Meta Attribution: Windows, Models, Real ROAS

Attribution is how Meta decides which ad gets credit for a sale. It sounds simple. It is not. Behind a single ROAS number sit several layers: a click window, a view window, an engagement window, an attribution model, and a conversion count. Change any one of them and the number moves, sometimes by half. Worse, the number Meta shows you is structurally inflated. Meta marks its own homework and has every reason to claim credit for sales it never caused. This guide breaks down each layer in plain terms, explains what changed in 2025 and 2026, and shows you why the ROAS in your Ads Manager is not the ROAS in your bank account. You will leave knowing how to read attribution like an operator, not a tourist who trusts the headline figure.

What an attribution window actually is

An attribution window is the period of time after someone clicks or sees your ad during which a conversion can be credited to that ad. If your window is seven days and a person clicks today and buys on day six, Meta credits the ad. If they buy on day eight, it does not. The window is a rule about time, nothing more. Yet it silently controls a huge share of what you see. A longer window catches more conversions and flatters your ROAS. A shorter one is stricter and reports less. Neither is the truth. Both are just different stopwatches measuring the same race, and the length of the stopwatch is a choice you make, not a fact about how your customers behave.

There are two flavors of window. A click window covers people who clicked the ad link before converting. A view window covers people who only saw the ad, never clicked, and converted anyway. These are very different signals. A click is intent: the person did something deliberate. A view is a guess: the person had your ad on screen, maybe glanced at it, maybe scrolled past while it loaded. Meta treats both as evidence the ad worked. You should not. The view window is where most of the inflation hides, and we will come back to it repeatedly because it is the single biggest reason your reported ROAS looks better than your real one. Every serious mistake in attribution traces back to confusing a view with a decision.

The 2026 default settings, decoded

When you build an ad set for website conversions and optimize for the number of purchases, Meta sets your attribution to a default: seven day click, one day engage through, and one day view through. That means a conversion counts if it happens within seven days of a link click, within one day of a non link engagement, or within one day of an impression with no click. Jon Loomer documents this stack clearly and notes Meta only gives you a few editable options. You can pick seven or one day click. You can keep or remove the one day engage and the one day view. You cannot invent a thirty day click. The defaults are chosen by Meta to maximize reported volume, which conveniently makes the platform look generous with the credit it hands itself.

In March 2026, Meta changed what counts as a click. Before, any tap on the ad counted, including likes, comments and other non link taps. Now click through attribution only credits conversions after a click on the ad link itself. Everything else, social clicks, reactions, engaged video views, moved into a new bucket called engage through attribution, still on a one day window. This matters because the same conversion can now land in a different column than it did a year ago. If you compare reports across that date, you are comparing apples to a slightly different apple. Always check which definition was live when the data was recorded before you draw a trend line, or you will mistake a definition change for a performance change and react to a ghost.

Why the 28 day window disappeared

For years the default was twenty eight day click and one day view. A click could be credited for almost a month. Then Apple shipped iOS 14.5 and App Tracking Transparency in April 2021. Users could refuse tracking, and most did. Meta lost a large slice of the signal it needed to follow a person for twenty eight days. Its response was to cut the default to seven day click and one day view. The data needed to support a long window simply was not reliable enough anymore. This was not a UX tweak. It was a forced retreat dressed up as a setting change, and it permanently reshaped how every advertiser reads performance. Anyone still benchmarking against pre 2021 ROAS numbers is comparing two different measurement worlds.

Then came a harder cut. On 12 January 2026, Meta permanently removed the seven day view and twenty eight day view windows from the Ads Insights API, leaving only one day view alongside the click windows. According to DOJO AI, accounts that had been leaning on those long view windows saw reported conversions drop fifteen to forty percent overnight. Read that carefully. The sales did not vanish from the business. They vanished from Meta’s report, because the rule that had been crediting them was deleted. If your dashboard fell off a cliff in mid January 2026 and nothing else changed, this is almost certainly why. The customers were always real or always imaginary. Only the accounting moved, which is the clearest proof that the metric and the money were never the same thing.

Standard versus Incremental attribution

Beyond windows, Meta now offers two attribution models. Standard is the old behavior: count every conversion that falls inside your windows, whether or not the ad truly caused it. Incremental, which rolled out in 2025, uses Meta’s own prediction models to estimate whether a conversion would have happened without the ad, and optimizes delivery toward conversions it judges to be caused. When you select Incremental you lose the ability to edit windows, because the model now decides what counts. The promise is a cleaner reflection of true impact at the cost of raw volume. The catch is that you are still trusting Meta’s model to grade Meta’s ads, which is the same conflict of interest moved one level up the stack.

