Guide / optimization
Analyze TikTok Ads Performance with AI: From Reports to Actionable Findings
Build a reviewable read-only analysis with TikTok Ads MCP, align account, dates, time zone, currency, and attribution, then drill down from campaigns to evidence.
This guide turns “review my ad performance” into a repeatable, auditable TikTok Ads MCP workflow. It assumes the MCP connection already works; otherwise, start with the TikTok Ads MCP Beginner Guide.
A complete review usually takes 10–20 minutes. The workflow reads reports and does not change ads.
What you will have when you finish
- A campaign report with account, dates, time zone, currency, and attribution context.
- Comparable current and previous periods of equal length.
- Evidence that drills from the account into campaigns, ad groups, and ads.
- No more than three findings supported by data and ready for human validation.
- An action list that does not execute campaign changes.
Before you start
| Requirement | Example | Why it matters |
|---|---|---|
| Advertiser ID | 1234567890123456789 | Prevents analysis of the wrong account |
| Account time zone and currency | America/Los_Angeles, USD | Defines date boundaries and money units |
| Primary conversion event | Complete Payment | Prevents mixing different outcomes |
| Attribution basis | Use the API response and account settings | Keeps CPA and ROAS comparable |
| Business target | CPA at or below 40 USD | Determines whether a change matters |
| Analysis period | Last seven complete days ending yesterday | Avoids partial data from today |
Do not let the AI invent missing context. Anything you cannot confirm at the start should be labeled “to verify.”
Step 1: Lock the account and reporting context
Start with a small read:
Use only read operations from TikTok Ads MCP. Retrieve the account name, status, time zone, and currency for advertiser ID 1234567890123456789.
Also state the primary conversion metrics and attribution information available for reporting.
Mark fields that cannot be retrieved as "unavailable." Do not infer them or change anything.
Checkpoint: The account name, advertiser ID, time zone, and currency match Ads Manager. Stop immediately if the account is wrong.
Step 2: Define two comparable periods
Use:
- Current period: the last seven complete calendar days ending yesterday in the account time zone.
- Comparison period: the immediately preceding seven complete calendar days.
The periods must not overlap. They must use the same account, currency, attribution basis, conversion event, and aggregation level. Do not include today or compare a seven-day total directly with a 30-day total.
Checkpoint: The output includes exact start and end dates instead of only “last seven days” and “previous week.”
Step 3: Start at the account and campaign level
Collect only the metrics needed to set direction:
| Metric | Purpose | Caution |
|---|---|---|
| Spend | Shows allocation and scale of change | Always include currency |
| Impressions, reach | Shows delivery scale | Do not estimate reach when unavailable |
| Clicks, CTR, CPC | Shows traffic-acquisition change | State the click definition |
| Conversions, CVR, CPA | Shows conversion efficiency | Tie them to a named conversion event |
| Conversion value, ROAS | Shows business return | Do not infer ROAS without value |
Sort campaigns by spend and include an account total. Pulling every dimension at once makes low-spend noise look important.
Step 4: Drill down only on material changes
Group campaigns into:
- High share of spend with declining efficiency.
- Material spend change that moved the account result.
- Improved efficiency with enough volume to study.
Only then inspect ad groups and ads. State why each drill-down was selected, such as “34% of account spend and a 28% CPA increase.”
For creative analysis, compare returned metrics such as spend, impressions, CTR, CVR, and frequency. Do not infer creative content from an ad name.
Step 5: Run the complete read-only analysis
Replace the advertiser ID and target CPA:
Use only read operations from TikTok Ads MCP to analyze advertiser ID 1234567890123456789. The business target is a primary-conversion CPA at or below 40 USD.
Scope:
1. First confirm the account name, advertiser ID, status, time zone, and currency.
2. The current period is the last seven complete calendar days ending yesterday in the account time zone.
3. The comparison period is the immediately preceding seven complete calendar days. The periods must not overlap.
4. State the primary conversion event and attribution basis used.
Layer 1 — account and campaigns:
- Return spend, impressions, reach, clicks, CTR, CPC, conversions, CVR, CPA, conversion value, and ROAS.
- Show current value, comparison value, absolute change, and percentage change.
- Sort by current-period spend descending and include an account-total row.
Layer 2 — material drill-down:
- Select no more than three campaigns that contributed most to the account change.
- Drill into ad groups and ads, and explain why each was selected.
- Show data supporting each finding instead of giving an unsupported label.
Output:
- Begin with account, dates, time zone, currency, conversion event, attribution basis, and filters.
- Mark metrics not returned by the API as "unavailable." Do not guess or treat missing values as zero.
- Separate data facts, analysis hypotheses, and items requiring human validation.
- End with no more than three recommendations. For each, include evidence, risk, and the next human validation step.
Do not create, edit, pause, or delete any campaign, ad group, ad, audience, budget, bid, or targeting setting. Do not execute recommendations.
Recommended output structure
Use this order:
- Query context: account, dates, time zone, currency, conversion event, attribution.
- Account overview: current, comparison, and change.
- Campaign table: sorted by spend with contribution noted.
- Focused drill-down: ad-group and ad evidence.
- Findings: facts, hypotheses, and open checks separated.
- Recommendations: next steps only, with no execution.
Step 6: Validate the important findings
Before accepting recommendations, confirm:
- The periods are equal length and do not overlap.
- Both periods use the same time zone, currency, event, and attribution basis.
- Percentage changes include the base values so a tiny denominator is visible.
- No data, unavailable, and numeric zero are handled separately.
- Totals reasonably reconcile with campaign rows; platform deduplication differences are explained.
- Findings come from returned data, not names or missing fields.
Manually reproduce one or two critical numbers in Ads Manager with the same filters.
Turn findings into controlled actions
Write each recommendation as an approval item:
| Field | Content |
|---|---|
| Finding | What changed |
| Evidence | Account, level, metric, current value, comparison value |
| Hypothesis | Possible cause, not a stated fact |
| Risk | Data delay, sample size, or attribution change |
| Next step | Additional query, creative review, or manual action |
| Approver | Person who must approve execution |
This separates analysis from execution and prevents unsupported campaign changes.
Fast troubleshooting order
A metric is missing
Check whether it applies to the objective and report level. Keep it unavailable instead of replacing it with zero or another metric.
A percentage change is extreme
Inspect the comparison-period base. When it is close to zero, prioritize the raw values and absolute change.
ROAS differs from Ads Manager
Compare conversion value, currency, attribution window, conversion event, time zone, and date boundaries. Do not calculate ROAS when conversion value is absent.
There are too many findings
Limit the answer to the three items with the greatest impact on total spend or the primary KPI, and require contribution data.
The query times out
Fetch the account and campaign overview first, then drill down in batches. Reduce the date range or dimensions instead of resubmitting one large query.
Credential and execution safety
- Use an advertiser ID in prompts; never paste passwords, cookies, or tokens.
- Every query in this guide permits read operations only.
- Reject any write-tool approval and inspect the prompt.
- A recommendation is not approval. A human must review budget, bid, targeting, or status changes.