AI Agents for Search Console Analysis: Our Workflow, Prompts and Guardrails

How we run Search Console analysis with AI agents and MCP: real prompts, real checks, and where the agent gets it wrong. Copy our workflow and test it.

Last updated: November 1, 2026

Summary

An AI agent connected to Google Search Console can pull, compare and explain your search data in minutes instead of hours. We run Claude connected to Search Console, Semrush, Screaming Frog and a browser through MCP (Model Context Protocol). The agent does the digging. A human makes every decision, and every number gets checked against the source before it leaves the building.

The speed is real. So is the risk: an agent that can’t say “I don’t know” will give you a confident, wrong answer.

What changed since our 2025 Manus test

In April 2025 we tested Manus, the Chinese AI agent, on our own Search Console data and wrote about it here. It was impressive for a single prompt. But each session started from zero, it had no live connection to our tools, and we couldn’t reuse the work.

Eighteen months later the setup looks different. The agent now has persistent tool connections:

The last row matters most. Without memory, the agent re-diagnoses the same problem every session and forgets what you already decided.

Example 1: picking 10 posts to refresh in under an hour

This week we decided to refresh 10 old posts on stradiji.com, five in English and five in Turkish. Every content archive decays, and ours is no exception. Here’s how the agent picked them.

Step 1. Compare two 90-day windows. The agent pulled page-level data for July 29 to October 27, 2026 and for the 90 days before. It sorted by impressions and flagged pages that lost ground.

Step 2. Pull the queries behind each candidate. For each page it pulled the top queries. One example: our English “best paid and free SEO tools” post had 12,303 impressions at an average position of 44 and zero clicks. The queries told the story. People found it for searches like “best free keyword tool 2014.” It was a 2020 post that still read like one.

Step 3. Add market data. Semrush gave US volumes for the realistic targets: “best seo tools” at 9,900 a month (keyword difficulty 62) versus “ai visibility tools” at 1,300 (difficulty 50). That changed the angle of the rewrite.

Step 4. Read the live page. Through Chrome, the agent pulled title length, meta description length, H1 count, internal links and word count for all 11 candidates. The tools post had a 67-character meta description and almost no internal links.

Step 5. Diagnose, then stop. The agent wrote an SEO diagnosis and a GEO diagnosis for each post, plus a list of questions only I could answer. Then it stopped and waited for my approval.

Total time from first prompt to a ranked shortlist: under an hour. Doing the same by hand used to take most of a day.

Example 2: spotting zero-click before it shows up in revenue

Our Turkish “What is a PWA?” post ranked at position 3.4 for its main query over the same 90 days. It got 1,715 impressions and 2 clicks.

A human scanning a dashboard sees “position 3, fine.” The agent flagged the click-through rate as an outlier for that position. The most likely explanation is that Google’s AI Overview answers the question above the results. That’s a hypothesis, not a fact. We’re testing it by moving the definition to the first sentence and tracking AI impressions in Search Console’s generative AI report.

Example 3: when the problem isn’t in Search Console at all

In October, a corporate site we manage showed a strange top country in GA4: a country the business doesn’t serve, with 943 sessions in a month, more than its home market. Only 2.3% of those sessions were engaged, and the average session lasted three seconds.

Search Console had nothing to say about it, because none of it was search traffic. The agent read the server logs instead and found two bot patterns: one scraping the WordPress API with a fake desktop Chrome user agent, the other loading pages as an iPhone. We blocked both at the hosting level.

The lesson for agent workflows: give the agent more than one data source, or it will explain everything with the one it has.

Where the agent gets it wrong

These are the mistakes we’ve caught, not theoretical ones.

1. Counting jump links as pages. Search Console reports URLs with #fragments (for example /post/#section-2) as separate rows. An agent that sums rows overcounts impressions. Our rule: strip fragments before any total.

2. Treating old and new URLs as different pages. If a site has changed its URL structure, Search Console splits history between the old and new address. The agent will happily report that a page “lost 90% of its traffic” when the traffic simply moved. Our rule: map old URLs to new before comparing periods.

3. Inventing a cause. Ask “why did traffic drop?” and you’ll get an answer, always. Sometimes it’s right. Our rule: the agent may propose causes, labeled as hypotheses, with the evidence for each. A cause becomes a finding only after two independent sources agree.

4. Doing instead of advising. Will Scott documented a case where an agent, asked to create pages for target keywords, cloned the homepage into two new URLs with new titles. They sat live for six months with zero impressions. His advice is the right one: use Claude for SEO, don’t let Claude do SEO. Our rule: the agent never changes a live page on its own. It drafts. I approve. Then it applies, and checks the live result.

5. Reading a number wrong. Models misread tables. Every number in a client report gets re-pulled from the tool, not copied from the chat.

Prompts you can copy

These work with any agent that has a Search Console connection. Replace the dates and site.

Refresh triage

Compare page-level Search Console data for [site] between [date range A] and [date range B].
Strip #fragments and merge old and new URLs using this map: [map].
List the 20 pages with the largest impression loss. For each, show clicks, impressions, CTR and position for both periods, and the top 5 queries in period B.
Do not suggest causes yet.

Zero-click check

Find queries where our average position is better than 5 but CTR is below 1%, with at least 200 impressions in the last 90 days.
For each, tell me whether the query is informational, and label any cause you suggest as a hypothesis.

Before-and-after check

For these URLs: [list], compare the 28 days before [change date] with the 28 days after.
Exclude [date ranges with known tracking issues].
Report the change only if both periods have at least 100 impressions.

How to set this up

1. Search Console MCP server. There are open-source servers that connect an agent to the Search Console API. Run it on a machine you control and connect it with your own Google account.
2. One market-data tool. We use Semrush’s MCP. Pick one, not three: every extra tool makes the agent slower and the answers noisier. (Our tool stack post explains why we chose Semrush.)
3. A project memory file. Decisions, open issues, pages not to touch. The agent reads it at the start of every session.
4. A verification checklist. Run it in a separate session, with the agent acting as an auditor rather than the author. A model reviewing its own work is too kind to itself.

Frequently asked questions

Can ChatGPT or Claude read my Search Console data?

Not by default. You need a connection, usually an MCP server or an integration that uses the Search Console API with your permission. Once connected, the agent can query your data directly instead of reading screenshots.

Is it safe to give an AI agent access to Search Console?

Read access is low risk if you run the connection yourself and use your own account. Write access to your site or CMS is a different story. Keep a human approval step for every live change.

Will an AI agent replace an SEO consultant?

It replaces the export-and-filter part of the job. It doesn’t replace judgment about what to do next, and it doesn’t take responsibility when the answer is wrong.

The one thing to do this week

Run the zero-click prompt above on your own site, even by hand in Search Console. Any query where you rank in the top five but get almost no clicks is a page where AI is answering for you.