Loomer’s honest read is worth repeating: in his own accounts he has seen little difference between Standard and Incremental results, partly because he already tunes windows to suppress junk view conversions. That is the practical lesson. Incremental is not magic. If you already remove one day view on remarketing and on non purchase events, you have done much of the cleanup by hand. Incremental makes sense by default for advertisers with high budgets and reliable volume, where trading some volume for honesty costs nothing painful. If you struggle to exit the learning phase, deliberately starving yourself of reported conversions can hurt more than the cleaner number helps. Choose based on your volume and your goal event, not on the marketing copy that sells the feature.

First conversion versus all conversions

A third lever is conversion count. By default Meta counts All Conversions, so one person who buys twice inside your window adds two to your numbers and two purchase values. First Conversion counts only the initial action per person. This sounds minor until you compare Ads Manager with your Shopify back end and the totals refuse to match. Often Ads Manager is not lying so much as double counting repeat buyers. Switching the reporting view to First Conversion frequently reconciles the gap. Meta introduced this as a reporting option in 2024 and later let you optimize on it inside the ad set. Use it to investigate discrepancies, but think twice before optimizing on it, since you usually do want repeat buyers counted and valued.

The core lie: in platform ROAS overstates reality

Here is the uncomfortable truth no Meta rep will volunteer. The ROAS in Ads Manager is over attributed by design. Meta credits itself for conversions that would have happened without a single ad. Across verticals, independent analysts put the over attribution of self reported platform ROAS at roughly twenty to fifty percent versus incrementality tested or marketing mix model figures. The mechanism is simple. If someone already wanted your product, searched for your brand, and would have bought regardless, but they happened to see or click your ad inside the window, Meta books the sale as a win. Multiply that across thousands of warm, intent loaded users and you get a number that feels great, decides nothing, and quietly bleeds budget into demand you already owned.

The cleanest way to size the gap is the incrementality factor, the ratio of incremental conversions to platform reported ones. If Meta claims five hundred conversions but a controlled lift test shows three hundred would not have happened without the ads, your incrementality factor is zero point six. Forty percent of the credited sales were going to happen anyway. The number gets brutal on retargeting. Industry lift studies cited by Seresa and others find that as many as seventy five percent of retargeting conversions would have occurred without the ad. That turns a glorious looking eight times ROAS into a roughly two times incremental ROAS. Same campaign, same spend, very different verdict on whether you should keep funding it or quietly turn it down.

Multi touch attribution and why it shifts the picture

Last click attribution gives one hundred percent of the credit to the final interaction before the sale. It is tidy and it is wrong for most journeys. A real customer might see a Meta video, ignore it, get a retargeting reminder, search your brand on Google, click an email, and finally buy. Last click hands the trophy to the email and tells you Meta did nothing, when Meta created the demand the email merely harvested. Multi touch attribution tries to fix this by spreading credit across several touchpoints, from the first impression to the final click. Data driven multi touch uses machine learning to assign credit by observed patterns rather than fixed rules, which is more defensible than any rigid first or last touch model that ignores the middle of the journey.

Multi touch is not a silver bullet either. It still relies on user level tracking, which privacy changes keep eroding, and it cannot see offline word of mouth or the brand search that an ad quietly triggered. This is why serious teams stack methods. A 2025 EMARKETER survey found roughly thirty five percent of US marketers planning to invest in multi touch attribution and close to half in marketing mix modeling. Meta’s own open source mix modeling tool, Robyn, explicitly calibrates its output against geo tests, lift studies and multi touch data, meaning even Meta’s engineers assume no single model is trusted alone. The grown up answer is triangulation: platform numbers for daily steering, lift tests for truth, mix modeling for budget allocation across the whole channel mix.

What incrementality testing reveals about Advantage Plus

Haus analyzed six hundred forty incrementality experiments and the results puncture a popular belief. Advantage Plus Shopping campaigns, Meta’s heavily promoted automation, over reported relative to manual campaigns by twelve percentage points on average. In head to head tests, fifty eight percent of brands saw higher incremental ROAS on manual campaigns than on Advantage Plus, even though Advantage Plus often shows the prettier in platform number. Advantage Plus delivered roughly twelve percent lower incremental ROAS at eighteen percent lower daily spend. The automation was not driving more real revenue per dollar. It was claiming more credit per dollar. That distinction is the whole game, and the in platform dashboard hides it perfectly behind a confident multiple.

There is a second finding worth burning into memory. When brands rebalanced away from putting more than seventy five percent of spend into a single favored tactic toward a roughly even split, Haus saw an average eighteen percent improvement in incremental ROAS. Concentration in the tactic the platform flatters is not free. The lesson is not that Advantage Plus is bad, it is that you cannot judge any tactic by its self reported number. The only honest scoreboard is a controlled test where some users are deliberately withheld from your ads and you measure the difference in actual sales. Everything else is the platform telling you, with great confidence, exactly how good the platform is at taking credit.

