Conversion Lift in Google Ads 2026: Incrementality for Search and Performance Max
Your Google Ads report says 1,240 conversions last month. The one question that report can never answer: how many of those happened because of the ads, and how many would have happened anyway — the person was going to buy, and would have arrived through organic search, a bookmark or a direct visit. The gap between those two numbers is incrementality, and measuring it is exactly what Conversion Lift in Google Ads 2026 is for — the conversions report cannot answer it in principle.
On 7 October 2026, Google opened user-based Conversion Lift to Search and Performance Max campaigns. Until then, those two campaign types needed a Google account representative to set a study up, which in practice meant only large accounts ever ran one. Now any advertiser who clears the thresholds can launch it from the interface. Here is what the thresholds are, how to size a holdback, how long a study has to run, how to read iCPA and iROAS, and what to do when the answer is worse than you hoped.
Conversion Lift in Google Ads 2026: how it differs from attribution
Attribution distributes conversions that already happened. A data-driven model decides how much credit goes to Search, how much to YouTube, how much to remarketing. Every one of those conversions is already in the bank, and no model can tell you whether it would have landed without the impression. We went through the limits of that in attribution models in Google Ads 2026: DDA vs last click.
Conversion Lift works differently. It is a randomised experiment: a share of users is placed in a holdback group that never sees the campaigns under test. Google compares the two groups and reports the difference as incremental conversions. That is a measurement, not a model estimate.
Google Ads offers three lift study types, and they get confused constantly:
| Study | What it measures | Method |
|---|---|---|
| Brand Lift | Recall, awareness, consideration, purchase intent | Surveys served to exposed and control users |
| Search Lift | Increase in brand or product searches on Google and YouTube | Comparison of search behaviour between groups |
| Conversion Lift | Incremental conversions and conversion value | Holdback by users, or control vs exposed geographies |
This article is about the third one, specifically the user-based variant. The geo variant splits regions instead of audiences; we covered that approach, including building your own matched-market test, in incrementality and geo experiments in 2026.
What actually changed in October 2026
Before the change, self-serve user-based Conversion Lift covered Display, Video, Demand Gen and App campaigns. Search and Performance Max were missing from that list, so those studies ran through a rep — a gate that excluded most accounts by default.
Search and PMax are now on the list. Google’s own documentation never uses the phrase “self-serve” — that framing came from trade coverage — but the practical effect matches it: if your account qualifies, the option appears in the interface.
One caveat worth stating plainly: availability is not universal. Google says Conversion Lift is not available to every account and can vary account by account. The only reliable check is opening the section and seeing whether the option is there.
The three eligibility thresholds
All three have to be satisfied at once.
| Requirement | Threshold | Detail |
|---|---|---|
| Observed conversions | 1,000 minimum | Conversions from supplementary data sources do not count toward this |
| Budget of participating campaigns | US$5,000 minimum | Google does not specify per campaign or combined; non-USD accounts should check the local equivalent |
| Conversion action | At least one compatible action | It has to fire unconditionally, not behind a site-side condition |
The detail that trips people up is “observed”. It means conversions your measurement captured directly. Modelled store visits and uploaded offline leads do not count toward the 1,000 — although you can still include them in the study itself, where they appear with a warning icon. If you are in the middle of building that offline pipeline, the sequence is in offline conversion import in Google Ads.
Two campaign types are excluded outright: App campaigns targeting iOS, and Travel Ads. A campaign can also belong to only one study at a time — if it is already in a Brand Lift or Search Lift study, remove it first. Campaigns showing “Needs attention” cannot be selected at all; they are inactive, already in a study, or incompatible with the chosen lift type.
Where it lives in the interface
Campaigns → Experiments → Lift studies tab → plus button → Conversion Lift → “Based on users”. As of 2026, lift studies sit inside Experiments rather than in a section of their own, alongside ordinary campaign tests. We covered that section and standard campaign experiments in experiments in Google Ads: testing hypotheses in Experiment Center.
Which campaigns to include
The instinct is to include one campaign to limit the downside. That instinct usually kills the study. Google recommends including every campaign of the same type that shares an audience and targeting strategy, for a mechanical reason: if a held-back user does not see campaign A but does see campaign B, they are no longer a clean control. The effect bleeds, and statistical power drops with it.
Performance Max adds a wrinkle: a single campaign often covers several product lines, so “the same audience” inside it is a simplification. Before putting PMax into a study, look at how your asset groups and listing groups are arranged — if one campaign holds segments with incompatible economics, an averaged lift number will not tell you much.
