Check the campaign numbers before you write the story.

Use this for the last numerical check before a report goes out.

You have a weekly export and a summary headed for a client, manager or business owner. Before explaining performance, check that the summary uses the same records and denominators as the export. This kit produces a review file you can keep with the report; it does not change a campaign or send a campaign report.

Start with the fictional sample to see the complete job. Then use an approved copy of your export, retain its stable account and campaign identifiers, and remove totals or overlapping breakdowns. The tool can flag repeated keys, but it cannot know whether two differently named segments describe the same impressions. Google documents that some segmented report totals are deliberately unavailable because adding the rows would double-count data.

If your reporting platform already gives you a dependable report with a review process, you may not need another tool. This checker is most useful when numbers have been copied, combined or rewritten, including by an AI assistant, and you want the calculation and exceptions to remain visible.

Choose the reporting basis before calculating.

The required detail fields are date, spend, impressions and clicks, plus columns that identify a row. Use the supplied headers or provide a field-to-column CSV map. Conversions and revenue are optional. If an optional value is missing, the affected currency total and related rates stay unavailable; the tool never treats an empty revenue cell as zero.

Date and currency are added to your selected row key. For an account-level campaign export, account_id,campaign_id may be enough. If the file includes several platforms or an additional non-overlapping dimension, include those columns too. Both identical and conflicting repeated key groups are excluded together so you can resolve the structure instead of accepting an arbitrary version.

1. Confirm that every row represents a mutually exclusive slice of the report. If you cannot confirm this, choose the option that withholds totals.
2. Choose the inclusive start and end dates and the detail date format. Check the original report's timezone and whether conversions are assigned to interaction date or conversion date; the checker cannot inspect those settings.
3. Map a supported currency code per row. For an export with no currency column, explicitly declare one verified currency. Keep different currencies separate; do not infer USD from a dollar sign.
4. Add the summary CSV only if you want to compare an existing claim. Use one row per currency, matching YYYY-MM-DD period dates, and only the metric columns you intend to check.
5. Run the check, resolve the findings and retain the downloaded review with the source export. Refreshing the page does not save your pasted inputs.

Worked answer: 1% and 2% do not average to this report's CTR.

The first fictional campaign has 10 clicks from 1,000 impressions: 1% CTR. The second has 180 clicks from 9,000 impressions: 2%. Taking the average of those rates gives 1.5%, but the second campaign supplied nine times as many impressions. The combined rate is 190 ÷ 10,000 × 100 = 1.9%.

The August 17–23 sample has USD 1,000 spend, 20 supplied conversions and USD 3,000 supplied revenue. Its cost per click is 1,000 ÷ 190 = 5.263158; cost per supplied conversion is 50; its click-based conversion rate is 10.526316%; and revenue divided by spend is 3. The included summary agrees on all supplied totals and disagrees only on CTR, producing one review finding.

These are fixed arithmetic reference answers. They are not observed client performance, proof that an AI product passed a test or an assessment of whether the campaigns were good. Change a sample input and the tool recalculates the result; the answer key documents the original inputs only.

A difference is a question for review, not an automatic verdict.

If an in-period row has an unresolved key, date, currency or required metric, the remaining figures are labeled as a subtotal. Complete-report summary comparisons are withheld. Rows outside your declared period are accounted for separately. A zero denominator produces an unavailable rate, not zero percent or an infinite return.

The result records the checked value, reported value and reported-minus-checked difference. Counts must match exactly. Conversion counts and ROAS allow a 0.000001 difference; monetary fields and percentage-point rates allow 0.01. Those are this tool's disclosed rounding tolerances, not advertising-industry standards.

Metric names can conceal different definitions. This tool's conversion rate divides supplied conversions by clicks. Google Ads can use eligible ad interactions, and cost-related conversion metrics can have their own eligibility rules. Its generic spend-per-click calculation also needs the matching spend scope. Compare definitions before treating a mismatch with a platform as an error; attribution, reporting delays and conversion settings are outside this check.

• Download the complete check for the calculation basis, metrics, comparisons and every finding.
• Use the normalized-row file to trace included, excluded and outside-period records back to the input.
• Keep the metrics, issues and basis files together when handing the report to another reviewer.
• Do not describe a subset as the full campaign result, or use an arithmetic pass as a budget recommendation.

Let AI draft the explanation after the numbers are checked.

The useful AI step is a bounded writing task: turn reviewed metrics and the exception list into a short draft for the intended reader. Give the assistant the metric definitions and the reporting period alongside the numbers. Use only fictional or approved data in an approved environment; the local checker does not grant permission to share an export with another service.

Try this brief: Draft a short campaign update from these checked metrics and this exception list. Preserve the period, currencies and metric definitions. Separate observations from possible explanations. Mark unresolved rows. Do not invent a prior-period comparison, target, cause or budget recommendation. Return a draft for review; do not send it.

Check the draft against the exported values, not against how plausible the prose sounds. If you want a week-over-week change, supply a separately reviewed prior period with the same scope. If the assistant supplies a reason performance changed, ask which evidence establishes that reason. This practice teaches a reusable business skill: defining what an assistant may summarize, what it may calculate and what still needs judgment.

Keep the check small enough to inspect.

Each supplied CSV is limited to 1 MB of UTF-8 data and 10,000 data rows. Supported numeric fields are nonnegative plain or explicitly selected US-formatted numbers, with at most six decimals and no more than 1,000,000,000,000 per field per row. Negative adjustments, comma-decimal locales and unsupported currency codes need a different preparation or review step. Malformed CSV and explicit records containing only empty cells stop the check.

The tool runs locally in the browser without a model call, account connection or transmission of the CSV contents. It neither saves your files nor reads live advertising accounts. The visible tables show at most 100 rows each; the downloadable check retains all output records. Source row references count parsed CSV records with the header as row one, ignoring blank lines.

Passing these rules does not prove that the source is complete, conversions are correctly attributed or the campaign achieved a business outcome. Preserve the source report and its settings. If you need scheduled extraction, account permissions, attribution reconciliation or large-scale joins, use the appropriate reporting system and its review process.