Claude for Media Buyers: A Practical Guide

Where Claude fits in a media buying week

Most of a media buyer's week is not pressing buttons in Ads Manager. It is reading: competitor ads, your own scripts, comment threads, last week's numbers, creator drafts. That reading is where Claude pays off first.

Here is how the work splits when we map a typical Meta ads week against what a language model can actually do.

The pattern: Claude is strong when the input is text and the output is text. It gets shaky when the answer depends on information it can't see, like auction dynamics, a pixel that broke on Tuesday, or the client call where the CMO hated the blue packaging.

The reading load is bigger than it looks

To see why reading is the bottleneck, look at what a competitor research pass involves. On the Ad Radar board on October 9, 2026 there were 2,443 live Meta ads with at least $50k in estimated spend, from 749 brands.

A median winning video runs 64 seconds, so its transcript is roughly 150 to 200 words. Reading 40 of them, tagging the opener, the angle and the proof, and writing down what repeats takes a person most of an afternoon. Claude does the first pass in a couple of minutes, and the first pass is the boring part.

What it can't do is decide which 40 ads deserve the read. That is a data question: estimated spend, days live, variant count. If you give Claude a random pile of ads from the Ad Library, it will find patterns in noise. Give it ads that have already survived months of spend and the patterns mean something. Our guide on how to find winning Facebook ads covers those filters.

Five jobs to give Claude this week

1. Monday: competitor hook sweep

Paste or pull 20 to 40 competitor ads in your niche, with spend and days live attached, and ask for a grouped breakdown.

Here are 30 Meta ads from competitors in [niche], each with estimated
spend, days live and the full script.

1. Classify each opener into one hook pattern (problem/benefit, number,
   story, question, offer-first, warning, authority).
2. For each pattern: count, median estimated spend, and the 2 strongest
   examples quoted word for word.
3. List 3 openers that appear in ads live 90+ days with 2+ variants.
Do not invent ads. If a field is missing, say so.

The last line matters. Without it, a model will sometimes fill a gap with a plausible ad that doesn't exist.

2. Tuesday: variants on a proven body

When an ad is working, the cheapest test is a new opener on the same body. Claude is good at this if you show it the real winner first.

Ask for openers in different hook patterns, not ten rewordings of the same line. "Ten hooks" without that rule usually returns ten versions of one idea.

3. Wednesday: the brief

Turn the test question into a brief with fixed fields: audience awareness, angle, hook pattern, proof, format, length, landing page, kill rule. Our creative brief template has the fields; how to write a creative brief with AI shows the prompts that fill it from a winning ad.

4. Thursday: script review

Paste your draft script next to two or three winners in the same niche and ask Claude where yours loses the viewer. It is useful at spotting a body that takes 15 seconds to name the problem, or a CTA that introduces a new claim at the end.

5. Friday: the review note

Export your week (spend, CPA or ROAS, hook rate, hold rate by ad) as CSV, paste it, and ask Claude to write the review note your team or client reads. Give it your rules, for example "flag any ad above target CPA for 3 days running", so the note follows your process instead of its own. The structure we suggest is in the weekly creative review.

Connecting Claude to data instead of pasting

Pasting works for a test. It breaks down by the second week, because the data you pasted is already stale and the context window fills with CSV.

The fix is MCP, the Model Context Protocol. Anthropic introduced it in November 2024 as an open standard for connecting AI apps to outside systems, and in December 2025 donated it to the Agentic AI Foundation under the Linux Foundation. An MCP server exposes a set of tools; Claude decides when to call them. We explain the protocol for non-developers in MCP for marketers.

For a media buyer, two kinds of connection matter:

• Your own account. Meta released its Ads AI Connectors, an MCP server plus a CLI, in open beta on April 29, 2026, according to PPC Land. Google's Google Ads MCP server is read-only in its current release.
• The market. Competitor ads, their copy and transcripts, and a signal of what is getting spend. That is what the Ad Radar MCP server does. In Claude Code:

claude mcp add --transport http ad-radar https://app.ad-radar.dev/api/mcp

In Claude Desktop and claude.ai it is a connector with an OAuth sign-in, no token to copy. Claude then has six tools: search_ads, get_ad, winning_patterns, advertiser_hooks, get_swipe_prompt and list_taxonomy. A full session, from first search to finished swipe, is in competitor ad research inside Claude.

A real example of a question worth asking

Here is the kind of ad that rewards a careful read, and how you would put it to Claude.

The prompt:

Using get_ad, pull this ad's full transcript. Then:
- Map the script into beats with timestamps estimated from word count.
- Name the job each beat does (anchor, credibility, mechanism, offer).
- Tell me which beats I could reuse for [my product] without
  copying wording, and which depend on facts I don't have.

That last request keeps the output honest. A comparison ad only works if your comparison is true, and Claude doesn't know your product's numbers unless you give them.

What to keep for yourself

Three things stay with the buyer, whatever the tooling.

1. Budget and status changes. If you connect a write-capable account connector, require confirmation before any change that moves money. An agent that can pause ads can also pause the wrong ones.
2. Claims. Claude will happily write "clinically proven" if a competitor's ad says it. Compliance is yours.
3. The read on results. Claude can summarize a week. It doesn't know that CPA jumped because a creator posted the discount code on Reddit. Context like that lives in your head and your Slack.

A starting setup

1. Connect one data source for the market (competitor ads) and, if you are comfortable with it, a read-only view of your own account.
2. Save three prompts as reusable snippets: the Monday sweep, the Tuesday openers, the Friday note. Our Claude prompts for Facebook ads has 25 more.
3. Always attach the evidence. Ask Claude to quote the ad, the spend estimate and days live for every pattern it claims.
4. Review the first month of output line by line. After that you will know which jobs you can trust it with and which still need a second look.

Figures marked as estimated spend come from Ad Radar's model of engagement on public Meta Ad Library ads. They are estimates, labeled as such, and are best used to rank ads against each other.