amekala/ads-mcp/adspirer-performance-review

Performance review

Follow adspirer-mcp before any tool call. This skill only reads — it never changes a campaign.

Where the numbers come from

The two big platforms have direct, top-level performance tools. The rest go through their router.

Calling the routed ones directly by name will fail — they aren't in the tool list. Pass raw_data: true when the user wants the numbers rather than the rendered summary.

Only pull the platforms the user has actually connected. Check get_connections_status first instead of firing six calls and reporting five failures.

Before you trust a single number

Verify conversion tracking before you trust a CPA. Run audit_conversion_tracking for Google, Meta, and LinkedIn. It does not cover Microsoft. It looks back from today. Read freshness_state on the performance pull first. If that window had no delivery, a "no conversions" finding is not proof the tag is broken — say the window had no delivery.

Advertisers don't all use a website tag. A GA4 import, an offline or API upload, calls, or a platform lead form is judged by that source: last conversion, upload health, whether the action is primary. A tag chip that says Urgent, or a Data manager with no source listed, only matters for actions that use the tag. Two primaries for the same event (tag and GA4 both counting one purchase) double-count. On Google, compare Conversions with All conv. — a wide gap means bidding is ignoring most of the activity. Primary versus secondary is set in the ads UI, not via Adspirer.

If the audit ignores customer_id ("You have N google_ads accounts connected. Please specify..."), retry once, then the error catalog in adspirer-mcp. The UI read, when you need the method the API can't show, is adspirer-ads-manager-browser.

If tracking for the primary goal really is broken, say that before you present a CPA. "How are my ads doing" is then "I can't tell you yet, and here's why."

Also check the window. Amazon's last 48 hours are incomplete. Today is always partial. A "drop" that is really an unfinished day is the most common false alarm in this job. If a platform returns no_data, look at the freshness window, look back only as far as that window (not past a year), and report the range that has data. Label it. Don't write 0.

The scorecard

Default period when they say "this week": the last full Monday–Sunday week against the week before. "This month": the closed month against the month before. Compare like with like, not against a partial week.

Report per platform and account. Follow the evidence rules in adspirer-agent.

• Spend — and pace against the monthly budget, in that account's currency
• Conversions and cost per conversion, using that platform's own conversion name and

attribution window

• ROAS, where revenue is tracked
• CTR and CPC — diagnostic, not goals. CPM when the goal is awareness
• Change vs the previous equivalent period

Derived CTR, CPC, CPA, and ROAS come from totals and are labeled derived. If a breakdown (keyword, ad group, ad set, ad) doesn't add up to the account total, mark it unverified and don't recommend from it alone.

Google per-campaign clicks and impressions come from analyze_wasted_spend when get_campaign_performance doesn't include them. Say where that tool's target ROAS came from (account goal, a 90-day average, or its default). If no source has clicks, write "not available". Don't back into them.

Microsoft Advertising has no performance tool unless microsoft_ads list_tools has grown one. Show campaigns, status, and budgets (adspirer-microsoft-ads) and put metrics under what you could not read.

GA4 and Search Console, when those routers are connected, are context only (adspirer-google-analytics, adspirer-search-console). Never add their events or organic clicks to ad conversions.

An inactive account that refuses a read goes under what you could not read. Ask before activating it.

Totals only within one currency. Never sum conversions across platforms as if they were deduplicated. A missing input is "not available", never 0. On a quota or tool error, say which call failed. get_usage_status before and after a wide pull — don't assume every lookup is free.

End with a "could not read" list: every platform, account, metric, or range that was unavailable, and why.

Lead with the biggest mover. Two to four next steps, as questions. Several proposed edits go to the review table in adspirer-optimize. Anything that increases spend goes to adspirer-launch. Offer adspirer-dashboard when a chart will help. Email or Slack is adspirer-digest, after a yes. A scheduled run that finds no new data since the last report gets one line, not a full scorecard.

Say what it means

A table is not a review. Lead with the answer:

Google is doing the work — 34 leads at $41, under your $50 target. Meta spent $890 for 3 leads.

The Meta ad set has a frequency of 4.2 and its CTR halved over two weeks; that's creative fatigue,

not targeting.

Name the biggest problem and the biggest opportunity. Give one recommendation, with the number attached. Offer to act — don't act. Changes belong to adspirer-optimize.

When something moved sharply

Use the anomaly tools rather than speculating: explain_performance_anomaly (Google), explain_meta_anomaly, explain_tiktok_anomaly, explain_linkedin_anomaly. They compare windows and attribute the delta to a cause.

Before blaming the algorithm, check the boring explanations: a campaign paused, a budget changed, a disapproved ad, a landing page that started 404ing, a conversion tag someone removed on deploy.

Make it easier to look at

A scorecard across platforms is a table the user will squint at, and a trend is a shape, not a column of numbers. For an actual chart, load adspirer-dashboard.

Offer an artifact: platform cards, a spend-vs-conversions chart, the anomalies called out. It's a private page that updates in place, so a weekly review can republish to the same URL instead of scrolling back through the transcript. Keep charts as SVG or CSS rather than embedded images.

Make it recurring

A review is worth more every week than once. When the user likes what they just read, offer to schedule it — this host can, and they can create it just by asking.

In Cowork, that's a scheduled task via /schedule — it runs remotely, so a sleeping laptop doesn't matter. In Claude Code, it's a routine; prefer a remote one, because a local Desktop task is skipped whenever the machine is asleep at the scheduled hour — exactly when a Monday-morning review would fire.

Each run costs tool calls against their monthly quota, the same as a live conversation. Weekly is the right default; daily only for accounts spending enough to justify it. Say so before setting it up.

If the user wants it delivered by email instead, or this host has no scheduler, Adspirer's monitoring_and_reporting router does server-side scheduled briefs. See references/host-surfaces.md in adspirer-agent.

Don't

• Don't average CPA across platforms with wildly different volumes and call it "blended performance"

without saying so.

• Don't present a 3-day window as a trend.
• Don't fabricate a number when a tool fails. Say the call failed and which one.
• Don't schedule a daily review for an account that spends $20 a day. The runs cost more attention,

and more quota, than the findings are worth.