Gains:
- Ability to understand performance metrics such as CTR, conversion, CPA, ROAS and attribution and use artificial intelligence to summarize data and draft insights
- Ability to anonymously give campaign data to artificial intelligence and produce comments, anomaly detection and optimization proposal drafts
- Understanding that it is mandatory to compare and verify every number and causality sentence produced by artificial intelligence with the source data
Once a campaign is live, the real work begins: understanding what it does. Measuring the ad's performance is essential both to shift the budget to the right place and to better set up the next campaign. But measurement is not a pile of numbers; Good analysis finds the story and reason behind the numbers. This is called insight. Artificial intelligence is a great accelerator at this stage: it summarizes hundreds of lines of campaign data in minutes, points out anomalies (unexpected deviations), writes draft reports. But the immutable rule of this unit is this: Every number and every causality sentence produced by the AI does not enter the report until it is compared and verified with the source data. Language models can produce fluent but made-up numbers when they do not fully access the data or calculate it incorrectly.
Key performance metrics
Several metrics form the language of performance analysis. CTR (Click-Through Rate) is how many people who saw the ad clicked on it (clicks ÷ impressions). Interest is the measure of attractiveness. Conversion rate is how many clickers take the desired action (purchase, registration). CPA (Cost Per Acquisition) is the cost per acquisition (spend ÷ number of acquisitions). ROAS (Return on Ad Spend) is the return on ad spend: revenue earned ÷ ad spend; For example, if you spent 10,000 TL and earned 40,000 TL in revenue, ROAS is 4. There is also attribution: the issue of which ad/channel to attribute a conversion to. If a customer first saw an Instagram ad, then searched on Google and bought, which channel will you post this sale on? The citation model makes this decision and directly affects the analysis.
The following table summarizes the key metrics and what they say:
metric
formula
what does it say
CTR (click-through rate)
Clicks ÷ Impressions
Does the ad attract attention?
Conversion rate
Conversion ÷ Click
Does anyone who clicks buy it?
CPA
Spending ÷ Acquisition
How much does a customer cost
ROAS
Income ÷ Expenditure
Is the expenditure profitable?
frequency
Exposure ÷ Reach
How many times did the same person see it?
Tip: Don't make your decision based on just one metric. High CTR but low conversion means "the ad attracts but the page/offer does not hold". Reading the metrics together gives the real story.
Step by step: from data to insight to report
Step 1 — Collect and anonymize data. Get data from platform dashboards (Google Ads, Meta) and analytics; Remove personal data (customer ID, email).
Step 2 — Calculate and validate metrics. Confirm the odds yourself; Specify the formula when giving it to AI.
Step 3 — Look for anomalies and patterns. Have the AI ask “which channel/day/segment is deviating from what is expected”; then confirm it with data.
Step 4 — Establish causality, but cautiously. Support “CTR dropped because…” sentences with data; Don't confuse correlation with causation.
Step 5 — Draft report and recommendation. Generate executive summary and optimization proposal draft with AI; Verify and confirm every number and claim.
three mini cases
Case 1 — Quick summary, verified issue. A performance expert anonymously gave 300 lines of weekly data from 6 channels to AI and requested a draft executive summary. AI summarized in 5 minutes; The expert compared each ROAS and CPA value to his dashboard, correcting two erroneous interpretations. Duration: 1 hour instead of half a day. AI gave the draft, verification remained with the human.
Case 2 — Made-up number trap. A team said to the AI, "our ROAS missed last month, comment" without giving raw data. "About 3.5, good performance," the AI wrote confidently. The actual ROAS was 1.8 and the campaign was actually at a loss. The report was corrected, but customer confidence was shaken. Error: waiting for metrics from AI without giving data.
Case 3 — Attribution fallacy. One brand attributed all sales based on last click, typed all sales into Google search, and cut its Instagram budget. Sales fell the following month; because Instagram created awareness and fueled search. Blindly relying on the single citation model produced the wrong decision. Lesson: attribution is a model, not absolute truth; It is necessary to compare different models.
