How to ask AI when your ROAS numbers don’t match

Four different ROAS for the same campaign

Monday, 9:50 a.m. Ten minutes before the weekly client call, you’ve got three tabs open. Meta Ads Manager shows a 4.2 ROAS, GA4 shows 2.8, and the back office shows 3.5—same campaign, same date range. Every week you have to pick which number to put on the slide, and every week someone asks why they don’t line up.

It gets trickier when you add up each platform’s attributed revenue. Meta claims 4.2 million KRW, Google says 3.6 million, and Naver says 1.8 million—but actual paid revenue in the back office might be just 7 million. The platforms aren’t lying. Each one credits itself generously within its own ad touchpoints and doesn’t deduplicate against the others.

This is also why even Meta shows up as two different numbers. The ROAS in Meta Ads Manager is the performance Meta credits itself under its own attribution rules; the Meta ROAS in GA4 is the share GA4 assigns to Meta across the full customer journey. GA4’s number usually comes out lower because it splits overlapping credit with other channels and applies its own reporting attribution model—last-click or data-driven, for example.

In a spot like this, asking an AI agent to “analyze my ROAS” doesn’t fix anything. AI can pull the numbers faster, but unless you tell it whether to use platform-reported figures, GA4’s channel allocation, or actual back-office revenue, all you’ve done is automate the same confusion.

Four things that drive the differences

The first is attribution—specifically, the conversion window. This sets how many days after a click or impression a purchase still counts toward that ad. Google Ads and GA4 let you set conversion windows and reporting attribution models, and TikTok breaks out separate windows for click-through, engaged-view, and view-through conversions. Naver Biz Advisor likewise reports last-click ROAS and +14-day estimated-contribution ROAS as two separate numbers.

The second is view-through credit and modeling. The numbers shift depending on whether you credit conversions where the user saw the ad but never clicked, and whether you use modeling to fill the gaps that privacy and browser restrictions leave behind. Google Ads says outright that it uses modeled conversions to cover attribution it can’t observe directly.

The third is how you define conversion value. Whether the value your pixel fires includes VAT or shipping, and whether it nets out cancellations and refunds, naturally drives a wedge between platform ROAS and back-office ROAS. Google Ads’ Target ROAS, after all, optimizes against the conversion value the advertiser reports to it.

The fourth is double counting. If a customer sees a Meta ad, clicks a Google Search ad, then runs a Naver Brand Search before buying, every one of those platforms can claim the order within its own touchpoints. Add up the platforms’ attributed revenue and you’ll top actual order revenue; add up their ROAS and you get a fake “blended ROAS” that means nothing.

The industry isn’t chasing one ‘true’ ROAS

In practice, the industry keeps platform ROAS, analytics-tool ROAS, back-office ROAS, blended ROAS, marketing efficiency ratio (MER), and incrementality as separate measures, each for a different job. Northbeam chalks up the ROAS gaps between platform reporting and its own multi-touch attribution (MTA) model to differences in attribution models, attribution windows, and customer journeys. Triple Whale, too, splits its data sources and attribution models by use case—first-party click data, deterministic view data, and post-purchase surveys.

Google likewise points teams toward using attribution, incrementality experiments, and marketing mix modeling (MMM) together. Attribution explains the touchpoints along the journey; incrementality experiments measure the extra conversions ad exposure actually drove; MMM uses aggregated data to factor in outside influences and long-term effects.

In other words, the working question in the industry isn’t “Which number is absolutely right?” You use platform ROAS to compare campaigns and creative within that platform, GA4 ROAS to weigh channel contribution on one standard, and back-office or blended ROAS for real P&L and budget calls. The ground rule: never call the sum of platform-reported revenue your total revenue.

Carry that same separation into Neuro when you hand ROAS to AI. Before AI tells you “Meta ROAS looks strong, raise the budget,” it has to say whether that Meta ROAS comes from Meta Ads Manager, from GA4’s allocation to Meta, or from a figure recalculated against actual back-office sales.

Don’t ask AI: ‘Did ROAS drop?’

Bad questions are short. Ask “Why did ROAS drop last week?” and AI has no idea which source to look at, which attribution window to apply, or whether to compare against the week before or the prior month’s average. You usually get back a summary that just rattles off a handful of metrics.

