Claude Code for Marketing Analytics: What I Automate

Claude Code for marketing is most useful on work you can do from a command line: its docs say that while it excels at coding, it can help with anything you can do there, including writing docs, searching files and researching topics (how Claude Code works). In my marketing analytics work that means read-only tag audits of public pages, keyword and results-page research with a spend cap, Search Console snapshots, writing and fact-checking manuals, and render checks, while every commit, push and judgment call stays with me. Every Claude Code detail below was checked against the docs on October 3, 2026.

What do I hand to Claude Code?

Jobs that only read, or that produce something a second check reviews before it ships.

The official prompt library has a marketing tag with broader prompts, such as reviewing a launch post "for unsupported claims, missing attributions, and brand-guideline issues" before it goes to legal. Mine are narrower, because the work is measurement.

How do I run a tag audit with Claude Code?

Claude Code runs commands, so a script can load a public page in headless Chrome, which runs the browser "without any visible UI", and capture every network request the page sends. Claude reads the capture and reports which tags loaded and where their requests went. These audits are read-only research for outreach and never submit a form. If you use the Chrome integration instead, it shares your browser's login state, so it can reach any site you are signed into, which is more access than a public page needs. The manual version of the same check is in checking conversion tags with Tag Assistant and the GTM audit checklist.

Why can a tag audit be wrong?

Because it sees one visitor's version of the site. A fact-checker caught an outreach finding that a Wix site sent consent "denied" and showed no banner. It was true only because the check ran from Bulgaria: Wix sets denied only for EU visitors, which fits Wix's help saying that without a cookie banner, a visitor from a country that requires consent sends no data to Google Analytics. The finding was withdrawn, and checking consent from a non-EU location is now a saved rule in Claude Code's memory. The method is in testing Consent Mode from Europe.

How do I keep keyword research from running up a bill?

Put the limit in the script, not in the prompt. My keyword data script has a cost gate: a run over a set limit is refused unless I approve it. Third-party SEO skills run in this project under a rules block of mine: no paid API call without my approval, and my positioning rules win over the skill's. Read a skill before you run it, because the docs warn that a skill "can grant itself broad tool access" through allowed-tools (pre-approve tools). For a skill with side effects, they recommend disable-model-invocation: true, so only you can start it (control who invokes a skill).

How do I pull Search Console data with Claude Code?

With a plain script that Claude Code runs. Mine calls the Search Console API with a service account on the read-only scope, https://www.googleapis.com/auth/webmasters.readonly (authorization), and the Search Analytics query returns clicks, impressions, CTR and average position for each row (query reference). Each snapshot is saved with its date, so Claude can compare it with the last one. The same pattern sends URL submissions to Bing through the Bing Webmaster API after a deploy. An MCP server is the other route, and the docs say to "verify you trust each server before connecting it" (MCP).

How do I keep API keys away from the agent?

Keep them out of the repo and out of chat, and add a Read deny rule for where they live. The rule stops Claude's file tools and commands such as cat from opening the file, but it does not apply to "a Python or Node script that opens files itself", so the script still reads its key (Read and Edit rules). A ~/path pattern starts at your home directory:

{
  "permissions": {
    "deny": ["Read(~/.config/keys/**)"]
  }
}

It is a seatbelt, not a lock: a command that opens the file another way is not matched, and for enforcement at the operating system level the docs point to the sandbox.

How do I write and fact-check content with Claude Code?

With separate agents for writing and checking. Each article here gets one writer agent, an independent fact-checker that re-reads the official docs and fixes or cuts claims, editor passes, and an agent that renders every page in a real browser at desktop and phone width. One run on October 2, 2026 used 32 agents to write 13 manuals, tighten 5 drafts and render-check 20 pages. The checkers caught a Google Ads claim that missed the newer AI rating of recorded calls and an invented-frequency sentence, "the case I see most". Claude Code can hold this kind of orchestration as a dynamic workflow, a script that runs many subagents and can be saved to .claude/workflows/ and rerun. Why the checker has to be a separate agent is in subagents that verify work.

What turns a rule into a check?

A script that fails. I rejected early drafts as padded, and the answer became a fixed manual format and a saved rule. The rules a script can test sit in scripts/content_lint.py, which fails the run on em dashes, links to removed or missing pages, over-long titles and descriptions, unknown categories, repeated FAQ questions, broken manual structure and banned phrasing; a dated baseline holds known errors in frozen pages. The docs give the reason: Claude treats CLAUDE.md and auto memory "as context, not enforced configuration" (memory). For checks that must run every time, see Claude Code hooks.

What does Claude Code remember between sessions?

Rules I wrote and rules it saved. My personal CLAUDE.md holds working rules, such as never commit or push without explicit approval; the project CLAUDE.md holds positioning and copy rules and imports AGENTS.md with an @AGENTS.md line (imports). Auto memory keeps a MEMORY.md index plus one file per memory, and the first 200 lines or 25KB of the index load every session (auto memory). My saved rules include running the dev server on port 3100, never 3000, and requesting recrawls after every deploy. More in Claude Code memory.

What stays with a human?

Three things. Every commit and push: I approve each one, and agents never push on their own. Any write to an ad account or a client system, such as a conversion upload, which should be tested the way testing a conversion upload safely describes. And every finding that depends on where the check ran or on consent. Dynamic workflows take no input mid-run, and for sign-off between stages the docs say to run each stage as its own workflow (behavior and limits), so an approval belongs between runs, not inside one.

Where to start

Pick one job that only reads, give Claude Code a check it can run, and keep every write with you. If your ad platforms, GA4 and CRM disagree and you want the numbers reconciled by someone who signs off every change, see marketing analytics.