[{"data":1,"prerenderedAt":433},["ShallowReactive",2],{"guide-facebook-ads-incrementality-testing":3},{"id":4,"title":5,"answer":6,"authorId":7,"body":8,"category":333,"ctaVariant":334,"dataset":333,"description":335,"examples":336,"extension":337,"faqs":338,"heroImage":363,"intro":364,"meta":365,"navigation":366,"path":367,"publishedAt":368,"seo":369,"sources":370,"stats":406,"stem":431,"updatedAt":368,"__hash__":432},"blog\u002Fblog\u002Ffacebook-ads-incrementality-testing.md","Facebook Ads Incrementality Testing (2027)","Facebook ads incrementality testing measures the conversions your ads actually caused, not the ones a dashboard takes credit for. You split your audience into an exposed group and a randomly held-out control group that sees no ads, run the test for 2 to 4 weeks, then read the gap in conversion rates as the true lift. It matters because attribution over-credits: people who click an ad were often going to buy anyway, and Apple's App Tracking Transparency cut measurable ad click-throughs by an estimated 37.1%, so platform-reported ROAS is now a modeled guess. Meta's in-platform Conversion Lift runs this as a free randomized controlled trial, and geo or audience holdouts do the same job when you want to run it yourself.","likit-sae-lee",{"type":9,"value":10,"toc":322},"minimark",[11,16,20,23,26,30,33,36,45,122,125,129,132,135,138,142,145,148,151,154,158,161,164,167,229,233,236,239,242,246,249,252,300,303,306,309],[12,13,15],"h2",{"id":14},"why-platform-reported-roas-overstates-what-your-ads-did","Why platform-reported ROAS overstates what your ads did",[17,18,19],"p",{},"Start with the uncomfortable fact underneath every ad dashboard: the platform credits itself. When someone sees or clicks your ad and then buys, the reporting counts that sale as a win for the ad, whether or not the ad had anything to do with the decision. The problem is not dishonesty. It is that the people most likely to click your ad are the people already most likely to buy from you, so the ad and the sale ride together even when one did not cause the other. Statisticians call this selection bias. Marketers feel it as a dashboard that always seems to be winning while the bank balance says otherwise.",[17,21,22],{},"Two large field experiments put hard numbers on how big that gap can be. Researchers ran 15 randomized advertising experiments at Facebook, covering 500 million user-experiment observations and 1.6 billion ad impressions, then checked whether the observational methods the industry normally relies on could reproduce the true experimental result. Published in Marketing Science in 2019, the study found they often could not: even after conditioning on extensive demographic and behavioral data, the observational estimates frequently missed the experimental truth, and in a large share of the cases they overstated the ad's effect substantially. A second experiment at eBay, published in Econometrica in 2015, paused brand-keyword search ads and watched what happened to sales. The measurable incremental return was near zero, because the shoppers searching a brand name were going to arrive anyway. The ad was harvesting a click it did not create, and the reported returns were a fraction of what the non-experimental numbers had claimed.",[17,24,25],{},"The measurement environment has only gotten murkier since. Apple's App Tracking Transparency, introduced in 2021, cut measurable ad click-throughs by an estimated 37.1% according to a University of Maryland study, because platforms lost the signal they used to tie ads to outcomes. Meta now models a growing share of the conversions it reports rather than observing them directly. Costs keep climbing on top of that: Meta reported its average price per ad rose about 9% across full-year 2025. So you are paying more for reach while the number that justifies the spend is built on thinner and thinner data. That is the case for measuring lift directly instead of trusting the credited total.",[12,27,29],{"id":28},"incrementality-versus-attribution-the-difference-that-changes-the-number","Incrementality versus attribution: the difference that changes the number",[17,31,32],{},"Attribution and incrementality are answering two different questions, and confusing them is how good marketers talk themselves into bad decisions.",[17,34,35],{},"Attribution is a reporting method. It looks