ROAS Rewards Harvesting, Not Growth

Your best-performing channel by ROAS is probably not growing anything. Branded search, retargeting and the rest of the bottom funnel report spectacular returns because they intercept people already on their way to you, which is demand capture, not creation. The canonical proof is eBay, which turned off its brand-keyword ads and kept 99.5 percent of the clicks for free because they flowed to organic search, at a measured return of negative 63 to 75 percent against a naive last-click read of positive 2,568 percent. That gap is the demand-capture illusion, quantified. A ROAS-maximizing optimizer always drifts down-funnel toward the last click, scaling the channels that harvest existing intent and starving the ones that create it, until it hits the ceiling of demand that already exists.

Why ROAS structurally rewards harvesting

The metric rewards intercepting a decision, not causing one. The last click before a purchase is where measured conversions concentrate, so a system optimizing for measured conversions always favors the channel closest to the transaction, whether or not it changed the outcome. Incrementality collapses on brand terms, where the user already intended to reach you, yet those post the highest ROAS. Attribution worsens it by crediting only the touch it can see. Refine Labs' twelve-month analysis of 21.5 million dollars in closed-won revenue found a roughly 90 percent gap between software attribution and self-reported surveys, with their podcast credited for 53 percent of revenue by buyers and zero by the software. The software credits the Google search that captured the buyer while ignoring the podcast that caused it. Treat the vendor figure as directional, but the mechanism is real.

Why eBay proves it, except when a rival is bidding

When eBay stopped bidding on its own brand keywords, it lost almost nothing. In the 2015 Econometrica field experiment, halting brand-keyword ads left 99.5 percent of the forgone paid clicks immediately recaptured by organic, a near-perfect substitute for the paid version. Once the researchers removed the self-selection of intent that makes branded search look good, the measured return was negative 63 to 75 percent against a naive last-click read of positive 2,568 percent, buying clicks eBay would have had for free. Nothing was broken, the channel was simply harvesting demand the brand had already created. This is the extreme case, so read it as the upper bound rather than proof that branded search never works, because branded search is genuinely incremental when a competitor is bidding on your name. A replication at Edmunds found only about 50 percent substitution rather than eBay's 99.5, and under heavy conquesting Edmunds lost up to 72 percent of its branded traffic when it stopped bidding, because rivals were there to catch the customer it had earned. Practitioner holdouts agree, with only 10 to 20 percent of platform-reported branded conversions truly incremental where the auction is uncontested. The rule is not to kill branded search but to split it, hold out the uncontested terms and cut them if incrementality is low, and keep bidding where you block a rival. Defensive branded search is real, uncontested branded search is mostly a tax to intercept your own free traffic.

Why capture has a hard ceiling

No amount of bid optimization creates a new buyer, so capture is bounded by demand that already exists. The Ehrenberg-Bass estimate popularized by the LinkedIn B2B Institute is that only about 5 percent of business buyers are in-market in a given quarter, from a roughly five-year repurchase cycle, leaving 95 percent who will not buy for months or years. Capture channels can only reach that 5 percent. Growth comes from building memory in the other 95 so that when they enter the market later, you are already in their consideration set, which capture cannot do, since brands grow by reaching new and light buyers, not squeezing the ones already converting. The 5 percent figure is a heuristic, not a law, and it shifts by category, but the direction holds, the pool you can harvest is finite and the one you can create in is far larger.

Why share of search is the growth metric to watch

If ROAS measures harvesting, share of search measures creation, and it leads the outcome you care about. Share of search, a brand's share of all category-branded search volume, correlates with market share and tends to move ahead of it, by roughly twelve months in cars and six in phones in Les Binet's work, and an IPA analysis of 30 case studies found an 83 percent correlation between the two. The gap between your share of search and share of market, the extra share of search, is the strongest single indicator of where market share is about to move. The key shift is that branded-search volume is an output of upper-funnel work, not a channel to optimize, the footprint of demand you created, not the creation itself. This is measurable cheaply with AI, an LLM with code execution computes your monthly share of search from Google Trends, charts it against revenue, and calculates the lead time. Be honest about the limits, it is correlational, a brand crisis spikes search the wrong way, not every search is demand, and it breaks down in dynamic-pricing categories and for low-volume brands.

Why the market inverted the rule and paid for it

The evidence-based split favors brand, and the market has been moving the other way at a measurable cost. The IPA analysis of around 1,000 campaigns found a roughly 60 to 40 brand-to-activation split maximizes combined short and long-term profit in consumer marketing, and nearer 46 to 54 in business marketing, which leans more on activation. Yet WARC and Deloitte data show budgets went from 59.9 percent performance in 2023 to 68.8 percent in 2024, cutting brand to 31.2 percent, though the same marketers named 50 to 50 ideal. The cost is quantified. Analytic Partners found that moving from performance-only to a brand and performance mix improved total revenue return by an average of 90 percent, while the reverse cut it 40 percent. Airbnb is the proof of the flip side. When it cut marketing 58 percent in 2020, mostly from performance, its traffic returned to about 95 percent of the prior year with marketing near zero, over 90 percent of it direct or unpaid, and revenue passed pre-pandemic levels within a year, its brand having become a noun and a verb whose demand no longer needed to be repurchased every quarter.

The capture-versus-creation audit (copy this)

Run this before you scale another bottom-funnel channel.

• Classify every line of spend as capture (branded search, retargeting, bottom-funnel, review sites) or creation (broad-reach video and audio, category education, PR, organic social, creator content), and compare the ratio against 60 to 40 for consumer or 46 to 54 for business.
• Compute share of search and extra share of search. Pull category and competitor volumes from Google Trends, compute your share monthly, and chart it against market share as the growth signal, not last-click ROAS.
• Hold out branded search, split by competition. Run a geo holdout separating terms a competitor bids on from terms only you bid on.
• Cut the uncontested waste. If a holdout shows under 20 percent incrementality on an uncontested cluster, stop paying and move the budget to creation.
• Rebalance gradually. Shift 10 to 15 points of budget per quarter from saturated capture toward creation over 12 to 18 months, judged by share of search and direct traffic, not campaign ROAS, which looks worse before growth compounds.
• If you are a pre-awareness startup, invert the sequence. Win non-branded category capture first, which can be genuinely incremental, and defer generation until you have margin and scale.

The full workflow, computing share of search and extra share of search from Trends and Search Console, classifying spend into capture versus creation against the benchmark, and separating created from harvested demand, is packaged as a reusable Claude skill. Get the free skill.

What to do Monday

Stop reading blended ROAS as a growth number. Split your spend into capture and creation and compare it against the 60 to 40 or 46 to 54 benchmark, because if you are inverted toward activation, you are heading for the performance plateau where immediate lift gives way to decline. Build the share-of-search time series for you and your competitors on the same dashboard as revenue. Then run one geo holdout on your branded search, split contested from uncontested, and move the uncontested waste to creation. The move will make your reported ROAS look worse and your business get healthier, and the only way to see that is to measure the demand you create, not just the demand you harvest.

Sources: Blake, Nosko and Tadelis, the eBay paid-search field experiment, Econometrica, 2015; Coviello, Gneezy and Goette, the Edmunds branded-search replication, 2017; John Dawes and the Ehrenberg-Bass Institute for the LinkedIn B2B Institute, the 95-5 rule; Les Binet, share of search, EffWorks Global 2020, and James Hankins's IPA share-of-search analysis; Binet and Field, the 60:40 brand-to-activation split, IPA; WARC and Deloitte on budget share, and Analytic Partners' ROI Genome; Refine Labs, the Attribution Mirage study; Airbnb Q4 2020 earnings and SEC filings.