Claude for paid media: what the DIY stack does, and where a marketing team outgrows it

Use it for one person's questions

For one person with a clear question, the stack is quick: no export and no SQL. Write the request so it filters the data and changes nothing:

• “Which search terms spent over €150 in the last 30 days with no conversions? Group them by intent and draft negatives. Don't apply anything.”
• “What changed in our Meta account this week, and who changed it?”
• “Turn last week's numbers into a client summary, and list every figure you couldn't verify.”

One person writing on one platform, with that platform's own limits switched on, may need little more on approvals. Of the six limits below, limit 4 grows with a second write platform or a second person; the other five apply to anyone. Prompts and setup for each platform are in the guides to the Meta, Google Ads and GA4 servers.

Six limits that cost a team money or time

1. On a broken day, Google can spend up to twice a daily budget before anyone reads an alert

Google can spend up to twice a campaign's average daily budget in a day, and Meta up to 75% over a daily budget (how each platform overspends). That is normal pacing. When a pixel breaks, the spend continues while bidding gets no conversion signal, until someone looks.

A scheduled check is the DIY answer, and it can pass without doing the work. Claude's cloud scheduled tasks run at most hourly, in the account of whoever set them up, and Anthropic's documentation says a run's status shows “how the session ended, not whether Claude completed the work your prompt describes”.

A threshold check is quick to write, in a prompt, a script or the platforms' own rules. An alert a team keeps acting on needs its own baseline for each market and season, rare false alarms and the most likely cause named. For twelve markets on four platforms, that is at least 48 baselines for every KPI before seasons are counted, each needing an update when campaigns change.

The model adds variance of its own. Atil et al. ran five models ten times on the same tasks at temperature 0 with a fixed seed, and accuracy moved by up to 15% (2024 study). A performance marketer rules out tracking first, then delivery, the auction, creative and demand. A skill can list that order, or bundle a script that runs it, but the model still decides whether to run the script and how to read what comes back. It can stop at the first plausible cause and report it as confidently as a checked one.

2. Budget moved on platform credit can move the wrong way, in either direction

Each ad platform's server returns that platform's own attributed conversions, which aren't what the ads caused. For a brand as well known as eBay, a geographic test on non-brand search measured a short-term return of −63%, where observational data had shown over 1,400% (Blake, Nosko and Tadelis, Econometrica 2015). The gap can run the other way too: in an August 2025 analysis of 640 Meta tests, Haus, which sells incrementality tests, found 7-day-click reporting under-crediting Meta on average.

Changes inside a platform also pass for performance. Switching a Google Ads conversion action to data-driven attribution moves credit between campaigns, and changing which actions count as primary changes what the Conversions column counts, while orders stay the same. Only data the platform doesn't control catches it: your shop's orders first, then GA4 (why Google Ads and GA4 differ). The assistant makes that comparison only when someone connects the data and asks, every time.

3. A checkout bug read as a channel problem gets the wrong fix

When checkout breaks over a weekend, Google, Meta and email conversions fall together. That pattern points to the site, checkout, tracking or demand. Only a drop confined to one platform sends you to its pixel and tags. Cutting budget or swapping creative in that moment treats the wrong thing.

You can add several servers to one chat. Each reports under its own attribution settings, Meta per ad set and Google per conversion action, and the model lines them up again in each conversation, from whatever rules a project or skill holds. Nothing runs that comparison unless someone asks.

Promotion calendars, stock plans, margins and targets mostly sit in the shop, ERP and finance systems. Merchant Center gives Google prices and availability, and catalogues give Meta the same, but neither holds the plan behind them. Pasted into a chat or a project, the plan serves one person until it goes out of date.

4. Nothing tells a 50% budget jump from a 5% one, or holds one ceiling across platforms

A team usually wants rules like “budget moves up to +15% go ahead, larger ones wait for the head of performance”, or “total spend stays under €60k this month”. Claude's and ChatGPT's approval settings can't express either; enforcing them means writing your own server or hooks.

In Claude, approval is set per tool, whatever the size of the change: always allow, needs approval or blocked (Claude Help Centre). ChatGPT decides when to ask from the app's permissions and the action's context. Meta's admin rules can block budget edits, or budgets above an amount, on Meta alone, and no setting limits the size of one change. Supermetrics' campaign management, in early access, adds approval levels by change type and revert; its guide mentions no spend cap. If only Meta writes, Meta's own limits cover a lot. They stop at Meta's edge, so a second write platform, or one ceiling for all of them, falls outside. Claude's cloud scheduled tasks start in auto mode by default, where a classifier decides on actions that would otherwise wait for your approval, so keep write tools out of unattended runs.

Platform change logs record who changed what. The reason and the approval stay in someone's chat, and when that person leaves a Claude Team or Enterprise organisation, the others lose access to it.

5. A connector bug can turn growth into decline, and someone has to notice

A popular Search Console server reported growth as decline in some period comparisons; the bug was reported on 30 July and fixed in mid-September (issue 42). The platform's API, the MCP server, the SDK it is built on and the AI client each update on their own schedule, and each can break a working setup. In June a user found Meta's server rejecting every reporting call (GitHub). In July a new MCP SDK release crashed fresh installs of Google's GA4 server (issue 199).

The client layer fails more quietly. An open bug report says that when a Claude custom connector's sign-in expires, its tools drop out of the conversation with no prompt to reconnect (issue 702). ChatGPT keeps a frozen snapshot of a connector's tools, so calls can fail after the server changes, and OpenAI says admins “are not proactively notified when an app needs review” (OpenAI Help).

