Attribution
Marketing Attribution — who gets credit for the sale
Attribution says who got credit. Incrementality says who actually drove the decision. Attribution is fast and cheap, incrementality is slow and expensive. Most companies use both.
Marketing concepts
Incrementality testing: what would have sold without your ads
Incrementality testing is an experiment that measures how much revenue would arrive even if you didn't run an ad. The only method that measures actual cause-and-effect instead of assigning credit after the fact.
Principle
Incremental lift = revenue with ads − revenue without ads (in a control group)
Example
You run Meta Ads in NYC but not in LA. After four weeks: NYC $60,000 in revenue, LA $40,000. LA is naturally 30% smaller. If you didn't run ads in NYC, you'd expect a similar drop. Actual lift from Meta Ads: ~$15,000, not the full $60,000 attribution claims.
Heads up: incrementality is the only truth. But you can't do it fast, cheap, and for every channel at once — you pick where it hurts most.
Why attribution isn't enough and what incrementality does differently
Attribution tells you "campaign X got credit for a $100 sale." But it doesn't tell you whether that sale would have happened anyway. Maybe the customer was going to buy regardless. Maybe they'd have found the brand via Google. Maybe a friend reminded them. Attribution doesn't capture any of this.
Incrementality testing answers a completely different question: "What would have happened if I hadn't run this ad?" It's a classic controlled experiment — like in medicine, where half the patients get the drug and half get placebo so you can see the drug's real effect.
In marketing, the "control group" is a group of customers (or regions) where you deliberately don't run ads. After a few weeks you compare revenue in both groups. The difference is your true incremental lift. It's usually smaller — sometimes much smaller — than what attribution claims.
Main types of incrementality tests
There are five main methods for measuring lift. They differ in cost, speed, and accuracy:
Principle
Create two groups — control (no ads) and exposed (with ads). Compare results.
Five tests you should know
01
Geo holdout
You turn off a channel in selected regions (say, Boston) and keep it running everywhere else. After 4–8 weeks, you compare revenue.
Cheapest + most accurate if you have at least 3 comparable regions.
Good for ecommerce with multi-region footprint. Doesn't work if you sell purely online globally — no regional differences.
02
Conversion Lift Study (Meta)
Meta splits your audience into exposed and control. Neither you nor Meta shows ads to the control group. After 4–6 weeks, Meta sends you a report on actual lift.
Free from Meta, but typically requires higher spend and a Meta Sales contact — roughly $30,000+/month.
Great for Meta Ads. Unfortunately not publicly available to everyone — requires minimum spend and often a Meta sales contact.
03
Ghost bidding (Google + Meta)
Bid on a keyword or audience as usual, but the ad isn't shown (Google does this for Search lift studies). It tests whether the auction even mattered.
Requires a sales contact at Google or Meta. Not for small accounts.
Works mainly for large advertisers. Most marketers can't access it.
04
Brand vs no-brand test (Search)
Pause your brand campaigns in Google Search for 2–4 weeks. Watch how much brand search traffic ends up on your site for free via organic results.
You find out how much of brand spend was cannibalizing organic.
A classic experiment. Often reveals that 70–90% of brand search would happen without paid ads.
05
Synthetic control / statistical model
Instead of a regional holdout, you model "what would have happened" using historical data and a control group of similar companies or periods.
Requires a data analyst or specialized tool.
Least accurate, but available when you can't run a classic A/B test. Suitable for brands with a single region.
Which test fits your situation
Pick a test based on size, channel, and how much measurement you can stomach:
| Situation | Recommended test | Why |
|---|---|---|
| US ecommerce, multi-state, multi-channel | Geo holdout | Pause in a single state, compare 4–8 weeks |
| D2C brand with $30k+/mo on Meta | Conversion Lift Study | Meta builds the test for free, delivers a report |
| B2B with long sales cycles | Brand vs no-brand search test | Short test reveals how much brand search is subsidizing agency fees |
| Large advertiser ($100k+/mo) | Ghost bidding or MMM | Enough data for a statistically credible model |
| Small ecommerce (under $5k/mo) | Pause test (4 weeks) | Turn off the most expensive channel and watch MER |
| Global ecommerce with no regional differences | Synthetic control | Geo holdout doesn't work; model from history |
Rule of thumb: run incrementality tests on your most expensive channels (where you spend most). Small channels aren't worth it — the cost of the test outweighs the savings.