at the touchpoints it can observe (an ad view, a click, a visit), applies a rule for splitting credit among them (last click, first click, or a data-driven model), and assigns the sale accordingly. It is useful for describing the path a known buyer took. It cannot tell you whether that buyer needed the ad at all, because it never observes the buyer who converts with no ad exposure.",[17,37,38,39,44],{},"Incrementality is an experiment. It creates that missing observation on purpose by holding out a control group, then counts only the conversions above the control's baseline. It is not a new way to divide a reported number, and it is not the same exercise as comparing your blended revenue to your platform-reported revenue. ",[40,41,43],"a",{"href":42},"\u002Fblog\u002Fblended-roas-vs-platform-roas","Blended and platform ROAS"," are both still ratios built from numbers the tools hand you. Incrementality throws out the credited total and rebuilds the answer from a controlled comparison, which is why it can contradict every ratio on your dashboard at once.",[46,47,48,63],"table",{},[49,50,51],"thead",{},[52,53,54,57,60],"tr",{},[55,56],"th",{},[55,58,59],{},"Attribution",[55,61,62],{},"Incrementality",[64,65,66,78,89,100,111],"tbody",{},[52,67,68,72,75],{},[69,70,71],"td",{},"Question it answers",[69,73,74],{},"Which ad touchpoint did this buyer see?",[69,76,77],{},"Would this buyer have converted without the ad?",[52,79,80,83,86],{},[69,81,82],{},"Method",[69,84,85],{},"Observes and assigns credit by a rule",[69,87,88],{},"Randomized experiment with a control group",[52,90,91,94,97],{},[69,92,93],{},"What it counts",[69,95,96],{},"Every conversion it can tie to an ad",[69,98,99],{},"Only conversions above the no-ad baseline",[52,101,102,105,108],{},[69,103,104],{},"Main weakness",[69,106,107],{},"Over-credits buyers who convert anyway",[69,109,110],{},"Needs volume, spend, and a clean test window",[52,112,113,116,119],{},[69,114,115],{},"Right use",[69,117,118],{},"Day-to-day creative and campaign signal",[69,120,121],{},"Deciding whether a channel or budget truly pays",[17,123,124],{},"The practical takeaway is not to throw attribution away. It is fast, it updates daily, and it is a reasonable signal for which creative to scale or cut inside a channel. But when the decision is bigger, whether a whole channel earns its budget, whether to increase spend by six figures, whether retargeting is doing anything a follow-up email would not, attribution is the wrong instrument. That decision needs a causal read, and only an experiment produces one.",[12,126,128],{"id":127},"the-control-versus-exposed-logic-at-the-core","The control-versus-exposed logic at the core",[17,130,131],{},"Every incrementality test, however it is run, rests on the same simple idea borrowed from clinical trials. Take one population, split it at random into two groups, expose one to the treatment (your ads) and withhold it from the other (the holdout), then compare the outcome. Because assignment is random, the two groups are statistically identical in every way that matters, age, location, past purchases, device, intent, on average. The only systematic difference between them is the ads. So any gap in their conversion rates is caused by the ads, not by who happened to land in which group. That is the whole trick, and randomization is what makes it work. Without it, you are back to comparing people who chose to engage against people who did not, which is selection bias wearing a lab coat.",[17,133,134],{},"The math you read out at the end is small. If the exposed group converts at a higher rate than the control group, the difference is the incremental conversion rate, and multiplying it by the exposed population gives incremental conversions. Lift is usually expressed as a percentage: the incremental rate divided by the control rate. A control that converts at 2.0% and an exposed group at 2.4% is a 20% lift, and every conversion in that 2.0% baseline is a conversion your ads did not cause.",[17,136,137],{},"One design detail matters. A clean control group is not simply \"people we did not target.