Scaling an account does not make it work better. It makes whatever is already true about it larger. If a store converts poorly and its product page stalls when someone taps the size selector, doubling the budget buys twice as much of that. A leak scaled is a leak funded.
So I treat a budget increase as a multiplication decision rather than a growth decision, and I want to know what is being multiplied before I sign off. Across 22 years and more than $8M in managed ad spend, that check has narrowed to four numbers. Not four dashboards. Four numbers, each capable of being wrong in a way the account’s headline metrics actively hide.
One: how much of the conversion data was observed, not inferred
Every bidding system is a compression of the conversion feed you hand it. Feed it something stale, partial or quietly estimated and it will optimise toward a picture of demand rather than demand itself. At low spend that is survivable, because low spend barely exercises the model. At high spend it is the entire mechanism.
The number I want is the split between conversions the platform observed and conversions it reconstructed. Google has been unusually candid about the size of that reconstruction. When it launched conversion modelling through Consent Mode, it reported that modelling “recovers more than 70% of ad-click-to-conversion journeys lost due to user cookie consent choices.”
Read that from the other end. Those journeys were lost. Seven in ten come back as a statistical estimate, and the rest do not come back at all. That is the vendor describing its own best case. I am not against modelling; the alternative is a hole. I am against scaling on top of it without knowing how thick the modelled layer has grown.
In practice that means opening the conversion action rather than the campaign column: when it last recorded an unmodelled conversion, whether the action bidding uses is still the one anyone believes it is, and whether offline imports are landing or failing silently on a stale click identifier. Tag managers do not announce their own funerals.
Two: contribution margin per order, not revenue ROAS
Revenue ROAS is measured at checkout. Almost everything that destroys the value of an order happens after checkout: returns, shipping, payment fees, the discount code applied, the cost of goods on the item that sold. None of it is in the number most accounts are optimised against.
Returns alone are not a rounding error. The National Retail Federation, with Happy Returns, put total retail returns in the United States at $890 billion for 2024, with retailers estimating that 16.9% of their annual sales would come back.
That matters specifically at the moment of scaling, because scaling changes the mix. Shopping and Performance Max push incremental volume toward whatever converts most easily, which is frequently the cheapest item in the catalogue, often also the thinnest-margin one and the likeliest to come back. Blended ROAS holds perfectly steady while profit per order falls underneath it, because the algorithm did exactly what it was asked to do.
A lead-gen client of mine ran two campaigns with nearly identical cost per lead. The CRM said one produced leads that closed at a small fraction of the other’s rate. Until closed-won values flowed back into bidding, every scaling decision there used the wrong number.
Three: how the page responds on a phone, not how fast it loads on your desk
I have argued the economics of site speed elsewhere, so here is the narrow version: before scaling, I check Interaction to Next Paint on mobile, on the specific template the ads land on.
INP is not load time. It measures how long a page takes to visibly respond when someone taps something, so it measures the elements a paid visitor actually touches: the variant selector, the quantity stepper, the add-to-cart button, the first form field. A load-time test touches none of them, which is how a page scores well and still feels broken to the person you just paid for.
The 2025 Web Almanac, working from July 2025 field data, found that 77% of websites have good INP on mobile against 97% on desktop, and that only 48% of mobile sites pass all three Core Web Vitals versus 56% on desktop. The gap is the finding. The device you test on is the device that reports the flattering number, and it is not the device your traffic is using.
Four: the share of spend sitting on your own name
Brand search converts at rates non-brand will never reach, because the intent was manufactured somewhere else and the ad merely collects it. That makes blended ROAS a weighted average whose weights you control, and brand weight is the easiest way in the business to make an account look efficient without being more efficient.
How much of that brand efficiency is even incremental is an old question with an uncomfortable answer. When eBay ran large-scale field experiments that switched off its own brand search ads, the result, as Chicago Booth Review reported it, was blunt.
The ads for eBay had almost no effect — 99.5 percent of the traffic that would have come to the site via the ads ended up there anyway.
That number does not transfer cleanly to every advertiser. eBay owns a brand almost nobody can compete with, and a competitor bidding on your name changes the arithmetic. But the direction of the finding has survived a decade of attempts to argue it away, and it means the most efficient line in most accounts is partly buying traffic that was already arriving.
The scaling implication is sharper than the incrementality debate anyway. Brand volume is capped by how many people search your name, and a budget increase does not raise that number. So the incremental dollar lands in non-brand, whose economics are deteriorating. Dreamdata’s benchmark of non-branded B2B Google search, covering August 2024 to July 2025, found cost per click up about 29% to $5.34 while click-through rate fell about 26% to 4.04%. That is the environment the incremental dollar walks into, not the one the blended average describes.
On one e-commerce account, the blended return sat flat for months, and the client read flat as stable and asked to double spend. Brand’s share of the budget had been climbing quietly the whole time, after a competitor started bidding on their name and defending it got more expensive. The average held because the mix shifted, not because acquisition was working. Doubling the budget would have poured every new dollar into the half of the account nobody had looked at.
The one that gets skipped
Of the four, brand share is the one I almost never find already checked. Not because it is hard: it is a label and a filter, twenty minutes of work. The other three force themselves onto someone’s agenda eventually. Broken tracking breaks loudly, and a column that drops to zero starts its own conversation. Margin gets ignored for a while, but finance asks sooner or later, because finance reads the bank balance rather than the platform. Page speed at least has a free tool that grades it.
The brand split has none of that. It never generates a complaint, because its effect on every reported metric is flattering. It is the only one of the four whose failure mode is that the account looks good.
Here is the position I will defend and might be wrong about: revenue ROAS should be removed from the primary reporting view entirely, not supplemented with a margin column next to it. Two numbers in one row means the better-looking one wins the meeting, and revenue ROAS is always the better-looking one. I have had accounts where that was the wrong call, catalogues with flat margins where forcing cost data through the feed consumed more attention than it returned. Those are rarer than the people who claim to have one.
None of the four takes long. Signal integrity is an afternoon, margin is usually a week of data work, INP is a morning, brand share is twenty minutes. What takes long is unwinding a quarter of scaled spend that funded a leak nobody measured first.