Budget allocation across ad platforms: why fixed ratios keep failing you

It is the 28th of the month and you owe someone a number. Next month's €40,000 has to land somewhere across Meta, Google and TikTok, and you have three tabs open. One guide says 70/20/10. Another says 70 to 80% Meta for a business your size. The third says it depends, then recommends its agency.

Meanwhile your own account history is sitting right there, telling a story none of those guides have read.

I spent years managing budgets around €150,000 a month, and the monthly split was the decision I sweated most. Not because the math is hard. Because every public answer to it is a number that was averaged across businesses that are not yours.

The takeaways

• The 70/20/10 rule found in most budgeting guides is an average of other people's businesses. Your trailing 12 weeks of per-platform results beat any borrowed ratio.
• Cross-platform ROAS is three different metrics wearing one name. Each platform's attribution claims credit differently, so build a composite score per platform and compare those.
• A new platform needs enough spend to exit learning (Meta's documentation puts it at roughly 50 conversion events per week) before its numbers mean anything. A 5% token budget proves nothing either way.

How should you split ad budget across platforms?

Start from your own trailing performance, scored the same way on every platform, and let the split follow the evidence. The ratio is an output of your data, not an input you copy from a blog post. If a platform earned 55% of your results over the last quarter, it has earned the first claim on next month's money.

That sounds obvious written down. In practice almost nobody does it, because "performance" arrives in three incompatible dialects. Meta reports one attribution story, Google another, TikTok a third. So buyers fall back on rules of thumb like 70/20/10 (70% proven channels, 20% testing, 10% wildcards), which most budgeting guides repeat to each other in slightly different proportions.

A fixed ratio is not wrong, exactly. It is a prior. It is what you should do in week one, when you have no history. The mistake is still running someone else's prior in month nine, when 12 weeks of your own per-platform data could be doing the deciding.

Why doesn't comparing ROAS across platforms work?

Because each platform measures ROAS against its own attribution model, the three numbers in your dashboards are not on the same scale. Google Search often takes credit for demand that Meta created upstream. Meta's view-through window books conversions Google would call organic. Comparing the raw numbers rewards whichever platform grades its own homework most generously.

There is a second trap stacked on top: the average and the margin are different animals. A platform showing your best average ROAS may already be saturated, so the next €1,000 there buys less than the same €1,000 on the platform ranked second. The question for an allocation decision is always what the marginal euro returns, and a single trailing average cannot answer it.

And a third: small samples. A platform can look like your winner off two lucky weeks, the same way a creative can look like a winner off forty conversions. Splits built on a hot fortnight get reversed a month later.

How much budget does a new platform need before the numbers mean anything?

Enough to get its delivery system out of the learning phase and keep it there. Meta's documentation puts that at roughly 50 conversion events per ad set per week; if your CPA is €40, that is around €2,000 a week before the algorithm is even optimizing properly, let alone producing results you can judge.

This is why the classic 10% "testing" sliver fails quietly. Spread €4,000 of a €40,000 budget across a new platform's minimums and you often buy a verdict of "inconclusive" at full price. The honest options are two: fund the test above the platform's learning threshold for long enough to read it, or do not run it this quarter. A platform you cannot afford to test properly is not a platform you are testing. It is a platform you are donating to.

What does an evidence-based split look like in practice?

The version I built into Adscalr works like this: every platform's ads are scored on the same composite basis, the trailing 12 weeks of those per-platform scores set the Meta/TikTok/Google split, and the output is 3 to 5 prioritized campaign plans, each with a conversion goal and an AI-estimated CPI (labelled as an estimate, because that is what it is). When an account has no history on a platform, the system says so and falls back to a transparent low-confidence default instead of dressing a guess up as analysis.

What it deliberately does not do: reallocate budget in real time. I would not trust a tool that claimed to, since the marginal-return picture moves on a weekly rhythm while intraday numbers are mostly noise. What runs continuously is pacing: every 5 minutes, spend is checked against plan, with staged alerts at 150% of cap (runaway), 110% (overspend) and under 70% after midday (underspend). The split is a considered weekly or monthly decision. The watchdog is the always-on part.

If you want the longer version of how the planning works, the budget intelligence page walks through it, including the parts that are still advisory rather than automatic. Either way, the principle costs nothing to adopt today: pull your last 12 weeks per platform, score them on one common basis, and let that, rather than a stranger's ratio, write next month's split.