What Marketing Spend Can and Cannot Buy

Key takeaways

• CAC rises structurally, not because you got worse: Meta delivered 12% more ad impressions in 2025 and its average price per ad still rose 9%, which it attributes to increased advertiser demand.
• Branded search mostly bills you for demand you already had: eBay's field experiments found brand-keyword ads produced no measurable short-term benefit.
• Measuring advertising honestly is harder than running it: across 25 large experiments, the median confidence interval on ROI was over 100 percentage points wide.
• Spend cannot buy demand that isn't there. Etsy spent 31.7% of 2025 revenue on marketing, raised it 6.8%, and consolidated gross merchandise sales still fell 5.3%.

What a marketing dollar actually buys

A marketing dollar buys attention. It does not buy demand. Everything difficult about paid acquisition follows from that one sentence, so it is worth being precise about it.

When you buy an ad, you are renting a moment of someone's notice. Whether that moment converts depends on things the ad did not purchase: whether the person has the problem, whether they have the money, whether your product is the obvious answer, whether they trust you enough to hand over a card. Advertising can find demand, it can remind latent demand, and over long horizons it can create preference. What it cannot do is manufacture a need in someone who does not have one, at a price that makes economic sense.

This is why two businesses can run the same campaign, on the same platform, at the same cost per click, and get results that differ by a factor of ten. The variable is not the media buying. It is the underlying demand, the fit of the offer to it, and what happens after the click. Operators who lose money on paid acquisition usually diagnose it as a media problem and hire a better buyer. Frequently it is a demand problem, a price problem, or a landing-page problem wearing a media costume.

The honest framing is that paid acquisition is a distribution mechanism with a market-clearing price attached. If your economics work, it lets you buy growth at a known rate. If they do not, it lets you discover that faster and more expensively than any other method available. Both of those are useful. Neither is a strategy.

Why CAC rises — the auction mechanism, with real numbers

Your acquisition cost rises over time even when you do nothing wrong, and the reason is structural. Ad inventory is sold at auction, and in an auction your price is set by the second-highest bidder, not by your own efficiency.

The cleanest public evidence sits in Meta's 2025 Form 10-K. Meta delivered 12% more ad impressions across its Family of Apps in 2025 than in 2024, and its average price per ad, which it defines as total advertising revenue divided by ads delivered, nonetheless rose 9%. In 2024 the same figures were +11% impressions and +10% price. Meta attributes the price increase to an increase in advertising demand, which it believes is mostly the result of improvements to its ad targeting and measurement tools.

Sit with that for a moment, because it inverts the intuition. Supply of ad inventory expanded by double digits. Price went up anyway. That only happens when demand is expanding faster than supply, and demand here is other advertisers, bidding against you, getting better at converting, and therefore able to pay more for the same impression. The platform's targeting improvements do not lower your costs. They raise the ceiling of what your competitors can afford to bid, and the auction hands that surplus to the platform.

The implication is uncomfortable and load-bearing: any efficiency you discover is temporary unless it is proprietary. A creative angle that halves your cost per acquisition is visible to competitors within weeks. A targeting configuration is replicable. The efficiency gets bid away, and the equilibrium re-forms at a CAC that leaves the marginal advertiser roughly breaking even. The advertisers who persistently earn above-normal returns are not the ones with better media buying. They are the ones with something the auction cannot arbitrage: a higher lifetime value per customer, a genuinely better conversion rate, a margin structure competitors cannot match, or a demand source that is not an auction at all.

The same dynamic now runs inside marketplaces. Amazon's advertising services revenue was $68.6 billion in 2025, up from $46.9 billion in 2023, a 46% increase in two years, paid largely by sellers bidding against each other for position in front of buyers the marketplace already had. Visibility that used to be earned by sales rank is increasingly rented, and the rent clears at auction there for exactly the same reason it does everywhere else.

Demand capture versus demand creation

Split your marketing into two jobs, because they behave completely differently and confusing them is the most common expensive mistake in the discipline.

Demand capture intercepts people who are already looking. Search ads on high-intent keywords, marketplace placements, comparison sites, retargeting. It converts well, it measures easily, and it is capped: you cannot capture more demand than exists. Its returns look spectacular in the dashboard precisely because it is standing in front of a queue that formed for other reasons.

Demand creation makes people want the thing before they were looking. Broad-reach advertising, content, PR, category education, distinctive brand assets. It converts poorly in-session, measures badly, and pays off on a lag measured in quarters or years.