\" It is people who were eligible to be targeted but were randomly withheld, sometimes shown an unrelated public-service ad in the ad's place so the two groups have an identical experience apart from your message. That is how platform-run tests keep the comparison honest, and it is the part a naive homemade test most often gets wrong.",[12,139,141],{"id":140},"meta-conversion-lift-a-free-randomized-controlled-trial","Meta Conversion Lift: a free randomized controlled trial",[17,143,144],{},"Meta builds this experiment into the platform as Conversion Lift, and it is the most accessible way to get a causal read on Facebook and Instagram ads. Meta randomly assigns eligible people in your target audience to a test group, which is eligible to see your ads, and a control group, which is held out. It then measures conversions in both groups through the signals you already send it: the Meta Pixel, the Conversions API, or offline event uploads for purchases that happen off-site. Because the split is randomized at the person level by the platform itself, the control group is a true counterfactual, not an approximation.",[17,146,147],{},"The report gives you the numbers that attribution cannot: incremental conversions (the sales your ads actually caused), the percentage lift over baseline, and an incremental ROAS calculated only on those incremental sales. It reports a confidence level alongside them so you know whether the result is real or noise. Meta treats a lift or holdout result as reliable around 90% confidence, a higher bar than the roughly 65% it uses to call a simpler A\u002FB test winner, precisely because a causal claim should clear a stricter test. There is no extra media fee: you pay for the ads you were going to run, and the experiment sits on top.",[17,149,150],{},"Decide what you are actually testing before you build it, because a lift study answers one question at a time. The highest-value questions are usually the ones attribution flatters most: is retargeting driving sales or reharvesting buyers who were coming back anyway, does a prospecting campaign grow the customer base enough to justify a lower attributed return, would a chunk of budget be better off in another channel entirely. Frame it as a hypothesis with a threshold (\"this campaign lifts purchases at least 10% over baseline\") so the result is a clear pass or fail rather than a number you can rationalize either way. Test the decision that moves real money, not a headline change you could settle with a cheap A\u002FB split.",[17,152,153],{},"The requirements are where most tests fail before they start. Meta advises running its experiments for at least 7 days and no longer than 30, so 2 to 4 weeks is the working window: long enough to cover full weekly buying cycles and your typical consideration period, short enough to stay inside the platform's guidance. You need enough conversion volume for both cells to accumulate a readable sample, which is a volume problem, not a settings problem. And you must leave the test alone while it runs. Changing budgets, swapping creative, editing the audience, or adding a promotion mid-test contaminates the comparison, because now the two groups differ by more than just exposure. Set it, resist the urge to optimize, and read it at the end.",[12,155,157],{"id":156},"geo-and-holdout-tests-you-can-run-yourself","Geo and holdout tests you can run yourself",[17,159,160],{},"Not every advertiser can or wants to run a platform-managed study, and there are two do-it-yourself routes that prove the same thing.",[17,162,163],{},"A geo test splits by geography instead of by person. You keep ads running in one set of regions and pause them in a comparable set, then compare sales across the two over the same window. It is the cleaner option when you cannot hold out individuals, when conversions happen offline or across devices the Pixel cannot see, or when you want a read that does not depend on the platform's own measurement. The hard part is choosing regions that would have behaved the same without the ads. Meta publishes an open-source R package, GeoLift, that handles this by building a synthetic control: it weights your untreated regions into a model that predicts what the treated regions would have done with no ads, then measures the gap. That synthetic baseline is what turns \"sales went up in the on regions\" into a defensible causal estimate.",[17,165,166],{},"An audience holdout is the simplest version of all. When you build a campaign around a custom audience, randomly carve off a slice of that audience and exclude it from delivery, then compare its