Four copyable templates
1) Data summarizer (with validation emphasis):
Below is anonymous campaign data (no customer ID):[paste data/table]Task: (1) Calculate and display CTR, conversion rate, CPA, ROAS with their formulas.(2) Point out the 2 best and worst performing channels.(3) Clearly state where you are unsure or missing in the data.Do not make up any numbers that are not in the data. I will confirm the account.
2) Anomaly/pattern hunter:
Below is daily data for the last 4 weeks: [data].Task: Detect unexpected deviations (sudden drop/rise).For each anomaly: which day/channel, how much it deviated, possible explanations (hypothesis).State that these are hypotheses, not definitive causes.
3) Causality checker:
The observation I have is: "[e.g. CTR dropped from 2% to 1.3%]". Task: List possible reasons for this drop (ad fatigue, season, targeting change, competitor activity, etc.). Tell me what data I can test against for each reason. Don't present correlation as causation; Do not claim definitive cause.
4) Executive summary + proposal draft:
Metrics I verified: [CTR, conversion, CPA, ROAS values].Task: Write a 1-page executive summary draft:(1) what happened this term (3 items), (2) what worked/didn't work, (3) 3 optimization suggestions for next term.Just use the verified numbers I gave; generating new numbers.
Weak prompt / Strong prompt
Weak prompt:
Evaluate my campaign's performance and tell me my ROAS.
No data given; AI can only make up “ROAS”. The output is smooth but unrealistic.
Powerful prompt:
Your role: analytics assistant to a performance marketing specialist. Below is anonymous weekly data (no names/identification): Advertising spend: 25,000 TL. Income: 60,000 TL. Views: 500,000. Clicks: 12,000. Conversion: 480.Task: (1) Calculate and display CTR, conversion rate, CPA and ROAS with formula.(2) Write a single paragraph comment draft.(3) Don't make up any numbers that are not in the data; State that you are not sure.
In this request, the expectation of data, formula and the prohibition of "making up" are clear. You again verify the output: CTR = 12,000 ÷ 500,000 = 2.4%; conversion rate = 480 ÷ 12,000 = 4%; CPA = 25,000 ÷ 480 ≈ 52 TL; ROAS = 60,000 ÷ 25,000 = 2.4.
Common mistakes
- Requesting metrics without providing data. AI cannot access your data; The number he gives is fake.
- Looking at one metric. If CTR is high and conversion is low, the problem may be with the page/offer, not the ad.
- Confusing correlation with causation. Just because two things act together does not mean that one causes the other.
- Relying on the single citation model. Attribution is a model; Different models produce different decisions.
- Putting AI's causality statement in the report without confirming it. “Because…” sentences must be supported by data.
- Mistaking the raw number for insight. The report does not list numbers; tells you what it means and what to do.
Caution: The most dangerous sentence in a report is an unsubstantiated "because." AI can come up with a fluent reason; If you present that reason to the management without proving it with data, you may lead to a wrong decision.
In summary
Performance analysis is the task of converting campaign data into decisions, and at its core are metrics such as CTR, conversion rate, CPA, ROAS and the attribution issue. Artificial intelligence is a powerful assistant in summarizing hundreds of lines of data, finding anomalies and drafting reports. But every number and every causality sentence it gives is not included in the report until it is verified with the source data; The language model can adapt when it cannot access the data. Reading metrics together, separating correlation from causation, and not tying attribution to a single model are signs of good analysis. The report produces meaning and suggestions, not a list of numbers.
Application task
Prepare real or hypothetical campaign data (spend, revenue, impressions, clicks, conversions). (1) Anonymize data. (2) Have the metrics calculated with the "Data summarizer" and confirm each one yourself with the formula. (3) Find a deviation with the "anomaly hunter"; connect it to a hypothesis. (4) With the “causation checker,” write down the possible causes of this deviation and how you would test each. (5) Produce a draft of the “executive summary” and check that it contains only the numbers you have verified; Support a “because” sentence with data.
checklist
- [ ] I anonymized the campaign data before giving it to the tool.
- [ ] I verified each metric with its formula myself.
- [ ] I read the metrics together, not one by one.
- [ ] I supported the causality statements with data; I didn't confuse it with correlation.
- [ ] I did not attribute the reference to a single model, I compared it.
- [ ] I put only verified numbers in the report; There are no fake numbers.