The attribution systems most operators rely on can see the first and are nearly blind to the second, which produces a predictable failure: budget migrates toward capture because capture reports better numbers, the pipeline of people who already want the product slowly empties because nothing is refilling it, and capture costs rise because you are bidding harder for a shrinking pool. It looks like an efficiency problem. It is a demand problem that has been building for a year.

The IPA's databank work by Les Binet and Peter Field is the most-cited attempt to put a ratio on this. Their 2013 report “The Long and the Short of It” put the optimum budget split at roughly 60:40 between brand building and sales activation; their 2018 follow-up “Effectiveness in Context” landed at 62:38 on a later set of campaigns. Treat those as a centre of gravity, not a prescription: the right split moves with category, growth stage, and margin. What the work establishes is directional and robust: an all-activation budget maximizes this quarter and erodes the next several, and the erosion is invisible in the channel reporting that caused it.

A useful diagnostic, and one you can run today: what share of your paid acquisition is branded search plus retargeting? Those are pure capture: they bill you for intent you already generated. If they are most of your spend and your unattributed or direct traffic is flat, you are harvesting a field nobody is planting.

The measurement problem is worse than the industry admits

Almost everyone measuring advertising is measuring correlation and reporting it as causation, and the peer-reviewed work on this is blunter than any vendor will be.

Start with the sharpest result. Blake, Nosko and Tadelis ran a series of large-scale field experiments at eBay, published in Econometrica in 2015, that turned paid search ads off for randomized populations. Their finding on branded keywords is unambiguous: brand-keyword ads produced no measurable short-term benefit. The people clicking the ad for eBay after searching for eBay were going to eBay regardless; the ad was a toll on traffic the company already owned. For non-brand keywords the picture was more nuanced and still sobering. New and infrequent users were positively influenced, but frequent users whose behaviour the ads did not change accounted for most of the spend, producing negative average returns overall.

That is one company in one period and should not be generalized into a rule that branded search is always waste. It should be generalized into a method: the only way to know what an ad caused is to withhold it from a randomized group and compare. Everything else (last-click attribution, multi-touch models, platform-reported conversions) measures who saw the ad and also bought, which is exactly the population most contaminated by selection. Targeting makes this worse, not better: the better a platform gets at showing your ad to people likely to buy, the more your attributed conversions consist of people who would have bought anyway.

Then there is the hard part. Lewis and Rao, writing in the Quarterly Journal of Economics in 2015, analysed 25 large field experiments with major US retailers and brokerages representing $2.8 million of digital ad spend. The median confidence interval on return on investment was more than 100 percentage points wide. Individual-level sales are volatile relative to per-capita ad cost, with a coefficient of variation of 10 being common, so an informative experiment can require more than ten million person-weeks. Their conclusion is the one operators need: for many firms, precisely measuring advertising ROI is not merely difficult, it is infeasible at their scale.

So what does a small operator do with that? Not despair, and not fake precision. Three things. Run holdout tests on the largest single line of spend, where the effect is big enough to detect. Watch aggregate ratios that do not depend on attribution at all: total marketing spend divided by total new customers, tracked monthly, is crude and honest, and it cannot be gamed by a platform's conversion window. And treat any channel report claiming a 9x return as a hypothesis, especially when the platform selling the ads is also the one grading them.

The arithmetic to run before you spend

Before a campaign, you need one number: the most you can pay for a customer and still make money. Everything downstream is a comparison against it.

Work it in the order that keeps you honest. Start with contribution margin per order: price minus product cost, payment processing, packaging, shipping, and a realistic returns allowance. Then decide the horizon you are willing to underwrite: first order only if you are cash-constrained, first 90 or 180 days of margin if you have real repeat data.

Illustrative only, with round numbers chosen for legibility. Say average order value is $80 and contribution margin is 40%, so $32 of margin per order. Say a customer places 2.5 orders in their first year, an observed figure from a real cohort, not an aspiration. First-year contribution is $80. If you require a 3:1 return on acquisition, break-even CAC is about $27. That $27 is the ceiling, and it is the number to compare against every channel.

Now apply the funnel. At a 2% landing-page conversion rate, a $27 CAC means you can pay $0.54 a click. If clicks in your category cost $1.80, you are underwater by a factor of more than three, and no creative testing closes a 3x gap. What closes it is conversion rate, order value, repeat rate, or margin. Lift conversion from 2% to 3.5% and the affordable click goes to $0.95. Add a second purchase to the first-year cohort and the ceiling moves again. This is the actual lesson of the arithmetic: paid acquisition is usually won upstream of the ad account.

