Incrementality Testing for Lead Gen: Run a Test on CRM Revenue

If you only need the arithmetic, the incrementality calculator takes test and control counts and returns the lift, iROAS and a significance read. If you are deciding whether to test at all, start with incrementality vs attribution.

Attribution assigns the conversions that happened to the touchpoints before them. An incrementality test estimates the conversions that would not have happened without the ads, by comparing a group that could see them with a group that could not.

Three numbers come out of every test. The lift is the relative increase. The incremental conversions are the absolute count the ads added. Divided into the spend, they give the incremental cost per conversion; divided into incremental revenue, the spend gives iROAS. The glossary entry on incrementality defines each one, and return on ad spend covers the attributed version iROAS is compared with.

Five designs cover almost every test a marketing team runs. They differ in who does the split, what outcome they can read and how much volume they need.

The user-based platform study is the most rigorous, because the platform randomises individuals and nobody else can. Its weakness for lead gen is the outcome: it reads what the platform's pixel or conversion API reports. A geo design is noisier, as Google says of its own, but it can read any outcome you can count by region, and that is the one lead gen needs. A pause test has no control group at all, so seasonality and everything else that changed that month land in the result; use it to decide whether a proper test is worth running.

Google offers Conversion Lift based on users and based on geography. Both are set up inside Google Ads, and Google says plainly that the feature is not available to every account: you ask your Google account representative. Requirements as read on Google Ads Help on 7 October 2026.

Two details matter for lead gen. Google warns of up to a 17% drop in measured absolute lift in studies with a long conversion lag that run for less than 14 days, and a B2B lead has a long lag almost by definition. And the geography version is the one Google describes as measuring online and offline conversions: if the outcome you care about is a deal in the CRM, that is the study type to ask about, with your offline conversion data by ZIP or city.

Meta runs Conversion Lift tests in its Experiments tool, where the test creates a test group that can see your ads and a holdout that cannot. Meta suggests starting with an account-level test that measures all of your Meta advertising at once.

• Spend minimums are shown in the toolMeta says some conversion lift tests display spend minimums or recommended spend when you create them, and that your account representative knows which minimums apply. It publishes no single threshold on the pages we read, so we print none.
• The Conversions API is required for web eventsConversion lift tests measuring standard web events since 1 October 2021 require the Conversions API, according to Meta's page on Facebook-managed tests. A pixel alone is not enough.
• 80% power before, 90% confidence afterMeta suggests an estimated power of 80% or more before a test starts, and treats a lift result at 90% confidence or higher as statistically reliable.
• Check the outcome eventWhether your test can read an offline or CRM event, such as a qualified lead sent back through the Conversions API, depends on the study type. Meta's help pages we read do not say, so check in Experiments before you plan around it.

Meta also offers an incremental attribution setting, which changes what Ads Manager optimises for rather than running a test; Meta incremental attribution covers it. For a geo test you design yourself, Meta's open source GeoLift package picks the markets and reads the result with a synthetic control.

A lift study answers exactly the question it is given. Give it form fills and it tells you how many form fills the ads caused, which is not the same as how many customers.

• Form fills and qualified leads move apart. A campaign optimised for cheap leads can raise form fills in the test group while adding few qualified leads, because the extra fills are the ones sales rejects. A lift on form fills then overstates the channel.
• The platform never sees the deal. A deal is closed in HubSpot, Salesforce or Pipedrive weeks after the click. Unless the qualified lead and the won deal are sent back to the platform, its lift study cannot read them.
• Retargeting and brand search look best on the wrong outcome. Both reach people who were already on their way, so both collect form fills that would have arrived anyway. Measured on deals, the gap between attribution and lift is usually widest exactly there.

There are two ways to put the CRM outcome into a test. Send qualified leads and won deals back to the platform as conversions, through Meta's Conversions API and Google's offline conversion import, and ask whether your study type can read them. Or run the test yourself and count the outcome in the CRM by region, which works whatever the platform accepts. The rest of this guide does the second.