Good questions lock in the data source and the decision up front. For example: “Using the GA4 attribution model our team reports on, compare ROAS for May 11–17 against May 4–10. Show Meta Ads Manager ROAS alongside it for reference only, and don’t sum the platforms’ attributed revenue.”

The four parts of a good ROAS question

1. Pin down the source. Say up front whether GA4, Meta Ads Manager, Google Ads, or the back office is the standard.
2. Pin down the attribution standard. Spell out the rules behind the number—7-day click, 1-day view included, data-driven attribution, last click.
3. Name the comparison period. Pick a baseline: same weekdays last week, the last 7 days, last month’s average, or the campaign’s first week live.
4. Name the decision. AI needs different data depending on whether you’re walking through a report, planning a budget increase, pausing a campaign, or checking for tracking issues.

If spelling out all four every time is a hassle, lock them into a ROAS-analysis Gem in Neuro instead. In the Gem instructions you can bake in team rules like “Decision-grade ROAS uses back-office paid revenue, net of refunds,” “Platform ROAS is for internal optimization only,” and “Every report shows its source and attribution standard.”

Five prompts you can use right now

These prompts all have one thing in common: none of them asks AI to jump to a conclusion. They pin down the standard, compare within it, then ask for follow-up checks. Teams that work this way spend less time explaining away ROAS gaps and can move to next steps right away.

In Neuro, ROAS standards live in your workflow

If you run ROAS analysis on a regular basis in Neuro, split the work between workspace standards and Gems rather than rewriting a long prompt each time. Lock your decision-grade ROAS into the workspace—for example: “Budget calls run on Cafe24 or back-office paid revenue, net of refunds,” and “Platform ROAS is for in-campaign optimization only.”

Then build a ROAS-analysis Gem. In the Gem instructions, set the review order, the numbers to leave out, and the reporting rules—for example: “Check decision-grade ROAS first, then GA4, and platform ROAS last,” “Never sum ROAS from different sources into one row,” and “When AI proposes a budget change, it must keep the standard ROAS and the reference ROAS separate.”

1. Line up the sources: for the same campaign, period, and time zone, view GA4, platform, and back-office ROAS side by side.
2. Don’t aggregate: never add up attributed revenue or platform ROAS and call it total revenue or total ROAS.
3. Break down the gap: form hypotheses about which factor explains it—attribution window, view-through credit, modeled conversions, deduplication, refunds, or time zones.
4. Verification steps: spell out concrete next actions for a human—compare Meta attribution settings, audit GA4 events, recompute back-office refunds, or confirm what Naver’s cost figures cover.

This flow plays to how Gems work in Neuro. A Gem isn’t just somewhere to park a long prompt; it’s how you turn your team’s ad-operations standards into a repeatable mode of working. Build a ROAS-analysis Gem once, and next week you can ask the same way and review the answers in the same shape.

Define what AI shouldn’t do with ROAS, too

AI can’t settle every ROAS gap with certainty. There’s no way to audit, from the outside, how accurate each platform’s internally adjusted modeled conversions are. AI can say “this gap could be down to modeling,” but it shouldn’t claim that some exact percentage of revenue is the result of modeling.

Without back-office revenue and refund data connected, you can’t build decision-grade ROAS at all. AI can still explain the gap between platform ROAS and GA4 ROAS, but it shouldn’t make a budget-increase call on a P&L basis. The same goes for thin-data stretches like a new campaign’s first week.

A five-minute checklist before you present

• Did you state the ROAS source on the slide or in the chat reply?
• Are the attribution window and reporting attribution model the same as in the last report?
• Did you keep platform ROAS and decision-grade ROAS out of the same summed table?
• Do GA4, platform, and back-office time zones all point to the same calendar dates?
• Did you explain how refunds, cancellations, VAT, and shipping land in each ROAS?
• Are AI’s suggested budget changes grounded in the standard ROAS, or leaning on platform reference numbers?

ROAS gaps aren’t going away. What should go away is walking into a meeting unable to explain them. Handing ROAS to AI in Neuro doesn’t mean AI crowns one “correct” number. It means your team decides which ROAS drives which decision—and AI holds to that standard every time.