Evaluating Campaigns — KPIs, Test Duration, Budgets and Sensible Test Setups
Why “running well” and “running poorly” almost never cut it
The classic question in the performance review is: “Is the campaign working?” The classic answer is a gut feeling, garnished with whichever number happens to look good that week. Sometimes it’s CTR, sometimes CPC, sometimes the raw conversion count. The problem: every one of those numbers can be high while the campaign burns money — and every one can look ugly while the campaign is genuinely profitable.
Evaluating campaigns is therefore not a question of your favourite KPI, but of setup: which metric carries the decision, how long does it need to be observed, and is what you’re seeing a signal at all — or just noise? This article walks through the KPIs you actually need (and the ones you can ignore), the question of how long to test, sensible test setups — and the setups that are as widespread as they are useless.
The KPI pyramid — from click metric to profit
There are three layers of KPIs, and they don’t connect symmetrically: you can nail the lower layers and still lose at the top. The other direction rarely happens — anyone profitable at the top usually got something right below.
Layer 1 — delivery and click KPIs
This is where impressions, clicks, CTR, CPC, average position and impression share live. These metrics tell you whether your ads even get served and clicked. They’re useful for diagnosis (“we get no impressions because the daily budget is gone after three hours”), but they aren’t success KPIs. A campaign with 12 % CTR can ruin you if the clicks are all the wrong ones.
Rules of thumb that hold up in practice: a search CTR below 2 % on generic keywords is usually an ad-copy or match-type problem; a CTR above 15 % on brand is normal. A CPC that runs 30 % above account average is almost always a Quality Score problem — not necessarily a bidding problem.
Layer 2 — conversion KPIs
Conversions, CVR, CPA, CPL, CPO. Things get interesting here, because for the first time there’s a connection to business reality. But only if your conversion tracking is clean — and it rarely is without active maintenance. Double-counted thank-you pages, missing cross-device tracking, mis-set conversion actions (“Page Load” instead of “Form Submit”) are the classics.
CVR and CPA are only comparable if the conversion setup is identical across both compared periods. If you start counting a second conversion action (“newsletter”) into the same column halfway through a test, you’re comparing apples to a fruit basket.
Layer 3 — profit KPIs
ROAS, POAS, CAC, MER, contribution margin. These are the only KPIs that answer the question your CFO actually asks: does this campaign make money?
ROAS (Return on Ad Spend) is the most familiar — revenue divided by ad spend. ROAS 4.0 means: every euro invested returns four. Sounds great, but on a 20 % margin product it’s a loss-maker once you account for VAT, returns and warehousing.
That’s why POAS (Profit on Ad Spend) is the more honest metric in any serious setup: instead of revenue, it uses contribution margin. A campaign with ROAS 6 on a high-margin product can be more profitable than one with ROAS 12 on a loss-leader.
In lead-gen businesses, CAC (Customer Acquisition Cost) replaces ROAS — what may a customer cost in acquisition, measured against Customer Lifetime Value? The CAC < CLV/3 rule is conservative but healthy. With recurring-revenue models and high retention, CAC ≈ CLV/2 can also be defended.
Which KPI is the most important?
ROAS is the pragmatic default for e-commerce with homogeneous margins. POAS is the honest version once margins differ across products. CAC/CLV rules in B2B and SaaS. Knowing all three but steering by the wrong one means optimising your way successfully into bankruptcy.
The conversion-volume problem — the underrated one
The most important number when evaluating a campaign is not the conversion rate. It’s the number of conversions your statement is based on.
A typical scenario: campaign A has 8 conversions over two weeks at a CPA of €42. Campaign B has 11 conversions at a CPA of €38. Conclusion: B is better? No. At those volumes, statistical noise is larger than the difference. Three conversions either way are luck, not performance.
Rough rule of thumb: per test variant, you need at least 30, better 50 conversions before any CPA or CVR statement becomes reliable. In small-volume accounts with 20 conversions a month, that means tests run for months, not days. Anyone who refuses to accept that is optimising on noise.
That’s also why Smart Bidding has a data floor: Google recommends at least 30 conversions in 30 days per campaign for tCPA, 50 in 30 days for tROAS. Below that, the algorithm learns on randomness.
How long does a test need to run?
The honest answer: until it’s decided, not until a calendar date is hit. In practice, that means three lower-bound conditions, all of which must be met before you draw a conclusion:
At least one full weekly cycle. Search and conversion behaviour differs systematically between weekdays and weekends, B2B and B2C, morning and evening. A five-business-day test misses the weekend entirely. Plan for two full weeks as a floor, not a comfort zone — that’s the absolute minimum, not a good number.