One was fixed within a day; others took weeks or are still open, and every fix began with someone noticing.

6. A bigger account means a bigger API bill, or scheduled runs that stop

One question costs little. Watching every market, campaign and KPI means running the data through a model on every pass, aggregated or raw. On the API the bill grows with the data. On a subscription, scheduled runs share the chat allowance, and once it runs out, further runs are rejected until the window resets, unless usage credits are switched on (Anthropic).

Large reads get set aside. Claude Code moves text results over 50,000 characters into a file the model has to read back, and caps what lands in the chat at 25,000 tokens by default, a limit you can raise (Claude Code docs). A summary of part of the data looks complete; the GA4 guide has a published case.

Six platforms run an official server, and three can change campaigns

MCP (Model Context Protocol) is the open standard that lets an assistant call outside tools. The DIY stack is an assistant plus one server per ad platform.

Google Ads, Meta, GA4, TikTok, Amazon Ads and Microsoft Advertising documentation; LinkedIn's Marketing API changes and Google's hosted MCP list. Checked 25 October 2026.

Acting on what the assistant finds takes a different route on each platform. Meta, TikTok and Amazon write through their own servers, each with its own sign-in and token lifetime. Google Ads needs a third-party server holding a write token, or code that calls the Google Ads API. Google's own Ask Advisor (beta) makes changes inside Google Ads with your approval, apart from your assistant.

What Claude and ChatGPT add

Claude Help Centre on connectors, custom connectors and skills; Anthropic's scheduled tasks docs; OpenAI Help on custom MCP servers, projects and tasks, and OpenAI's skills guide. Checked 25 October 2026.

Skills and Claude Code for paid media

A skill is a folder of instructions and scripts, with a SKILL.md file, that Claude loads when a task calls for it. Public skill packs for paid ads bundle audit checklists, scoring rules and report formats. A good one makes runs more consistent by writing down the definitions and the order of checks. Claude still reads it as instructions, so a step can be skipped, and two runs can read the same data differently.

Claude Code adds what a chat lacks. It runs servers on your machine, such as Google's, works on large exports in code, and runs headless with claude -p from scripts, CI or a scheduled job (Claude Code docs). Run that way on a server, it is an in-house build, with the upkeep described below.

Settle six things before a team or agency connects it

1. Who may connect which accounts. On Claude's Team plan, only Owners add custom connectors, and each person then signs in with their own access. A connector set up with a shared credential, such as an API key, reaches whatever that credential can access, for everyone who uses it (Claude Help Centre). In ChatGPT Business, only admins and owners create and publish connectors.
2. One login, many clients. Meta's connector inherits the user's access to the business portfolios picked at sign-in, including any client ad accounts shared with them, as Jon Loomer points out. A Google manager-account login reaches every client below it, and the model picks the customer ID on each query. Use a sign-in per client, or a sub-manager holding only the accounts in scope, and agree it with each client.
3. What a write may do. Owners on Claude Team and Enterprise can set each tool to always allow, needs approval or blocked for the whole organisation, and members can't override it. ChatGPT Enterprise admins choose which actions an app may take, and new actions arrive switched off. Add the platforms' own controls: Meta's admin rules on every ad account, and a Read-only user behind any Google token.
4. Offboarding. Claude's scheduled tasks “belong to your individual claude.ai account” and aren't shared with teammates (Anthropic). When a member leaves a Claude Team or Enterprise organisation, the others lose access to that person's chats and private projects, while projects shared with the organisation stay (Claude Help Centre). The Meta and Google sign-ins are that person's grants, so revoke them on the platforms as well, and rebuild any scheduled checks in someone else's account.
5. Client data and the AI provider. Every answer passes account data to the AI provider under your plan's terms. Anthropic's consumer terms let each Free, Pro and Max user choose whether chats can train its models. Team, Enterprise and the API run under its commercial terms, where that choice doesn't apply (Anthropic). OpenAI says it doesn't train on Business, Enterprise, Edu or API data by default (OpenAI). Connected services process data under their own terms. Some client contracts restrict AI tools or require consent, so read yours; contract questions go to your counsel.
6. A shared decision log. For every change, record the date, what changed, why, the data behind it and who approved it, somewhere the whole team can read it. Our change approval policy template has a format to copy.

Settling them leaves seven recurring jobs for named owners: a connector admin in each AI workspace, a sign-in per person and per client, tool settings per connector, platform rules per ad account, an offboarding routine, a data-terms check per client, and a log kept by hand. At 40 clients, each buyer's AI account holds its own sign-ins to the clients they handle, scheduled checks sit in several people's accounts, and the log works only if everyone writes in it.

Build it in-house?

A team with a data warehouse can take the assistant out of fetching the data. Fixed SQL on the warehouse removes the errors that come from the assistant choosing dates, filters and row limits, paging through results or summing rows in the chat. The assistant only explains numbers the SQL has already fetched. For reporting, that is a sound setup.

A quick build can stall. At a listed company with billions in revenue, the data team built an optimisation tool, mostly with Claude, and delivered it quickly. The search team barely uses it: most of its suggestions aren’t useful, improving them needs the team’s feedback, and other priorities always win.

Clean warehouse data leaves the hard parts in place:

• Baselines per market and season, tuned until false alarms are rare.
• Choosing and maintaining root-cause checks that run in the same order every time, across channels and the shop's orders.
• Acting within guardrails: a write integration per platform API, limits per campaign and action type, a spend ceiling across channels, and a stop when a conversion feed goes stale.
• An approval trail with the reason for each change, which outlasts the people who made it.

Three ways to run it, side by side

Check these before connecting Claude to an ad account