Holdback: how many people you are willing to not advertise to
The control group is configurable from 1% to 50%. If Brand Lift and Search Lift run inside the same study, the holdback is fixed at 30%.
This is a real trade-off and it belongs in a conversation with whoever owns the revenue number, before launch:
- A larger holdback accumulates statistics faster, so the study is shorter and more reliable — but more people never see the ads, and you give up conversions during the test window.
- A smaller holdback barely dents current sales, but the study either drags on or ends inconclusive because there was never enough power to separate signal from noise.
A rough frame: a 10% holdback on campaigns turning over $50,000 a month costs roughly $5,000 of forgone revenue across a one-month study — a directional figure, not a rule, since the real number depends on how much of that revenue was incremental in the first place. That is the price of the answer. If the price is unacceptable, skipping the test is more honest than running it at 1% and reading noise.
Duration, and why seven days is almost always too short
The minimum duration is 7 days. Google recommends more than 14, and longer still when the conversion lag is long. One number worth remembering: studies with long lags compressed under 14 days showed up to a 17% drop in absolute lift — treat that as a directional benchmark from Google’s guidance rather than a constant.
The logic is simple. If you sell B2B and the median path from click to closed deal is three weeks, a seven-day study measures response speed, not incrementality. Before you pick a duration, look at the actual lag distribution in your conversion actions — the “days to conversion” column and the path reports give it to you in a minute.
Study power: the number that decides the study before it starts
The setup screen shows projected study power from 50% to 95% in 5% steps. That is the probability the study detects an effect if an effect exists.
- 90% or above — the target. A result at this level can carry a budget decision.
- 50–90% — directional only. Useful as a hypothesis, not as grounds to cut or double a channel.
- Below 90% — the tool tells you what budget would get you there. That is free guidance, and ignoring it is the most expensive way to spend two weeks.
Clean measurement is half the result
Conversion Lift compares two groups on conversions. If conversions are counted badly, the comparison inherits the problem. Google recommends having these in place first:
- Enhanced conversions for web and for leads — hashed first-party data recovers matches that would otherwise be lost. Setup is in enhanced conversions in Google Ads.
- Consent mode, including advanced consent mode — without it, a slice of your audience never enters the measurement at all. See Consent Mode v2 and privacy in 2026.
- Google tag gateway — server-side signal collection that reduces loss to blockers.
Saving the study triggers automated diagnostics. You can save despite warnings, but a warning means exactly what it says: part of the signal is not arriving, and the lift number will be biased by that gap. A related read on how Google scores the quality of your first-party setup is Data Strength in Google Ads 2026.
Reading the results
Depending on the report focus you choose, you get one of two metric pairs:
| Focus | Metrics | Typical fit |
|---|---|---|
| Conversions | Incremental conversions and iCPA | Lead gen, subscriptions, installs |
| Value | Incremental value and iROAS | E-commerce and anything with variable order value |
Pay attention to cool-down — the projection of delayed incremental conversions that arrive after the study ends. It is included in results by default; you can view results without it through the metric tooltip or the reporting dashboard. The gap between the two versions is a useful diagnostic in itself: if lift halves without cool-down, weekly readings on this channel were never meaningful.
When iCPA comes back higher than your reported CPA
That is the normal outcome of a first study, not a failure. Incremental CPA is always higher than attributed CPA, because some of the conversions in your report would have happened anyway. The question is not “why is it worse” but “does this iCPA still fit the allowable CPA my unit economics support”. The arithmetic for that is in unit economics and LTV for media buying.
Three scenarios follow a study:
- iCPA below allowable — the channel is underfunded. Raise budgets and loosen the tCPA target.
- iCPA near allowable — you are at equilibrium. Further growth comes from signal quality and landing pages, not from bids.
- iCPA well above allowable — before cutting, check the composition. Incrementality is usually destroyed by one segment, most often brand search and remarketing, while the rest holds up fine.
Conversion Lift or a geo experiment?
| Situation | Better fit |
|---|---|
| Measuring one channel inside Google Ads | User-based Conversion Lift |
| Answering a question about the whole mix, including non-Google channels | Geo experiment |
| A meaningful share of sales closes offline | Geo experiment or geo-based Conversion Lift |
| Fewer than 1,000 observed conversions | Geo experiment — looser threshold, noisier answer |
| You need audience-level breakouts | User-based, which reports more granular splits |
If your account does not clear the thresholds
A thousand observed conversions and a $5,000 budget rule out a fair number of mid-sized accounts. That is not a reason to give up on incrementality — just a reason to use a different instrument.
- Geo experiment. The closest equivalent: split regions into test and control, switch ads off in control. The volume bar is lower, but noise is higher and the timeline longer. It needs a carefully matched pair of regions, comparable on seasonality and organic share.