Enough conversions per variant. See above: 30+ per variant. If your account sees 15 conversions per week in the test campaign, you need four weeks, not two.
No open learning phases. Smart-bidding strategies need 1–2 weeks of learning after any meaningful change (new strategy, target-CPA shift > 20 %, doubled budget). Data from the learning phase is unusable for comparisons — it’s systematically biased because the algorithm is still searching.
That gives you a pragmatic table:
| Account volume (conv./month) | Realistic test duration | What you can test | |---|---|---| | < 30 | 6–8 weeks or don’t | Only blunt levers — strategy, structural shift | | 30–100 | 4 weeks | Bid strategies, ad copy, match types | | 100–500 | 2–3 weeks | Ad variants, audiences, landing pages | | 500+ | 1–2 weeks | Finer levers, fast iteration possible |
Anyone running weekly “A/B tests” in a 30-conversion account isn’t testing — they’re playing roulette and calling it optimisation.
Budget for a test — the simple math
Instead of guessing by gut, the test budget can be derived directly from two known numbers:
Test budget = (target conversions per variant) × (expected CPA) × (number of variants)
Example: you want to test two bid strategies, your historic CPA is €35, you want 50 conversions per variant. Math: 50 × €35 × 2 = €3,500. Plus 15–20 % buffer for learning phases and outliers — call it €4,000.
That’s the honest floor. Anyone starting with €1,500 because “more isn’t available” isn’t running a test — they’re buying anecdotes.
When the test budget isn't there
Better not to test than to test too small. An undersized test produces a number that looks like a result but isn’t — and real budget decisions get made on top of it. That’s worse than no test, because at least gut feel still doubts. A well-grounded heuristic from industry knowledge beats an undersized test almost every time.
Sensible test setups
Not every test is an A/B test, and not every A/B test works in Google Ads. Three setups cover most practical questions.
Drafts & Experiments (Google-Ads-native A/B)
Google Ads offers its own experiment feature: you duplicate a campaign, change one single lever (e.g. the bid strategy) and split traffic 50/50. It’s clean, because Google divides the auction evenly between the two versions — no self-bidding, no double serving.
Good for: bid-strategy switches (Manual CPC vs. tCPA, tCPA vs. tROAS), landing-page tests at campaign level, structural comparisons.
Not for: ad variants within an ad group (Google handles that natively via RSA), date-precise promotions, tests with less than 30 conversions per week of volume.
RSA-internal asset performance
Inside an RSA (Responsive Search Ad), Google tests headlines and descriptions automatically. You don’t need any extra setup — you need enough volume for Google to even rate the asset-performance labels (“Best”, “Good”, “Low”) in the first place (rule of thumb: 5,000 impressions per asset).
Good for: headline variants, hook tests, USP comparisons.
Not for: comparing two completely different ad concepts (use separate ad groups or Drafts & Experiments for that).
Geo-lift / hold-out tests
For brand campaigns, Performance Max and anything where classic A/B tests fail because of tracking: switch the campaign on in region A, off in region B, compare the incremental traffic/revenue. Requires several weeks of baseline data and regions with similar baseline volume.
Good for: incrementality measurement, brand-bidding-yes/no, Performance-Max contribution.
Not for: small accounts with less than 5–10 % market coverage per region, businesses with strong regional seasonality.
Test setups that mostly aren’t worth the time
The most common pseudo-tests in day-to-day account work — and why they fail:
Sequential before/after comparisons. “We had CPA 40 before the switch, now CPA 32 — the new strategy works.” Problem: seasonality, competitor activity, tracking changes, Quality-Score maturation. Almost any change in the account drags along external effects you can’t see. Two weeks before and two weeks after is not a test — it’s a story.
Tests during campaign launch. A new campaign goes through 2–3 weeks of learning phase and Quality-Score maturation. Any test that starts inside that window compares a learning system to a stable one. Result: noise.
Pulling several levers at once. New ads and new bid strategy and new landing page on the same day. If CPA drops — what helped? If it rises — what was it? Three levers at once is three times not a test.
Tests during Black Friday or sale weeks. Seasonal spikes distort CVR and CPA so much that any test effect drowns in the noise. Run tests in standard weeks, never in hot phases.
