---
title: "The Search Console Tool That Does the Math Before the Model Sees the Data"
description: "Tired of LLMs fumbling exports, an SEO veteran built an MCP where the server computes and the model only interprets"
author: "Walter Pine"
published: 2026-09-25T05:47:59Z
modified: 2026-09-26T17:26:29Z
url: https://rews.cc/a/the-search-console-tool-that-does-the-math-before-the-model--697024
language: en
tags: ["llm", "seo", "mcp", "clickhouse", "data", "tech"]
publisher: "Rews (https://rews.cc)"
---

# The Search Console Tool That Does the Math Before the Model Sees the Data

*Tired of LLMs fumbling exports, an SEO veteran built an MCP where the server computes and the model only interprets*

By Walter Pine · September 25, 2026 · https://rews.cc/a/the-search-console-tool-that-does-the-math-before-the-model--697024

## In brief

- Bobbink says wrapper MCPs hand raw Search Console rows to LLMs, which truncate data and mangle aggregations
- Google’s API caps rows, filters anonymized queries and keeps only 16 months of history
- GSC Wizard stores data in ClickHouse on dedicated hardware and backfills Google’s full 16-month window
- Server-side analysis returns conclusions — “30 relevant rows instead of 30,000 raw ones”
- The tool targets enterprises with content groups, multi-property comparisons and fixed-cost infrastructure

Every SEO Jan-Willem Bobbink knows has run the same experiment in the past year: connect Claude or ChatGPT to Google Search Console, ask “why did my traffic drop last month?”, and wait for magic. What arrives instead, he writes on Medium, is a model that pulls 1,000 rows, sums the wrong column, misses most of the long tail because it never made it into the context window, and confidently recommends that you “improve your meta descriptions.”

Bobbink has been building websites for 30 years and spent much of the last decade on sites with more than 50 million URLs. His diagnosis is architectural, not a matter of model quality. Nearly every Search Console MCP — the connector standard that lets a language model talk to an external service — is a thin wrapper: a tool that forwards your parameters to Google’s Search Analytics API and hands back raw JSON, in the hope that the model will behave like a data warehouse. It does not.

## Four ways a pipe fails

He lists four failure modes. First, the context window, not the API, is the bottleneck: a mid-sized ecommerce site easily has hundreds of thousands of query-and-page combinations per month, so the wrapper truncates, the model sees the head of the data, and every conclusion about the tail is a guess. Second, LLMs are bad at aggregation: ask for a weighted average position across 5,000 rows or a click-through delta between two periods per URL and “you will get plausible numbers that are wrong” — often enough that they cannot go in a client report. Third, the API’s hard limits: caps on rows per property per day, anonymized queries filtered out, and only 16 months of history, which makes year-over-year analysis on a 17-month window impossible. Fourth, amnesia: every prompt triggers fresh API calls, re-downloading data you already had yesterday and burning quota and tokens.

The conclusion he reached was that an MCP for Search Console “should not be a pipe. It should be an analyst.” His is called GSC Wizard.

## The inversion

The stack is, in his words, deliberately boring where it can be and specialized where it has to be. ClickHouse — a columnar database built for billions of narrow rows like date, query, page, device, clicks, impressions, position — is the analytical core, with tables partitioned and sorted around the questions SEOs actually ask. It runs on a dedicated Hetzner machine rather than metered cloud, which keeps his infrastructure cost fixed no matter how many properties users connect. TypeScript covers the API and MCP layer, Supabase handles auth, Coolify and Traefik handle deployment, and the whole thing is a remote MCP server with OAuth: no local install, you sign in with Google inside Claude or ChatGPT, where GSC Wizard also exists as an app. Once a property connects, it backfills the full 16 months Google allows and keeps syncing, so your history stops expiring on Google’s schedule. Bing Webmaster Tools, Yandex Webmaster, GA4 (including LLM referral traffic), Google Merchant Center and IndexNow all feed the same layer.

Then the division of labor inverts: the server computes, the model interprets. His example is content decay. The find\_decaying\_content tool runs a statistical analysis in ClickHouse over the full page history, scores every URL on the site, and returns the pages with a significant, sustained decline, the numbers already computed — “30 relevant rows instead of 30,000 raw ones.” That buys correctness, because a database computes the z-scores, regressions and forecasts rather than a language model approximating them; completeness, because the long tail gets analyzed even though the model never sees it; token efficiency, cutting context use by orders of magnitude; and composability, since the model can chain ten compact tool results in one conversation without running out of room. “The model is excellent at the part it should do,” he writes. “It is terrible at being a spreadsheet.”

For enterprise sites — his reference point is classifieds and marketplaces with tens of millions of URLs — the same design carries the load. Whether a property holds 10,000 URLs or 50 million, the tool call looks the same and returns a result of the same size; the heavy scan happens on 256 GB of RAM and NVMe, not in the chat session. Users can define content groups, topic clusters, tags and saved filters, compare ccTLD and subfolder properties side by side, load redirect maps to compare pre- and post-migration performance per URL pair, and produce shared reports with per-client access control. Because a customer connecting 500 properties doesn’t blow up his fixed-cost bill, he says, predictable pricing is possible on the customer’s side too.

On Hacker News, where the post was shared, one commenter, nuc1e0n, arrived at the same conclusion from another direction: “It seems LLMs are better at writing SQL queries than doing the data summarizations themselves. Makes sense as SQL queries occupy much less space in a context window.”

Bobbink calls the result, “in my admittedly biased opinion,” the best Search Console MCP you can use in 2026. The bias is declared. The arithmetic, at least, is no longer the model’s problem.