Agentic SEO that ends in a draft, not a dashboard

Most agentic SEO setups have the same ending. The AI pulls live data, applies a careful method and produces a beautifully formatted report. Then someone has to do something with it, and that’s usually where things stall.

This guide is about the other half. It’s a practical, six-step workflow that takes you from live search data to a reviewed draft sitting in your CMS, and judges the result by enquiries rather than impressions. We’ve kept the tooling sensible, flagged where the free tiers run out and been honest about what an agent can’t do for you.

What is agentic SEO?

Agentic SEO is handing a defined, repeatable SEO workflow to an AI agent that retrieves its own data, follows a documented method and asks for approval before it changes anything.

Three parts of that definition matter:

• Retrieves its own data. The agent queries Search Console, your analytics and the live SERP itself, so it’s working from this week’s numbers rather than an export you downloaded in August.
• Follows a documented method. The steps live outside the conversation, so the checks happen whether or not you remembered to ask for them.
• Asks before it acts. It drafts, recommends and flags. A person publishes, redirects and deletes.

If a task is different every time, it isn’t a workflow and an agent won’t help much. If you just need to check 50,000 URLs for a status code, a script will beat an agent every time. Agentic SEO earns its keep on recurring jobs that need judgement: finding opportunities, diagnosing drops, briefing content and checking whether it worked.

Where most agentic SEO setups fall over

The tooling is rarely the problem. The setups we see tend to fail in three predictable places.

There’s no conversion data. An agent can only optimise for what it can see. If your analytics doesn’t record enquiries, bookings or sales, the agent will happily optimise for clicks, and clicks don’t pay invoices. Even Semrush hit this in their own agentic SEO walkthrough: their content portfolio workflow came back largely empty because the test site didn’t have conversion tracking set up. If it can catch out a team that size, it can catch out anyone.

The workflow ends in a report. Analysis is the easy bit to automate. The value appears when an insight becomes a page, an update or a fix. If your agent’s final output is a table, you’ve built a faster way to generate homework.

The stack costs more than the problem. Plenty of guides assume an enterprise SEO suite before you’ve proved the workflow is useful. For a small team, that’s often the wrong order.

So our version adds two things. It checks conversion data before content is briefed, and it finishes by creating a draft in your content platform for a human to edit.

The stack

You don’t need all of these on day one. Each tool has one job and one area where it’s the source of truth.

A few practical notes before you start connecting things.

The Search Console server runs locally on your machine, so it needs the Claude desktop app or Claude Code rather than the web app. Setup is a single command, npx search-console-mcp setup, which handles the Google sign-in. GA4 uses a service account and a separate setup step.

SerpApi’s free plan currently gives you 250 searches a month, which is plenty if you’re checking a few dozen priority queries rather than tracking thousands. It also supports Markdown output, which keeps responses far smaller than raw JSON.

Firecrawl’s free allowance is modest, and accounts of exactly how modest vary. Check the current limits on your account before you point it at an entire competitor site.

If you already use GA4 and trust it, the Search Console server covers it and you can skip Usermaven. We use Usermaven for conversions and attribution because that’s where our goals live. The principle matters more than the tool: pick one analytics platform as the source of truth for conversions, and tell the agent which one it is.

Before you run anything: three prerequisites

1. One conversion goal that means money

Define what a successful visit looks like for this business. For a trades firm, that’s a quote request or a call. For an accountancy practice, it’s a consultation booking. For a coaching business, it’s a paid booking or a discovery call.

If that goal isn’t tracked, fix it before you build anything else. Otherwise every recommendation further down the line rests on traffic alone.

2. A project with your business context

Set up a Claude Project and add short context documents: who you sell to, what you sell, which services matter commercially, who you actually compete with in search, and your editorial rules. Keep each under a page, because they load on every run.

The test for a good context file is whether a new freelancer could read it and avoid your three most common mistakes.

3. Approval rules, written down

Decide what the agent is never allowed to do without you. Ours is short: it doesn’t publish, schedule, redirect, delete, edit live pages or contact anyone. It can read anything it’s connected to and create drafts.

A word of caution: several of these connectors can write as well as read. Search Console tools can submit URLs for indexing, and content platforms can schedule posts. Put “drafts only, no publishing” in your project instructions, and keep it there.

The workflow: six steps from data to draft

Run these steps manually a couple of times first. Once a step produces the same quality of output twice, turn it into a Claude skill so the method is fixed and you aren’t rewriting the prompt each time.

Step 1: Find the opportunity in Search Console

Start with pages you already rank for, not keywords you wish you ranked for. Queries sitting just off page one with decent impressions are usually the cheapest wins available.

The Search Console server has built-in audit types for this, including striking-distance queries, quick wins, cannibalisation and lost queries. Pair it with Ubersuggest so you can see demand and difficulty alongside your own performance.

Using Search Console, find queries for [site] from the last 90 days ranking in positions 8–20 with at least [number] impressions. Group them by landing page. For the top 15, pull UK search volume and difficulty from Ubersuggest. Flag any query where two of our pages are competing.Prompt

What good output looks like: a short list of pages with a clear reason each is worth attention, plus a cannibalisation warning if there is one. If the agent hands you 200 rows, tighten the threshold.

Step 2: Check what the results page actually shows

Rankings don’t tell you what a searcher sees. For many informational queries, Google now shows an AI Overview above the organic results, and the pages it cites aren’t always the ones ranking first.

SerpApi can pull Google’s AI Overview data for a query, including the sources it cites.