Tests without a hypothesis. “Let’s see what happens if we go to tROAS 400 % instead of 350 %.” That’s not a test, that’s gambling. A hypothesis should exist in writing before the test: “At tROAS 400 % we expect 15 % fewer conversions but 25 % higher revenue per conversion — net +8 % revenue.” That gives you an evaluation grid. Without it, no result.
Three worked-out examples
To make this concrete — three constellations that come up in accounts every day.
Example 1 — B2B lead gen, 25 leads per month
Account volume: 25 leads/month, CPL ≈ €80. Question: does switching from Manual CPC to tCPA €70 deliver more leads at the same price?
Test requirement: 30 leads per variant × 2 variants = 60 leads. At 25 leads/month total and a 50/50 split = roughly 12 leads per variant per month. Realistic test duration: 2.5 months, budget around €4,800.
Honest recommendation: rather than running an undersized test, switch the strategy on the entire campaign, plan for 6 weeks of learning, then compare against the prior period — knowing it’s seasonally distorted. Or use an external benchmark if available.
Example 2 — e-commerce, 800 conversions per month
Account volume: 800 conv./month, ROAS 5.2. Question: does an RSA with three new headlines deliver more CTR and more revenue?
Test requirement: inside the ad group — asset testing runs automatically. At 25,000 impressions per week and six headlines = roughly 4,000 impressions per headline per week. Asset-performance labels available after 2 weeks.
Honest recommendation: RSA with 8–10 headlines (mix of old and new), replace the “Low”-labelled assets after 2 weeks. Not a classic A/B test, but an iterative optimisation loop — fits the platform’s mechanics.
Example 3 — Performance Max, 200 conversions per month
Account volume: 200 conv./month, of which 60 % via Performance Max. Question: does Performance Max actually add over plain Search & Shopping — or does it just cannibalise?
Test requirement: classic A/B is impossible here (Performance Max can’t be tested against Search inside the same market, since both compete for the same auctions). The only clean route is a geo-lift test: region A with Performance Max, region B without, three weeks of observation, comparison of total revenue (not just the PMax share).
Honest recommendation: lift test with two comparable regions, at least 6 weeks of baseline data, then 4 weeks of test. Budget: an extra €4,000–6,000, because the off-region keeps running, just without PMax. Heavy lifting — but the only method that gives an honest answer.
Reporting cadence — which KPI how often?
A trap many accounts fall into: looking daily at KPIs that only make sense at week or month level. Anyone staring at ROAS daily sees 30 % swings — without anything actually having happened.
| Cadence | KPIs | Purpose | |---|---|---| | Daily | Daily-budget utilisation, spend, anomalies (≥ 50 % deviation) | Early-warning system, not an optimisation trigger | | Weekly | CTR, CPC, CVR, CPA, search terms, top ads | Operational optimisation rhythm | | Monthly | ROAS / POAS / CAC, structural analysis, budget reallocation | Strategic steering | | Quarterly | LTV, contribution to overall margin, channel mix, attribution-model review | Business level |
Daily reporting at conversion level is mostly noise in most accounts. Sensible daily reporting is limited to spend drift, disapprovals and zero-impression alerts.
Conclusion
Evaluating a campaign is not a question of your favourite KPI — it’s a question of setup. Three points carry most decisions: first, know the KPI pyramid, optimise at the profit layer, diagnose at the layers below. Second, accept the volume problem: statements based on fewer than 30 conversions per variant are not statements, they’re anecdotes. Third, pick a test setup that fits the question — sequential before/after comparisons are not tests in 80 % of cases.
In practice that means: before any planned test, answer three questions in writing — what is my hypothesis? how many conversions do I need? how much will it cost? If one of those answers isn’t there, what runs is not a test but a hobby. And for the monthly review: ROAS or POAS as the lead KPI, everything else as diagnostic — not the other way round.
Internalise that and you stop interpreting every weekly swing, and start pulling the levers where leverage is actually measurable.
Entdecke mehr
CPA (Cost per Acquisition)
Advertising KPI: average cost per conversion. Central efficiency metric for lead gen and performance marketing, and the steering input for tCPA bidding.
LexikonUntangling SEA metrics — how CPC, CPA, ROAS, POAS and CIR connect
How CPC, CPA, CPL, CPO, ROAS, POAS, CIR, ROI, CTR and CVR fit together mathematically — with a reference table and concrete worked examples.
NewsBVDW guide 2026: two new KPIs for the AI-agent inbox
Germany's BVDW guide from 18 March 2026 introduces AI Agent Approval Rate and Meaningful Interaction Rate. What the KPIs are worth and what they change.