A geo holdout switches one channel off in a set of regions and keeps it on in comparable ones. Everything below can be done with location targeting in the ad platform and a CRM export.

1. Pick the outcome before anything else. Qualified leads as the primary read, because they arrive within the test; deals and revenue as the secondary read, later. Write down the CRM stage that counts as qualified and do not change it during the test.
2. Make sure the CRM knows the region. Every lead needs a region field you trust: the state, the ZIP code or the city from the form, the billing address or the company's location. The IP's location is a fallback, not a source. If a third of your leads have no region, fix that first.
3. Pick the regions. Use the smallest unit the ad platform can target and the CRM can record: US states or DMAs, German Bundesländer, UK regions. Split them into a test set and a holdout set with similar volume, similar seasonality and no shared big city. Twenty or more units is what GeoLift's own best practices recommend; fewer works, with a noisier result.
4. Record a pre-period. Count qualified leads and deals per region for several weeks with the channel on everywhere. GeoLift recommends four to five times the test duration of stable history. The ratio between the two sets in this period is what predicts the holdout's counterfactual later.
5. Size the test. Estimate how many qualified leads the holdout will see in the test period and check them against the rule of thumb in the next section. If the numbers are too small, test a bigger channel, use more regions or run longer.
6. Run it. Exclude the holdout regions from the channel under test, and only that channel, for four to eight weeks. Keep budgets, bids and creative in the test regions as they were, and keep every other channel unchanged in both sets.
7. Read qualified leads at the end of the test. Expected test result = holdout result times the pre-period ratio. Incremental leads = test result minus expected. Lift = incremental divided by expected.
8. Read deals and revenue once the leads have had time to close. Usually one sales cycle after the test ends. Count the deals that came from leads created during the test, by region, with the same ratio.

The pre-period ratio is the simplest honest way to build the counterfactual. Meta's GeoLift and Google's matched_markets replace it with a synthetic control or a time-based regression, which handles regions that drift apart; for a first test, the ratio is easier to explain to a finance team and a fair start.

Most failed incrementality tests are not wrong, they are too small: the noise between regions is larger than the effect. A rough rule of thumb tells you before you start whether that will happen.

• About 16 divided by the lift squared, per sideFor counts like leads, a test has a reasonable chance of detecting a lift L when each side expects roughly 16 / L² conversions. A 20% lift needs about 400 per side, a 10% lift about 1,600. Regions add noise of their own, so treat this as the floor.
• Four to eight weeks, and at least one buying cycleGoogle recommends more than 14 days for its own studies, GeoLift at least 15 days on daily data and one full purchase cycle. For a lead gen account, four weeks is the realistic minimum for qualified leads.
• Deals take a sales cycle longerA test that ends in week six has its qualified leads in week six and its deals in month four. Plan the deal read before the test starts, so nobody stops waiting.

The arithmetic is unforgiving for B2B. An account with 100 qualified leads a month cannot detect a 10% lift in any reasonable time, and should either test a bigger effect (switching a whole channel off rather than a campaign) or accept a directional answer. The incrementality calculator runs the significance test on your own counts.

An invented B2B services company runs Meta, Google Ads and LinkedIn across 24 regions. It splits them into 12 test regions and 12 holdout regions, records 8 weeks with Meta on everywhere, then switches Meta off in the holdout for 6 weeks. Every figure below computes.

Without Meta, the test regions would have produced what the holdout did, scaled by the pre-period ratio: 240 x 1.25 = 300 qualified leads. They produced 360, so Meta added 60, a lift of 60 / 300 = 20%. For deals, the expected figure is 40 x 1.25 = 50 against 58 actual: 8 incremental deals, a lift of 16%.

With an average deal of €9,000, the 8 incremental deals are €72,000 of incremental revenue on €24,000 of spend: an iROAS of 3.0, half the ROAS attribution showed. The holdout found 60 of the 96 attributed qualified leads (0.625) and 8 of the 16 attributed deals (0.5). Meta was worth running; it was not worth what either Ads Manager or the attribution report said.