conversion rate to the slice that saw the ads. It needs no special tooling, just discipline in how you split and exclude. The trade-off is that self-built holdouts are easier to contaminate than a platform RCT: organic exposure, cross-device leakage, and imperfect randomization all creep in, and you have to trust your own splitting. Use a geo or audience holdout when access or setup rules out Conversion Lift, and treat the result as strong evidence rather than a lab-perfect number.",[46,168,169,185],{},[49,170,171],{},[52,172,173,176,179,182],{},[55,174,175],{},"Test type",[55,177,178],{},"How the split works",[55,180,181],{},"Best for",[55,183,184],{},"Main limitation",[64,186,187,201,215],{},[52,188,189,192,195,198],{},[69,190,191],{},"Meta Conversion Lift",[69,193,194],{},"Platform randomizes people into test and control",[69,196,197],{},"A clean RCT read on Meta spend, no extra fee",[69,199,200],{},"Needs volume and a Meta-supported setup",[52,202,203,206,209,212],{},[69,204,205],{},"Geo test \u002F GeoLift",[69,207,208],{},"Ads on in some regions, off in comparable ones",[69,210,211],{},"Offline or cross-device sales, platform-independent read",[69,213,214],{},"Requires comparable regions and a synthetic control",[52,216,217,220,223,226],{},[69,218,219],{},"Audience holdout",[69,221,222],{},"You withhold a random slice of a custom audience",[69,224,225],{},"Quick, self-serve directional read",[69,227,228],{},"Easier to contaminate; randomization is on you",[12,230,232],{"id":231},"what-it-takes-to-get-a-significant-read","What it takes to get a significant read",[17,234,235],{},"Here is the reality that a lot of incrementality enthusiasm runs into: many accounts do not have the volume to prove lift, and no amount of budget-splitting fixes a small numerator. Statistical significance depends on how many conversions the control group gathers, because that baseline is what your incremental result is measured against. A control with a handful of conversions produces a baseline swamped by random noise, and the test cannot separate a real 15% lift from a lucky week.",[17,237,238],{},"Work the arithmetic before you spend a cent. Suppose your account runs 100 conversions a week and you hold out 10% of the audience. The control now gathers roughly 10 conversions a week. To accumulate the few hundred control conversions a reliable read needs, you would run for many weeks, well past the 30-day ceiling Meta recommends, and by then the market has moved under you. Shrinking the holdout to 5% makes the exposed side look better but starves the control further, which is the opposite of what significance needs. This is why a 5% holdout is often a trap on anything but a very high-volume account, and why 10% to 20% is the safer default.",[17,240,241],{},"Budget follows from the same logic, and it is why there is no honest flat \"minimum spend\" to quote. With blended Meta CPM near $8.19 in 2025, reaching enough people to generate several hundred incremental conversions over 2 to 4 weeks costs real money, and the exact figure scales with your conversion rate and order value. A high-volume store might clear the bar comfortably inside a two-week window; a low-volume, considered-purchase brand may not reach significance at any budget it can stomach, and for that business a geo test or a longer seasonal read is the more honest tool than a person-level lift study it will never power. Run the conversion-count math first. If the control group cannot plausibly reach a few hundred conversions, do not run the test, because an underpowered experiment gives you false confidence, which is worse than the attribution number you distrusted.",[12,243,245],{"id":244},"reading-incremental-results-and-acting-on-the-number","Reading incremental results and acting on the number",[17,247,248],{},"A finished test hands you three numbers that matter: the lift percentage, the incremental conversions, and the incremental ROAS. Read them together, and read them against the attributed figures you started with, because the gap is the entire point.",[17,250,251],{},"Walk a worked example all the way through. Say you hold out a control group and the two cells come back like this:",[46,253,254,266],{},[49,255,256],{},[52,257,258,260,263],{},[55,259],{},[55,261,262],{},"Exposed group",[55,264,265],{},"Control (holdout)",[64,267,268,278,289],{},[52,269,270,273,276],{},[69,271,272],{},"People",[69,274,275],{},"100,000",[69,277,275],{},[52,279,280,283,286],{},[69,281,282],{},"Conversion rate",[69,284,285],{},"2.4%",[69,287,288],{},"2.0%",[52,290,291,294,297],{},[69,292,293],{},"Conversions",[69,295,296],{},"2,400",[69,298,299],{},"2,000",[17,301,302],{},"The incremental conversions are 2,400 minus 2,000, which is 400. The lift is 0.4 divided by 2.0, or 20%. Now sit with what that means: of the 2,400 conversions the exposed group produced, only 400 were caused by the ads. The other 2,000 would have happened anyway, so about 83% of the \"attributed\" conversions were not incremental at all. If each sale is worth $50 and you spent $8,000, the platform might proudly report 2,400 sales times $50, a $120,000 return and a 15x ROAS. The incremental truth is 400 sales times $50, or $20,000, which is a 2.5x incremental ROAS on the same spend. Same campaign, same money, two numbers an order of magnitude apart. One of them is real.",[17,304,305],{},"The same split gives you an incremental cost per acquisition, which is often the more sobering read. Divide the spend by the incremental conversions, not the attributed ones: $8,000 over 400 incremental sales is a true acquisition cost of $20, while the attributed math ($8,000 over 2,400) flatters it to about $3.33. If your product carries a $25 margin, that $20 incremental cost is a thin but real profit, whereas the $3.33 figure would have told you to pour in budget you cannot actually afford. Judge the incremental cost per acquisition and the incremental ROAS against your break-even, which is one divided by your gross margin, and ignore the attributed versions of both once you have the experimental read in hand.",[17,307,308],{},"Check the confidence interval before you act. A point estimate of 20% lift with a wide interval that crosses into low single digits is telling you the test was underpowered, and you should treat the result as directional, not final. A tight interval comfortably above zero is a result you can bank. This is where the volume math from the last section pays off: you built enough sample to get an interval you can trust.",[17,310,311,312,316,317,321],{},"Then act on the incremental number, not the attributed one. Set your budget and your break-even against incremental ROAS, because that is the return your margin actually receives. A ",[40,313,315],{"href":314},"\u002Fblog\u002Ffacebook-retargeting","retargeting campaign"," that shows a 2% lift is quietly harvesting demand you already had, and its budget is probably better spent on prospecting that grows the customer base. A prospecting campaign with a modest attributed ROAS but a strong incremental lift is doing the real work, and it deserves more money even though the dashboard makes it look worse. Re-test on a schedule rather than once, because creative fatigues, audiences saturate, and a channel that was incremental last quarter may be coasting on brand awareness this one. Incrementality is not a one-time audit you file away. It is the periodic reality check that keeps the rest of your reporting honest, and it pairs naturally with a working loop where you research the angle, generate and ",[40,318,320],{"href":319},"\u002Fblog\u002Fhow-to-run-a-facebook-ad","run the next creative",", then measure what it truly moved. Platforms like AdPlay.ai keep that loop in one place, but the discipline stands on its own: prove the lift, fund what causes sales, and stop paying full price for conversions you were always going to get.",{"title":323,"searchDepth":324,"depth":324,"links":325},"",2,[326,327,328,329,330,331,332],{"id":14,"depth":324,"text":15},{"id":28,"depth":324,"text":29},{"id":127,"depth":324,"text":128},{"id":140,"depth":324,"text":141},{"id":156,"depth":324,"text":157},{"id":231,"depth":324,"text":232},{"id":244,"depth":324,"text":245},null,"neutral","Prove whether your Facebook ads actually caused sales with a holdout test: Meta Conversion Lift, geo tests, and how to read incremental results.",[],"md",[339,342,345,348,351,354,357,360],{"question":340,"answer":341},"What is incrementality testing for Facebook ads?","It is an experiment that measures the sales your ads actually caused, rather than the sales an attribution model claims. You randomly split a population into an exposed group that can see your ads and a control (holdout) group that