Why Meta Ads reviews consume so much time
Most weekly reviews follow the same shape: compare campaigns, compare ad sets, inspect creatives, check frequency, explain winners, then translate the findings into next actions. None of those steps individually take long, but doing all of them for every account, every week, adds up to hours that rarely show up as "real work" on a timesheet.
That is exactly the type of repeated diagnostic work Claude can compress once it has live access to the account through MCP Ads, rather than working from a CSV export that is already a day stale by the time someone opens it.
How to connect Meta Ads to Claude
The setup mirrors any other platform connection: authenticate once through browser-based OAuth, choose which ad account or accounts Claude should see, and confirm access with a simple verification prompt before moving to anything strategic.
- Connect your Meta Business account through MCP Ads using browser-based OAuth — no app secrets to copy anywhere
- Select the specific ad account(s) Claude should have access to
- Verify the connection with a basic prompt, like spend and impressions for the last 7 days, before running anything more complex
What AI should automate in Meta Ads
The analysis layer is the right place to start. Claude can summarize changes, point to the creative or audience segments that need attention, and package the account story clearly enough to hand straight to a client or stakeholder.
That gives the buyer back time for strategy, testing, and approval instead of account spelunking. The accounts that move fastest are the ones where someone is reviewing daily, not monthly, and that only scales with AI doing the first pass.
- Creative fatigue checks before frequency tanks performance
- Audience and placement comparisons across ad sets
- Weekly account summaries that replace manual dashboard checks
- Client-ready recaps without a separate reporting step
What should stay human
Offer direction, messaging shifts, and aggressive budget moves still need judgment. MCP Ads works best when Claude prepares the recommendation — "frequency on this ad set crossed 4.2, here are three creative angles that haven't been tested" — and the operator decides whether and how to act on it.
That is a practical workflow, not a fantasy of fully unsupervised ads management. The same boundary applies across every platform we connect, including Google Ads: Claude is built for safety, not autopilot.
Worked prompt examples and what comes back
The fastest way to understand what "Claude reads the account" means in practice is to look at a few real prompts and the tools they trigger behind the scenes. You never call the tools yourself — you type the question, Claude picks the tool, and the result lands back in the conversation as a table or a written summary.
"Which campaigns are showing creative fatigue right now?" Claude calls meta_ads_frequency_watcher, which scans every active campaign over the last two weeks, flags the ones whose frequency has crossed the threshold you set (the default is 3.0), and attaches a recommendation per campaign: refresh the creative, reduce the budget, or pause. What you get back is a short list of at-risk campaigns rather than a frequency column you have to sort through yourself. Follow up with "show me the frequency and reach trend for the top one" and Claude switches to meta_ads_frequency for the per-campaign detail.
"Break down the retargeting ad set by placement for the last 14 days." Claude calls meta_ads_adset_performance with a placement breakdown. The response comes back as spend, impressions, clicks, CTR, and cost per result split across Feed, Stories, Reels, and the rest of the publisher platforms, and Claude summarizes which placements are doing the work and which are quietly absorbing budget. The same question at the account level routes to meta_ads_placements instead.
"What changed on the prospecting campaign this month, and who changed it?" This one is meta_ads_campaign_change_history. The result is a dated list of edits — budget changes, status flips, renamed objects, bid adjustments — with the person who made each one. It is the single most useful tool for the "why did performance move on Thursday" conversation, because more often than not the answer is a change nobody mentioned in Slack.
"Give me a one-paragraph account summary for the client." Claude combines meta_ads_overview for the headline spend, CTR, CPC, and ROAS numbers with meta_ads_daily_trend for the day-by-day shape, then writes the recap in plain language. If you want the breakdown by age and gender folded in, say so and it adds meta_ads_demographics to the same pass.
A step-by-step weekly Meta review with Claude
Once the connection is verified, the review below is the routine we recommend running every Monday. Each step is a single prompt, and the whole sequence fits comfortably in one conversation so Claude carries context from one step to the next.