- Move the event up the funnel. With few purchases, measure lift on a more frequent event — a lead, a signup, an add to cart — while remembering that such a proxy flatters upper-funnel activity.
- A time-based campaign holdout. The weakest method: pause the campaign for a period and compare with the one before. It controls for nothing — seasonality, competition, promotions — so it answers only the crude question of whether any effect exists at all.
- Accumulate and come back. If you are 20–30% short of the threshold, waiting a quarter usually beats running something unreadable.
What not to do: infer incrementality from a ROAS difference between periods, or from brand search that happened to rise. Neither controls for outside factors, and both routinely produce conclusions with the wrong sign.
Five mistakes that waste a study
- Editing campaigns mid-study. Changing bid strategy, targets or budgets breaks comparability between groups — and resets algorithm learning on top of it. Which edits do that is covered in the Smart Bidding learning period.
- Choosing a rare conversion action. “Request a quote” at twelve events a month will not produce power at any holdback.
- Testing a subset of same-type campaigns. The control group sees a neighbouring campaign and the effect smears.
- Acting on a 60% power result. A number exists, but it is directional; a budget decision on that basis is a bet, not a conclusion.
- Ignoring the diagnostics. Missing enhanced conversions or consent signal means lift is measured on an incomplete conversion set.
What to do this week
- Open Experiments → Lift studies and check whether Conversion Lift is available in your account at all.
- Count observed conversions over the last 30 days, excluding supplementary sources, and compare against 1,000.
- Pull your real conversion lag and set a duration with room to spare.
- Audit measurement hygiene: enhanced conversions, consent, duplicates. On duplicates, see why your report overstates results.
- Agree the cost of the holdback with whoever owns the sales plan before launch, not after.
Incrementality is the one metric worth two weeks of discomfort: it answers a question neither attribution nor a flattering ROAS column will. If you are running paid traffic at a volume where this test makes sense, the rest of the stack deserves the same standard — stable ad accounts included. What we put together on that side is on the Google Ads agency account page and in the wider PPC Rebels services list.
FAQ
How much does a Conversion Lift study cost?
The study itself is free. You pay in holdback: a share of your audience never sees the ads, so you give up their conversions for the duration. At a 10% holdback that is roughly 10% of the channel’s incremental sales over the study window.
Is Conversion Lift now available to every account?
No. Google states availability varies by account and may still require a representative. Check your own account under Campaigns → Experiments → Lift studies.
What counts as an “observed” conversion for the 1,000 threshold?
A conversion your measurement captured directly. Supplementary sources — modelled store visits, uploaded offline leads — do not count toward the threshold, though they can be included in the study itself.
What holdback should a first study use?
Do not pick a comfortable number. Pick the number that pushes projected study power to 90% on the setup screen. If even 50% does not get you there, the volume does not support a study yet.
Can Conversion Lift and Brand Lift run at the same time?
Yes, but the holdback is then fixed at 30%, and a campaign cannot sit in two separate studies — remove it from the current one first.
Why is incremental CPA 50% higher than my reported CPA?
Because reported CPA is attributed and includes conversions that would have happened anyway. The gap is expected, especially with a lot of brand search and remarketing in the mix.
Can I make changes to campaigns while the study runs?
Technically yes, practically no. Any material change to budget, bid strategy or targeting makes the two groups non-comparable and voids the point of the study.
What is cool-down in the report?
A projection of incremental conversions that arrive after the study ends because of lag. It is included by default; results without it are available through the metric tooltip or the reporting dashboard.
Does Conversion Lift work for lead generation?
Yes, if lead volume is sufficient and the lead flow is clean. If bots and junk submissions are filling your forms, the study compares two groups on noise — clean the intake first, using the sequence in fake leads in Google Ads.
What if projected power will not reach 90%?
Three levers: larger holdback, larger budget for participating campaigns, longer duration. A fourth, less popular one: choose a more frequent conversion action — add-to-cart rather than purchase, for example — while being honest about the limits of that proxy.
How is user-based different from geo-based Conversion Lift?
User-based splits people, reports more granular breakouts and works only inside Google. Geo-based splits regions, handles cross-channel and offline questions, but takes longer and returns a noisier answer.
How often should a study be repeated?
Quarterly, or after a structural change in mix or seasonality. Incrementality is not a constant: it moves with brand traffic share, auction competition and audience maturity.
Attribution answers “who gets credit for this conversion”. Conversion Lift answers “was the ad necessary at all”. The second question is more expensive and less comfortable, which is exactly why it usually goes unasked.
