Internal tool · Automotive retailSEO + GEO · AI agentsTrial run · November 2026

SEO + GEOPlanner.

One place for a car dealer network’s search team to see which 50 pages to work on this month, why, and who is doing what. Data from five tools and two AI research agents, turned into a plan people can act on.

Monitor→Research→Plan→Report→Team decides
RoleProduct owner and integrator: workflow, requirements, data plumbing, testing
Built withAI coding agents (Claude Code), directed by me
UsersThe company’s SEO team
StatusTrial run; used to plan the month’s 50 pages
01 / ProblemFive tools, no shared view

The data was there.The plan wasn’t.

Every number the team needed existed somewhere. Nobody could see in one place which pages to work on, why, and what had already been done. Almost nothing was written down, and SEO reports arrived as PDFs no one could act on.

  • AhrefsSearch demand and rankings
  • Search ConsoleClicks and impressions
  • GA4Traffic and leads
  • Google TrendsWhat’s rising
  • Checked by handCompetitor pages and Google’s AI Overviews
02 / How the data worksHover or tap any box to trace it

Seven sources in,one plan out.

Sources
Collect
Agents
Outputs
People

The planner’s score weighs search volume, how close a page is to page 1 and its Search Console clicks, plus a bonus from the research findings, with quotas for product, non-product and local pages. GA4 traffic and leads are shown for context but never scored.

03 / What I builtThe part the team actually opens

The agent suggests.People decide.

The first version felt like a copy of the spreadsheet. I pushed back and asked for a clean dashboard instead: cards and charts, fewer words, and the URL, category and keyword cluster visible on every row. The team works across six pages: Overview, Monthly plan, Keyword bank, Reports, Downloads and History.

Key decisionThe agent’s score is never overwritten. Human decisions live next to it, with a reason and a name.
50 / 50Nothing changes silently

Pages can be added, removed or replaced, always with a reason. A counter keeps the plan at 50, and the full history is kept.

Keyword bankBuild a page from keywords

Every keyword with volume, rank, rank change and whether it triggers an AI Overview. Changes are queued for the agent.

Honest labels“Not available” is not zero

Missing data says so, estimates are labelled, and caveats show where they matter, like partial trial click data.

The real app · anonymized dataScroll to walk through
01 · Overview

The month at a glance.

Pages in plan, progress by status, what still needs review, and the biggest gaps.

02 · Monthly plan

Fifty pages, each with an owner.

The agent’s top 50 by URL, with category, keywords, score, SEO and GEO signals, and estimated extra clicks.

03 · One page, opened

Why it’s on the list.

The reason in plain words, the keyword cluster, live rank history and what to change.

04 · Keyword bank

Every keyword, one table.

Volume, rank, rank change and AI Overview presence, for product, non-product, local and Search Console queries.

05 · SEO report

Where competitors beat us.

The research agent’s report as a dashboard instead of a PDF.

06 · GEO report

Who Google’s AI cites.

AI Overview citations by keyword, and which sites are cited instead of us.

Fictional brand and dealer names; private traffic figures are blurred. Team names, decisions and rank checks are demo entries.

Open the demo ↗
04 / The GEO angleBeing cited, not just ranked

Search now hasa second front page.

Trial run · keywords checked
33keywords checked
25showed an AI Overview
16cited a competitor, not us

GEO, generative engine optimisation, means being cited in Google’s AI Overviews, not only ranking in the blue links. I built it into the same workflow as SEO instead of a separate report.

  • 01For each keyword: is there an AI Overview, who is cited and in what order, and are we cited, cited below a competitor, or not at all.
  • 02Cited sources are sorted into groups, so the team sees which other dealers are being cited instead of us.
  • 03Our page is compared with the cited page (dated price tables, direct answers, FAQ, images), and the gaps become actions.
  • 04GEO findings feed the same score and the same dashboard, so SEO and GEO sit side by side for every page.
UsOther dealersBrand sitesMarketplacesSocial / other
05 / My roleI directed the agents; I didn’t hand-write the code

Putting it together,and making it usable.

  • DesignThe end-to-end workflow

    Monitor, research, plan, report: what each agent receives, what it outputs, and where a human decides.

  • RequirementsThe rules

    Which metrics matter, the scoring logic, the 50-page limit and quotas, calendar-month data windows, a full Ahrefs refresh only every 3 months to save budget.

  • IntegrationThe plumbing

    Search Console and GA4 through Apps Script into Sheets, the email sender, the tracker sheet, Ahrefs and Trends.

  • CoordinationAgents with separate jobs

    Several AI agents with distinct roles, kept from overwriting each other’s work.

  • TestingReal data, real problems

    Caught a feed that covered only about 30% of clicks and a sheet tab-name restriction, and made the call on fixes.

  • ProductUsable by the team

    Hosted online with a database and a shared password, so the team opens a link and installs nothing.

Python · FlaskSQLite · PostgresPlain JavaScriptApps Script · SheetsAhrefs · Search Console · GA4 · TrendsClaude agentsAutomated tests
06 / ResultA trial, reported as a trial
StatusNovember trial run

The full loop ran end to end on real data, from the data pull to the plan in the app.

UseThe month’s 50 pages, planned

The team used the app to review the agent’s picks and plan the month’s 50 pages.

Not claimedNo traffic or revenue gains

It’s too early to measure outcomes from one trial month, so this page doesn’t claim any.

07 / What I learnedFor the next tool
  • A dashboard is a set of decisions.The spreadsheet-style first version had every number and helped nobody choose. Showing fewer things, with the URL and reason up front, is what made it usable.
  • Keep the machine’s answer and the human’s apart.Storing the agent’s score untouched and the team’s decisions separately made the plan easy to trust and easy to audit.
  • Honest gaps beat confident zeros.Labelling missing and estimated data, and naming caveats like partial trial clicks, mattered as much as any feature.