- Step 1 — Account pulse: "Summarize spend, results, CPA, and ROAS for last week vs the week before." Claude pulls the overview and the trend and tells you whether anything moved enough to matter.
- Step 2 — Fatigue sweep: "Flag any campaign with frequency above 3 in the last 14 days and tell me what you would do about each." This is the creative-refresh queue for the week.
- Step 3 — Ad set comparison: "Rank the active ad sets by cost per result and show me the audience and placement each one is using." Look for the ad sets that are winning on a narrow placement and the ones losing on a broad one.
- Step 4 — Creative winners and losers: "Within the top three ad sets, which individual ads have the best and worst CTR and cost per result?" Claude uses meta_ads_ad_performance per ad and lines them up.
- Step 5 — Change audit: "List every change made to the account in the last 7 days." Anything you did not expect goes on the list of things to ask about.
- Step 6 — Recommendation draft: "Based on all of that, write the three actions you would take this week and the reason for each." You review, edit, and either apply the changes yourself or tell Claude to apply the ones you approve.
- Step 7 — Client recap: "Turn this into a short client update, no jargon, with the two things we are doing next." Paste it into email and you are done.
Where automation is safe, and where the guardrails go
Reading is always safe. Every prompt in the previous section is read-only: nothing changes in the account until you ask for a change. When you do want Claude to act, it is worth knowing which tools carry the change so you can scope your instructions precisely.
Pausing and resuming objects runs through meta_ads_update_status, which takes a campaign, ad set, or ad and sets it to active, paused, or archived. Budget changes on a non-CBO ad set use meta_ads_update_adset_budget; campaign-level budgets in a CBO campaign use meta_ads_update_budget. If you want Meta to keep spending inside a range without you watching, meta_ads_set_adset_budget_guardrails sets a spend floor and a hard ceiling per ad set inside a CBO campaign — a good middle ground between "fully manual" and "let the algorithm run."
For genuine unattended automation, Meta's own automated rules are the right tool, and Claude can create them through meta_ads_create_automated_rule. A schedule-based rule that pauses any ad whose frequency exceeds a threshold, or a trigger-based rule that emails you when cost per result spikes, runs on Meta's side even when nobody is in the conversation. The advantage of setting rules this way is that Claude can explain the rule back to you in plain English before it is created, and list the existing ones with meta_ads_list_automated_rules so you do not stack conflicting logic.
Our recommendation is a simple hierarchy: Claude reads freely, Claude proposes changes in writing, you approve the material ones, and only low-stakes hygiene (pausing an ad that has clearly failed, nudging a budget by a small percentage) gets a standing instruction. Everything about that hierarchy is a choice you make; the tools do not force it.
Mistakes to avoid
A few patterns come up repeatedly with teams new to running Meta through an assistant. None of them are hard to avoid once you know to look for them.
- Asking for a verdict before a date range is set. Meta's numbers swing enough day to day that "is this campaign good" without a window is a coin toss. Always name the period, and always ask for the comparison period alongside it.
- Treating frequency as a single number. Account-level frequency hides a retargeting ad set at 8 behind a prospecting ad set at 1.5. Ask for frequency per campaign or per ad set, which is exactly what meta_ads_frequency returns.
- Comparing ad sets with different optimization goals on the same metric. A reach ad set and a purchase ad set will never look alike on cost per result. Ask Claude to group by objective first.
- Reading Meta's attributed conversions as ground truth. Cross-check against GA4 before you move budget on the strength of a ROAS number, especially for anything view-through heavy.
- Letting Claude apply a budget change you have not read. Ask for the recommendation, read the reasoning, then say "apply it." The extra ten seconds is the whole safety model.
- Ignoring the change history. When performance moves, the first question should be "what changed," and meta_ads_campaign_change_history answers it faster than any theory you can form.
Bottom line
The goal is not replacing the media buyer. The goal is letting the buyer spend more time on decisions and less time gathering the context needed to make them.