How third party tools cap the over counting

If you sum the orders that Meta, Google and TikTok each claim, the total routinely exceeds the orders you actually see in Shopify. Every platform self attributes, so the same sale gets counted three times over. Third party attribution tools handle this differently. A tool like Northbeam assigns credit across channels so that the total can never sum to more than your real sales, which structurally solves the over attribution problem. A tool like Triple Whale offers a model that hugs the platform numbers more closely, useful for steering but less corrective. Neither replaces a lift test, but a tool that refuses to invent revenue is a far better daily compass than three dashboards each pretending to have single handedly driven the same order.

Keep a sense of scale while you do this. Triple Whale’s 2025 dataset, drawn from roughly thirty five thousand brands, put average ecommerce ROAS for Meta around one point nine times on a platform reported basis. Remember that this figure is already inflated by the over attribution we just discussed, so the incremental reality for many of those brands sits lower still. If your own Meta ROAS is sitting at two times in Ads Manager, your incremental ROAS could be close to break even or below it. That is not a reason to panic and cut everything overnight. It is a reason to run a lift test before you decide whether the spend is building your business or just decorating your dashboard with sales you would have made anyway.

How to actually use attribution without fooling yourself

Start with hygiene. Keep the default seven day click for purchases, drop the one day view on remarketing and on any non purchase event like leads, because that is where view through inflation does the most damage. Use the Compare Attribution Settings view to see how your numbers swing across windows; if seven day click and one day click diverge wildly, your customers take time to decide and your creative or offer may need patience, not panic. Switch the reporting view to First Conversion whenever Ads Manager and your back end disagree, to catch double counted repeat buyers. None of this costs money. All of it stops you optimizing toward a number that was always partly fiction.

Then build the truth layer. Run a Meta conversion lift test or a geo holdout at least once per quarter so you have a real incrementality factor to discount your reported ROAS. Validate retargeting hardest, since it is the worst offender. Use a third party tool that caps total credit at real sales for daily steering, and treat the platform ROAS as a directional signal, not a verdict. If budgets are large, lean toward marketing mix modeling for allocation. The mature operator holds three numbers in their head at once: the inflated platform figure for speed, the lift adjusted figure for truth, and the bank balance for sanity. When those three disagree, the bank balance wins, every single time, without exception.

The mindset that separates pros from spenders

Most advertisers treat the Meta ROAS as gospel and scale whatever shows the biggest multiple. That is exactly how you scale the campaigns that steal credit best, not the ones that grow your business. The professional move is to assume every reported number is generous until a controlled test proves otherwise. Windows are choices, not facts. Models are estimates, not verdicts. View through is a hint, not evidence. The single skill that compounds over a career is the discipline to keep asking one question of every dashboard: if Meta had never run this ad, how many of these sales would still have happened? Answer that with a real test, and you stop paying for conversions you already owned.

The brand search trap that fools everyone

Picture the most common over attribution story in ecommerce. A shopper sees your Meta ad in the morning, feels mild interest, and scrolls on without clicking. That afternoon they remember the brand, open Google, type your name, click the first result and buy. Inside Meta, this is logged as a one day view through conversion, a clean win for the ad. Inside Google Analytics, it is logged as branded search. Both platforms claim the same sale, and your blended spreadsheet now contains a sale that exists once but is counted twice. The Meta ad genuinely planted the seed, so it deserves some credit, but nowhere near the full purchase value it quietly absorbs in the report.

This is exactly the confound a lift test is built to remove. Withhold the Meta ad from a random slice of users, and the branded searches that still happen in that holdout group were never caused by Meta in the first place. The difference between the test group and the holdout is your true incremental impact, stripped of the demand that existed anyway. Until you run that test, every view through conversion sitting next to a strong organic or branded search channel deserves suspicion. The bigger and better known your brand, the worse this gets, because more of your buyers were already coming. Strong brands often have the most flattering and the least trustworthy in platform ROAS.

Sources

Jon Loomer, How Meta Ads Attribution Works in 2026 (jonloomer.com). DOJO AI, Meta Ads Attribution in 2026: What Changed (dojoai.com). mbuzz, Meta Killed 7-Day and 28-Day Attribution (mbuzz.co). Meta Business Help Center, About Attribution Models and Attribution Settings (facebook.com/business/help). Haus, The Meta Report: Lessons from 640 Incrementality Experiments, and Is Meta Incremental (haus.io). Seresa, Meta Incremental Attribution: Real ROAS Guide 2025 (seresa.io). Measured, Meta Advantage Plus: Why High ROAS is a Red Flag (measured.com). Eightx, Average Ecommerce ROAS by Vertical 2026 (eightx.co). Triple Whale, Meta Conversion Lift Tests (triplewhale.com). Stormy AI and TrueROAS on Northbeam versus Triple Whale attribution (stormy.ai, trueroas.com). EMARKETER 2025 attribution investment survey, via inBeat and Improvado. Apple iOS 14.5 App Tracking Transparency, April 2021.