Each of these can turn a clean design into a wrong answer, and most are invisible in the result table. Check them before the test and again before you read it.

• Too small to say anythingA holdout with 40 qualified leads cannot tell a 20% lift from noise. Size it first, and say so when a result is directional.
• Contamination between regionsPeople commute, work remotely and see ads targeted at their office's region. Leave a border region out of both sets where you can, and expect location targeting to leak a little.
• Seasonality and one-off eventsA trade show, a press mention or a holiday in one set breaks the comparison. GeoLift's best practices say to keep local marketing constant across test and control markets.
• Sales capacityIf the sales team cannot follow up every lead, qualified leads and deals in the test regions are capped by people, not by ads. Check the response times in both sets.
• Brand search leakageSwitching off Meta in the holdout lowers brand searches there too, so Google brand campaigns quietly lose leads in the holdout. That is part of Meta's effect; do not credit it to Google, and do not change brand bids during the test.
• Moving the goalpostsChanging the definition of a qualified lead, the budget or the creative mid-test ends the test. Write the plan down and let it run.

A single test is not a new dashboard. It is a correction factor for the dashboard you already have, valid until the channel, the creative or the budget changes a lot.

• Calibrate attribution with the ratio. In the example, multiply Meta's attributed qualified leads by 0.625 and its attributed deals by 0.5 when comparing it with other channels. Keep using attribution for daily decisions inside the channel; incrementality vs attribution covers how the two combine.
• Set budget on incremental cost. €400 per incremental qualified lead and €3,000 per incremental deal are the numbers to compare with your target cost per acquisition, not €125 or €250.
• Feed the platform the outcome you tested on. Sending qualified leads and won deals back with their value makes the platform optimise for what the test showed matters. Sending CRM data back to Meta and Google covers how.
• Retest when something big changes. A new market, a doubled budget or a new offer can move the lift. Twice a year for the largest channel is a reasonable rhythm.
• Test the suspicious channels next. Retargeting and brand search are where attribution and lift disagree most, which makes them the cheapest tests to learn from.

For the budget split across many channels at once, a sequence of tests becomes slow, and teams look at marketing mix modelling; MMM vs MTA covers when that is worth it for a lead gen budget, and why lift tests are what calibrates both.

LeadJourney is an attribution platform, not an incrementality tool. It does not run lift studies, geo experiments or marketing mix models, has no holdout feature and does not report by region. What it does is the half of the test that lead gen usually lacks: the outcome. One script on the site or in GTM records every visit server-side on your own domain, first-party, at 95%+ tracking accuracy, captures the click ids (gclid, gbraid, wbraid, fbclid, li_fat_id, msclkid) at the click, and joins the anonymous history to the person at the form fill, call or booking.

Through native CRM integrations with HubSpot, Salesforce, Pipedrive, Zoho and more, the source, campaign and journey are written onto the CRM record, and stage changes and won deals are read back. That gives you two of the three columns in the worked example: the qualified leads and deals attributed to each channel, under any of five attribution models, and the spend pulled from the ad platforms beside them. The region comes from the CRM record, so the geo read is a CRM export.

It also closes the loop a platform study needs. Won deals go back to Meta through the Conversions API, to Google Ads through offline conversion import on the gclid, to LinkedIn through its Conversions API and to Microsoft Ads, each with the deal value. That lets the platforms optimise on deals, and if your Google or Meta lift study type accepts offline events, it gives the study a deal to read instead of a form fill. Consent is respected: a visitor who declines is not measured. Setup takes about 21 minutes, with a 14-day free trial.

The definitions: incrementality, attribution model, view-through conversion and return on ad spend. The decisions around a test: incrementality vs attribution, Meta incremental attribution and MMM vs MTA. The arithmetic: the incrementality calculator and the ROAS calculator. Why platform numbers run high: click-through vs view-through conversions and cross-channel attribution. On the product side: offline conversion tracking, multi-touch attribution and B2B marketing attribution.