by Trivas.aiOne bad day on one channel can rewrite your whole media plan if you're not careful. A single spike or crash in daily ROAS gets averaged into the weekly number, and suddenly you're moving budget based on noise instead of signal. Knowing how to identify ROAS outliers across paid channels before they hit your reporting is the difference between a real optimization and a coin flip.
Here's the mechanic: one abnormal data point, a 12x spike from a branded search term or a 0.3x crash from a broken pixel, gets folded into a blended average like it's a normal day. It isn't. It's noise wearing a signal's clothes.
Say Meta spends $500 on a Tuesday and lands a 15x ROAS because one customer placed a $7,500 wholesale order. That single order can make an entire week of mediocre Meta performance look like your best channel. Meanwhile the campaign that's actually been steady at 3.2x for a month gets ignored because it never spikes.
The cost isn't abstract. Teams shift budget toward a channel that only looks good because of one outlier, or they cut a channel that was actually stable, just because a bad day happened to land during the reporting window. Either mistake compounds over a quarter.
This article walks through the actual methods, rolling benchmarks, z-scores, and order-level cross-checks, for catching these outliers before they warp a budget decision.
Not every spike is bad data, and not every crash is a broken campaign. The trick is telling a true outlier apart from a real performance shift.
True outliers usually trace back to something mechanical: a broken UTM parameter, a delayed conversion attribution window catching up all at once, a mismatch between what the ad platform reports and what actually landed in your bank account, or refunds that haven't been reflected yet. None of these describe how the campaign actually performed. They describe a data problem wearing a performance costume.
A real performance shift looks different. It's a campaign that's been climbing steadily for two weeks because you tightened the audience or swapped in a better creative. That's signal, not noise.
Channel quirks make this messier. Amazon Ads ROAS can spike from organic halo effects getting misattributed to the ad. Meta can inflate numbers through view-through attribution, crediting a sale to someone who scrolled past an ad and bought two days later through an unrelated path.
A decent rule of thumb: if a data point sits more than 2 standard deviations from the trailing 30-day channel average, treat it as a candidate for review, not a confirmed outlier. Flag it, don't act on it yet.
Comparing Meta ROAS straight across to Amazon Ads ROAS is a mistake almost everyone makes at some point. Different fee structures, different attribution windows, different definitions of what counts as a conversion. A "good" Amazon ROAS and a "good" Meta ROAS aren't the same number, and they shouldn't be judged against one shared target.
The fix is building a rolling baseline per channel, per campaign type. Run a 14-day and a 30-day rolling average separately for prospecting and retargeting campaigns on each platform. Prospecting will naturally run lower and noisier. Retargeting should sit higher and steadier. Blending them into one number hides both.
Once you have that baseline, flag a day or campaign as an outlier when it deviates by a set percentage, say plus or minus 40 percent, from its own rolling average. Not from a company-wide target. A TikTok prospecting campaign at 1.8x isn't underperforming if its 30-day baseline is 1.5x. It's actually up.
This only works if the underlying data is unified first. Spreadsheet exports from four or five platforms rarely land on the same day, use the same currency rounding, or define "spend" the same way. Blending Meta, Google, TikTok, and Amazon data into one consistent source, which is exactly what a BI reporting layer on something like Redshift is built for, removes that sync problem before you even get to the math.
Percentage thresholds are a good start, but they miss the subtler cases. That's where z-scores come in.
A z-score on daily ROAS is simple: (value minus mean) divided by standard deviation. It tells you how many standard deviations a given day sits from its own channel's normal range.
A practical threshold: z-score above 2 or below negative 2, calculated over a 30-day lookback, warrants a manual look. This catches things a flat percentage rule misses. TikTok, for example, naturally swings harder day to day than branded search does. A 40 percent move on TikTok might be entirely normal. The same move on a branded search campaign, which usually sits in a tight band, is a real red flag. Z-scores account for that natural variance instead of applying one blanket rule to every channel.
The limitation: small sample sizes break this. A campaign that's only been live for eight days doesn't have a stable enough mean or standard deviation to produce a trustworthy z-score. Pair this method with a minimum spend or impression threshold, something like $1,000 in spend or 14 days of history, before you trust what the number is telling you.
A ROAS spike on a dashboard is a claim, not a fact. Verify it against actual order data before you believe it.
Say a campaign shows 8x ROAS for the day. Pull the order-level detail and you find 90 percent of that attributed revenue came from one $4,000 wholesale order that has nothing to do with the ad creative or targeting. That's not the campaign performing well. That's one order happening to land on the same day.
Attribution windows cause a similar problem. A last-click model will happily assign a sale to a retargeting ad that showed up at the very end of a journey that actually started with an organic search or an email click three days earlier. The retargeting campaign gets credit it didn't really earn.
Pulling GA4 funnel data alongside your ad platform numbers is the check here. If the funnel shows a multi-touch path that started outside the flagged channel, the "outlier" is really an attribution artifact, not incremental revenue. This step is slower than the first two, but it's the one that actually confirms whether a flagged number is real.
Doing all three of these methods by hand, every week, across four or five platforms, is the part nobody actually enjoys. It's also where most teams quietly give up and just eyeball the dashboard instead.
Trivas blends Meta, Google, TikTok, and Amazon spend and revenue data on Redshift, so rolling benchmarks get calculated on one consistent dataset instead of five separate exports that never quite line up on date ranges or currency handling.
On top of that, the AI Wingman layer surfaces anomalies directly, campaigns deviating from their own baseline show up in the dashboard instead of requiring someone to rebuild a z-score formula in a spreadsheet every Monday.
The forecasting and simulation modeling piece adds one more layer: it can indicate whether a flagged outlier is likely to persist or revert, so a team isn't shifting budget off a single unusually good or unusually bad day.
Here's a process you can run without any special tooling, if you'd rather do it manually first:
Run this weekly, not daily. Daily reviews just mean you're reacting to single-day noise before it's even had a chance to average out. Performance marketers managing several channels at once tend to get more signal out of one focused Monday review than five rushed daily glances.
Catching outliers isn't a gut check on a dashboard number. It takes a rolling baseline per channel, a statistical threshold like a z-score, and an order-level check to confirm the number is even real.
Rebuilding that process by hand across four or more ad platforms, in spreadsheets, takes hours every single week, and it breaks the moment one platform changes its export format on you.
If you want to sanity-check a specific campaign right now, run the numbers through our ROAS calculator. Or if you'd rather see automated outlier flags running against your actual channel data, start a trial and watch it happen without the spreadsheet.
explore more insights
3 min read
3 min read
3 min read
.d53b12e5.png&w=3840&q=75 "Trivas.ai")