The cash constraint deserves its own line. CAC is paid today; margin arrives over months. A 3:1 first-year ratio with an eleven-month payback is a fundamentally different business from a 3:1 with a two-month payback, and the difference is whether growth funds itself or consumes your balance sheet. Compute payback in months alongside the ratio, always. Then check statistical power before you act on a test: a difference between 2.0% and 2.4% conversion needs thousands of visitors per variant to distinguish from noise, and acting on an underpowered readout is how operators talk themselves into permanent changes that were coin flips.

What spend cannot buy

There is a class of problem no budget solves, and recognizing one early saves more money than any optimization.

Etsy's 2025 results are the clearest recent illustration in public filings. The company spent $914.8 million on marketing, 31.7% of its $2.88 billion of revenue, an increase of 6.8% over the prior year. Over that same year, consolidated gross merchandise sales fell 5.3% to $11.9 billion and active buyers fell 2.0% to 93.5 million. Nearly a third of revenue committed to marketing, spending up, buyers down. Airbnb shows a gentler version of the same shape: sales and marketing rose about 20.5% in 2025 to $2.59 billion while revenue rose about 10.3% to $12.24 billion, taking marketing from roughly 19.3% to 21.1% of revenue.

Neither company is badly run, and neither number is a scandal. They are what mature demand looks like: spend rises faster than results because the cheap demand was captured years ago and what remains costs more to reach. If two very large, very sophisticated marketing organizations cannot spend their way past that, a smaller operator should be sceptical of a plan whose central assumption is that more budget fixes it.

The specific things marketing spend cannot buy are worth naming, because each has a different remedy. It cannot buy retention: customers who leave after one purchase will keep leaving at whatever volume you acquire them, and a retention fix multiplies the value of every dollar you have already spent. It cannot buy product-market fit; advertising an offer people do not want simply reaches the disinterested faster. It cannot buy margin, and a business with structurally thin margins cannot outbid one with fat margins for the same customer, forever, no matter how clever the creative. And it cannot buy trust at speed, because reputation accrues on a schedule that money accelerates only slightly.

What spend can buy is genuinely valuable and worth being clear-eyed about: reach you would not otherwise have, speed you could not otherwise achieve, information about which messages and segments respond, and volume on top of economics that already work. That last clause is the whole discipline. Paid acquisition is a multiplier applied to unit economics. Applied to a business that makes money per customer, it compounds. Applied to one that does not, it is a faster way to run out of cash, and the dashboards will look encouraging most of the way down.

Put it to work

Compute your break-even CAC this week (contribution margin per order times first-year orders, divided by your required return) because that is the ceiling every channel is judged against. Translate it into an affordable cost per click at your real conversion rate; if the gap to market rates exceeds about 30%, fix conversion or order value instead of bids. Then run a one-month holdout on your largest line of spend.

Sources & references

Linked entries open the named source directly. Entries without a link say exactly what kind of reference they are — and how to check them yourself.

• Meta Platforms, Inc. — Form 10-K for fiscal year 2025 (ad impressions +12%, average price per ad +9%)
• Blake, Nosko & Tadelis, "Consumer Heterogeneity and Paid Search Effectiveness: A Large-Scale Field Experiment," Econometrica 83(1): 155–174 (2015) — NBER Working Paper 20171
• Lewis & Rao, "The Unfavorable Economics of Measuring the Returns to Advertising," Quarterly Journal of Economics 130(4): 1941–1973 (2015) — abstract and citation record
• Etsy, Inc. — Q4 and Full Year 2025 results (marketing expense, revenue, GMS, active buyers)
• Airbnb, Inc. — Q4 2025 Shareholder Letter (sales and marketing expense against revenue)
• Amazon.com, Inc. — Form 10-K for fiscal year 2025 (advertising services revenue by year)
• IPA (Institute of Practitioners in Advertising) — on Binet & Field, "The Long and the Short of It" (60:40) and "Effectiveness in Context" (62:38)
• U.S. Census Bureau — Quarterly Retail E-Commerce Sales, 4th Quarter 2025
• Worked examples are illustrative — The $80 order value, 40% margin, 2.5 orders and 2% conversion rate in the break-even section are made-up round numbers chosen to make the arithmetic legible. They are not benchmarks and should be replaced with your own measured figures.