cannot, run both for a fixed window, then compare conversion rates. The difference is the incremental lift, the conversions that would not have happened without the ads. Everything else in the exposed group would have converted anyway, which is exactly what platform-reported ROAS cannot separate.",{"question":343,"answer":344},"What is a Meta Conversion Lift study?","Conversion Lift is Meta's built-in randomized controlled trial. Meta randomly assigns eligible people in your target audience to a test group that is eligible to see your ads and a control group that is held out, then measures conversions in both through the Pixel, the Conversions API, or offline event uploads. The report shows incremental conversions, the percentage lift, and an incremental ROAS based only on the sales the control group did not produce on its own. It is offered inside Meta at no extra media cost beyond the spend you were already running.",{"question":346,"answer":347},"How is incrementality different from attribution?","Attribution divides credit for a conversion among the ad touchpoints it can observe, then assumes the touchpoint caused the sale. Incrementality makes no such assumption: it uses a control group to estimate what would have happened with no ads at all, and counts only the extra conversions above that baseline. Attribution answers 'which ad did this buyer touch,' incrementality answers 'would this buyer have converted anyway.' The two can disagree by a wide margin, because the people most likely to click your ad are often the people most likely to buy without it.",{"question":349,"answer":350},"How long should a Facebook incrementality test run?","Plan for 2 to 4 weeks. Meta advises running its tests for at least 7 days and no longer than 30, so a 2 to 4 week window captures full weekly buying cycles while staying inside that range. A test needs to cover your typical consideration period end to end, so if shoppers usually take ten days from first click to purchase, a one-week test will miss most of the conversions it caused. Do not stop early because the numbers look good on day three, and do not change budgets or creative mid-test.",{"question":352,"answer":353},"How big does the holdout group need to be?","Big enough to produce a reliable baseline, which usually means more than the 5% many advertisers reach for first. A tiny holdout on a low-volume account gathers too few control conversions to tell a real lift from random noise. A 10% to 20% holdout is a safer starting point unless your conversion volume is very high, and the deciding factor is the number of conversions the control group accumulates, not the percentage itself. If the holdout cannot reach roughly a few hundred conversions over the window, the test cannot reach significance.",{"question":355,"answer":356},"How much budget do I need for a conversion lift test?","Enough to generate a statistically meaningful number of incremental conversions in both groups over the window, which is a function of your conversion volume and cost, not a flat fee. With blended Meta CPM near $8.19 in 2025 per Gupta Media, and a holdout that needs several hundred control conversions to read cleanly, low-volume accounts often cannot reach significance at any comfortable spend. The honest test is arithmetic: if your baseline is 100 conversions a week and you hold out 10%, the control gathers only about 10 a week, so you would need many weeks to accumulate a readable sample.",{"question":358,"answer":359},"What is incremental ROAS and how do I calculate it?","Incremental ROAS is revenue from incremental conversions divided by ad spend, where incremental conversions are the exposed group's conversions minus what the control group's rate predicts they would have produced anyway. Worked example: an exposed cell converts at 2.4% and the held-out control at 2.0%, so 0.4 points of that 2.4% is incremental, meaning roughly 83% of the exposed conversions were not caused by the ads. If the platform claims a 15x return on that campaign, the incremental ROAS on the same spend might be closer to 2.5x. That lower number is the one your margin has to clear.",{"question":361,"answer":362},"Can I run an incrementality test without Meta's tool?","Yes, with a geo or audience holdout. A geo test turns ads off in a set of regions and on in comparable ones, then compares sales; Meta's