For each of the top five queries from step 1, use SerpApi to check whether Google shows an AI Overview for UK searchers. If it does, list the cited sources. If not, list the top five organic results. Use Markdown output.Prompt

What good output looks like: a clear picture of who search engines currently treat as the authority on each topic. That shapes the next step more than any keyword score.

Step 3: Read the competition properly

Now read the pages that are winning. Firecrawl turns them into clean text the agent can analyse without the navigation, cookie banners and footer clutter.

Scrape the three most-cited pages from step 2 with Firecrawl. For each, give me the headings, the question each section answers, any data they cite and its original source, and what’s missing for a [audience] reader. Don’t summarise their prose.Prompt

That last instruction matters. You want the shape of the argument and the gaps, not a paraphrased version of someone else’s article. A brief built from summarised competitor copy produces content that sounds like everyone else’s, because it is.

Step 4: Check what converts before you brief anything

This is the step most guides skip, and it’s the one that stops you writing articles that bring traffic and nothing else.

Ask the agent which pages and channels actually contribute to your conversion goal, and which pages attract visitors who never convert.

In Usermaven, show which landing pages contributed to the [goal name] conversion over the last 90 days using linear attribution. Then list organic landing pages with meaningful traffic but no conversions or assisted conversions.Prompt

What good output looks like: a decision rather than a dashboard. Sometimes the right answer is to improve an existing page that already converts instead of writing a new one. Sometimes it’s to stop investing in a topic that attracts the wrong audience entirely. Both outcomes are worth more than another article.

Step 5: Brief, then draft into your CMS

Only now does the agent write. Have it produce a brief first, check it yourself, then let it draft.

Using steps 1 to 4 and our editorial rules, write a brief: target reader, the problem they’re trying to solve, search intent, recommended headings, questions to answer, evidence needed and internal links. Wait for my approval. Then write the draft, mark every factual claim that needs a source with [VERIFY], and create it in StoryChief as a draft. Do not schedule or publish.Prompt

The [VERIFY] markers are the most valuable part of the draft. They show the editor exactly where the agent is relying on its own assumptions, which is where AI-written content tends to go wrong.

A human still needs to edit the draft. The agent is good at structure, coverage and consistency. It’s less good at original opinion, real client experience and knowing which claims would make your solicitor nervous. That’s your job, and it’s also what makes the content worth citing.

Step 6: Measure against the brief, then loop

Six to eight weeks after publishing, run the same tools in reverse.

Compare [URL] in Search Console (impressions, clicks, average position, top queries) and Usermaven (conversions and assisted conversions) for the six weeks since publication. Judge it against the target in the brief and recommend whether to keep, update or merge it.Prompt

This closes the loop. The output from step 6 becomes the input for step 1 next time, and the workflow gets better because it’s measured against outcomes rather than activity.

What happened when we ran it on this article

We used parts of this workflow to plan the article you’re reading, and the results were humbling in a useful way.

Ubersuggest put UK search volume for “agentic seo” at around 20 searches a month. Global volume was higher, roughly 720 in October 2026, up from about 480 a year earlier. The related queries were more telling: “agentic seo tools”, “seo mcp claude”, “seo mcp server” and “what is agentic seo”. These are people looking to build something, not people browsing a trend.

Then we checked our own Usermaven workspace. We have exactly one conversion goal set up: a form submission. That’s the number this workflow should serve, and it’s a reminder that everything else is commentary.

So we’re not writing this expecting a flood of traffic. A topic with 20 UK searches a month is never going to be a volume play. It’s an authority play, aimed at the small number of marketing managers and business owners who are trying to make this work and would rather talk to someone who has actually done it. That’s a decision the workflow helped us make, and it’s worth more than a keyword score.

When agentic SEO isn’t worth it

Be sceptical if any of these apply:

• Your site has very little search data. If Search Console shows a few hundred impressions a month, the agent has almost nothing to reason over. Spend the time on the fundamentals first.
• You can’t track conversions. Fix that before building anything else.
• Nobody will edit the drafts. An agent publishing unreviewed content at scale isn’t a content strategy. It’s a liability with a schedule.
• The task is a one-off. Writing a skill for something you’ll do once is effort without return.

Copy-and-paste project instructions

Paste this into your Claude Project instructions and replace the bracketed parts.

This project runs agentic SEO workflows for [business name], [website].

Sources of truth
Search performance and indexing: Google Search Console
Conversions and attribution: [Usermaven / GA4], goal: [goal name]
Demand and difficulty: Ubersuggest
Live SERP and AI Overviews: SerpApi
Page content: Firecrawl
Drafts: StoryChief

Rules
Read freely. Never publish, schedule, redirect, delete, edit live pages or send messages.
Create content only as StoryChief drafts.
Show your plan before any multi-step workflow.
Use connected tools for evidence. Don’t estimate missing data; report it as missing.
Separate confirmed findings from hypotheses.
Mark any claim needing a source with [VERIFY].
If a tool call fails, retry once, then tell me.
Judge recommendations against the conversion goal, not traffic alone.
Use British English.Project instructions

The short version

Agentic SEO is useful when it’s structured, connected to real data and judged by commercial outcomes. It’s mostly theatre when it produces polished analysis that nobody acts on.

Start small. Connect Search Console, make sure your conversion goal is tracked and run steps 1 and 4 together. If the agent tells you something you didn’t already know about which pages earn enquiries, add the rest.

If you’d rather have this built around your own data, or you want a second opinion on whether it’s worth doing for your business at all, talk to us about SEO and GEO.