verify against a brand rubric in a fraction of the cost of freelance labour. Internal-linking hygiene is another quiet win: an agent that re-reads your sitemap weekly and proposes anchor updates does in an hour what a specialist would do in a quarter. And measurement — asking a verifier LLM "which of these 20 queries do you cite our client on today?" — is a workflow that simply didn't exist two years ago. It's how you build an A2A endpoint most B2B sites don't yet have and know it's working.
Fails. Agentic SEO fails hardest on YMYL topics (grief, medical, legal, finance) where a missed nuance costs a family or a case. It also fails on original-thesis content — a founder essay, a benchmark study, a counter-intuitive point-of-view piece — because none of that is in the training data yet, and an agent that can't find it can't write it. Google's own guidance on helpful-content signals is direct on this point: first-hand experience, unique perspective, and demonstrated expertise are non-negotiable, and none of the three is something an LLM has by default. The failure mode we see most often: sites that use agents to draft everything instead of using agents to handle everything except the perspective.
Agentic SEO is not a replacement for a strategist. It's a replacement for the 80% of an SEO team's day that was rote — the crawl, the brief, the audit, the internal link, the tracking spreadsheet. The strategist's job gets narrower and much more valuable, not obsolete.
How to start with Agentic SEO in the next 30 days
If you're a founder or head of growth reading this and wondering what "start" looks like, here's the version we walk clients through — no vendor pitch, just the sequence.
- Pick one narrow loop first. Not "agentic SEO for our whole site." Pick internal-linking-only, or SERP-tracking-only, or one page cluster. Ship it, watch it for two weeks, then compound.
- Wire in a verifier from day one. Even a cheap one — a second Claude Haiku call scoring the executor's output against a 5-item rubric. You can't retrofit quality control once you're publishing 200 posts a week.
- Audit Side B before you spend a dollar on Side A. If ChatGPT can't cite you today, publishing more won't change that. Fix machine-readability first — clean HTML, JSON-LD, an A2A endpoint if you're B2B.
- Measure share-of-model, not just rankings. Ask ChatGPT / Perplexity / Claude the 20 queries a buyer would ask before buying from you. Log which brands they name. That's your real dashboard now, and it's the one your board will start asking about in 2026.
If you'd rather not build this stack yourselves, that's the work an AI SEO agency built for this new stack — including us — does end-to-end. But whether you build in-house or hire out, the sequence above holds.
Agentic SEO is only half of the picture. A site that can't be read by AI agents on the receiving side won't get cited no matter how much content its sending-side agents produce. Fix the receiving side first — it's the cheaper half and the one that unlocks everything else.
- arXiv — Generative Engine Optimization (Aggarwal et al., 2023) — the academic paper that first defined GEO.
- A2A Protocol specification — the open standard for AI-agent ↔ website communication.
- OpenAI — Introducing ChatGPT Agent — vendor announcement, July 2025.
- Google Search Central — Perspectives and helpful content — Google's guidance on first-hand experience signals.
- schema.org — vocabulary agents parse when they read your pages.