When incrementality misleads
Incrementality is the most accurate method, but not infallible. Four situations where results can mislead:
Seasonality overwhelms the effect
If you run a test in December against November, you'll mostly measure seasonal variation — not the ad effect. The test must run during a stable period or in parallel across comparable groups.
Short measurement window
Ads have a long tail. A brand campaign you run in January will bring customers in March. If you measure only 4 weeks, you'll underestimate the real lift. For long decision cycles (B2B), you need 3–6 months.
You can't test everything at once
If you turned off Google Ads, Meta Ads, and email at the same time, you'd see the impact of none. Tests must be sequential, one at a time. That means months of work to understand the full mix.
Channel cannibalization
If you pause Meta Ads, some people who would have bought from Meta shift to Google Search or direct visits. Meta's measured lift looks smaller than it is. That's why you measure not only the paused channel but the others too — to see if they "picked up" the slack.
Related concepts and when to use them
Incrementality alone isn't enough. Metrics and methods that complement it:
MER
Marketing Efficiency Ratio — whole-marketing return
MER tells you whether the whole marketing engine is working. Incrementality then tells you which specific channel is pulling. MER is the quick indicator, incrementality the deep dive.
What is MER →MMM
Marketing Mix Modeling — statistical model
MMM models each channel's contribution from historical data. Less precise than classic incrementality but cheaper and faster. Often combined: MMM for understanding the mix, incrementality for the key channels.
ROAS
Return on Ad Spend
ROAS uses attribution. If you measure ROAS 4 but incrementality shows real lift is 30%, your true "incremental ROAS" is 1.2. Big gap.
What is ROAS →Three mistakes most companies make
- 01
Treating test results as final
An incrementality test is a snapshot in time. An ad with 30% lift today might be 50% or 10% half a year from now. Tests must be repeated — at least annually for key channels.
- 02
Running short tests and panicking
A four-week test shows noise, not effect. If you don't see at least a 10% difference between groups and at least 100 conversions in each, you don't have a statistically credible result. An 8-week test beats two 4-week tests.
- 03
Testing what's cheap, not what hurts most
Incrementality tests cost (lost revenue in the control group). It makes no sense to test a $500/mo email channel when you spend $20,000 on Meta Ads. Test where the biggest risk and biggest spend are — even if it's painful.
Frequently asked questions
- How long should a test run?
- Minimum 4 weeks, but most channels need 6–8 weeks. For brand campaigns and B2B, 3–6 months. The test must be long enough to capture ad tail, but not so long that seasonality overrides the effect.
- How much will the test cost me?
- The main cost is foregone profit in the control group. If you pause a channel that brought $20,000/month for 8 weeks, you're risking around $40,000 in lost revenue. It's worth it only on channels where you actually spend more than you'd risk.
- Can I run a test myself or do I need an agency?
- Geo holdout and brand vs no-brand tests you can do yourself — Google Sheets and a calendar are enough. Conversion Lift Study Meta sets up for free, but you need a Meta sales contact. Synthetic control and MMM usually require an analyst or specialized tool.
- How often should I repeat tests?
- Test key channels at least annually. When your mix changes significantly (new platform, major creative refresh, new target audience), test immediately. Lift isn't constant — it shifts with competition, season, and how well customers know you.
- What if the test shows zero lift on a channel?
- Unfortunately it means the channel did nothing for revenue that wouldn't have happened anyway. Reallocate budget, turn it off completely, or fundamentally change targeting and creative. Zero lift is the harshest feedback in marketing — which is why companies are reluctant to run these tests.
Want to know if your ads are actually pulling?
Lupli connects all your marketing data and shows you daily ROAS, MER, attribution, and anomalies. Plus we're building a template for your own geo holdout tests.
- Real view across all channels in one place
- MER as a safety net when attribution misleads
- 30 seconds to connect your first account