open-source GeoLift package builds a synthetic control from your untreated regions for exactly this. A simpler route is to hold out a random slice of a custom audience from a campaign and compare its conversion rate to the exposed slice. Both are less airtight than a platform-randomized RCT, but they need no special access and prove the same thing: what happens to sales when the ads stop.","\u002Fimages\u002Fblog\u002Ffacebook-ads-incrementality-testing-hero.webp","Your dashboard says the campaign returned six dollars for every one you spent, and you do not quite believe it. You are right to be skeptical. That number counts every sale it can loosely tie to an ad view or click, including the customers who were already halfway to checkout. Incrementality testing answers the only question that pays the bills: how many of those sales would not have happened if the ads had never run. This guide walks the control-versus-exposed logic, the free way to run the test inside Meta, the do-it-yourself geo and holdout routes, and how to read a result you can actually bank.",{},true,"\u002Fblog\u002Ffacebook-ads-incrementality-testing","2027-06-11",{"title":5,"description":335},[371,375,379,383,387,390,393,396,400,403],{"label":372,"url":373,"year":374},"Marketing Science (Gordon, Zettelmeyer, Bhargava, Chapsky), A Comparison of Approaches to Advertising Measurement (Facebook field experiments)","https:\u002F\u002Fwww.kellogg.northwestern.edu\u002Ffaculty\u002Fgordon_b\u002Ffiles\u002Ffb_comparison.pdf","2019",{"label":376,"url":377,"year":378},"Econometrica \u002F NBER (Blake, Nosko, Tadelis), Consumer Heterogeneity and Paid Search Effectiveness (eBay experiment)","https:\u002F\u002Fwww.nber.org\u002Fpapers\u002Fw20171","2015",{"label":380,"url":381,"year":382},"University of Maryland R.H. Smith School, Small Businesses Take Big Hit from Apple's Privacy Regulation","https:\u002F\u002Fwww.rhsmith.umd.edu\u002Fresearch\u002Fsmall-businesses-take-big-hit-apples-privacy-regulation","2024",{"label":384,"url":385,"year":386},"Meta Business Help Center, About Conversion Lift","https:\u002F\u002Fwww.facebook.com\u002Fbusiness\u002Fhelp\u002F221353413010930","2026",{"label":388,"url":389,"year":386},"Meta Business Help Center, About A\u002FB Testing","https:\u002F\u002Fwww.facebook.com\u002Fbusiness\u002Fhelp\u002F1738164643098669",{"label":391,"url":392,"year":386},"Meta Business Help Center, About confidence in your tests and experiments","https:\u002F\u002Fwww.facebook.com\u002Fbusiness\u002Fhelp\u002F239549606692303",{"label":394,"url":395,"year":382},"GeoLift (Meta \u002F Facebook Incubator), geo-experiment methodology (open-source R package)","https:\u002F\u002Fgithub.com\u002Ffacebookincubator\u002FGeoLift",{"label":397,"url":398,"year":399},"Gupta Media, The True Cost of Social Media Ads (CPM Tracker)","https:\u002F\u002Fwww.guptamedia.com\u002Fsocial-media-ads-cost","2025",{"label":401,"url":402,"year":399},"Meta, Fourth Quarter and Full Year 2025 Results","https:\u002F\u002Finvestor.atmeta.com\u002Finvestor-news\u002Fpress-release-details\u002F2026\u002FMeta-Reports-Fourth-Quarter-and-Full-Year-2025-Results\u002Fdefault.aspx",{"label":404,"url":405,"year":399},"Search Engine Land, Why incrementality is the metric that proves marketing's real impact","https:\u002F\u002Fsearchengineland.com\u002Fwhy-incrementality-is-the-only-metric-that-proves-marketings-real-impact-463439",[407,411,415,419,423,427],{"label":408,"value":409,"source":410},"Estimated drop in ad click-throughs after Apple's App Tracking Transparency","~37.1%","University of Maryland, 2024",{"label":412,"value":413,"source":414},"Ad impressions across 15 Facebook lift experiments in the measurement study","1.6 billion","Marketing Science, 2019",{"label":416,"value":417,"source":418},"Measurable incremental return from eBay brand-keyword search ads","near zero","Econometrica, 2015",{"label":420,"value":421,"source":422},"Blended Meta (Facebook and Instagram) CPM, full year","$8.19","Gupta Media, 2025",{"label":424,"value":425,"source":426},"Meta average price per ad change, full-year 2025","+9%","Meta, 2025",{"label":428,"value":429,"source":430},"Confidence level Meta uses to call a lift or holdout test reliable","90%","Meta Business Help Center, 2026","blog\u002Ffacebook-ads-incrementality-testing","ycVvR32nc2FTfDJjGokMGIOMAUrPHJTVsyv-8njIwj0",1786093700522]