Is branded search incremental? Conditions, not a verdict

Thomas Blake, Chris Nosko and Steven Tadelis ran a series of large-scale field experiments on paid search at eBay and published them in Econometrica in 2015. Their abstract reports that brand keyword ads had no measurable short-term benefits, and that returns from paid search overall were a fraction of non-experimental estimates¹.

Our view: that is the strongest public evidence in this debate, because it is a field experiment rather than a reading of attribution reports.

It still does not settle your case, for three reasons.

• eBay is not you. Your brand, your competitors and your customers are different. One marketplace's result does not carry over to every shop.
• Short-term only. The finding is about short-term benefits. Longer effects are a separate question that this result does not answer.
• Your search results page differs. What matters is what appears when someone searches your name. That includes who else bids and how your own listing ranks.

Take the eBay result as a reason to test, not as permission to switch off.

The results of single pauses point in different directions.

• A paid search agency, Closed Loop, reports a pause test for one client with no competitors bidding on the brand. It saw higher combined paid and organic traffic, sign-ups and purchases with paid brand ads on, and concluded that brand bidding was worth it². This is an agency that sells paid search, so it has an interest in the outcome.
• A marketer's newsletter post describes switching off branded search for 90 days for one lead-generation client. Total leads fell by about 5%, and most of the branded demand moved to the organic result³. The post reports no significance test.

Both are anecdotes. They differ because the brands, competitors, seasons and organic rankings differ. Both authors also point to the landing page: a brand ad can send people to a page built to convert, while the organic result usually leads to the home page²³. A pause on one site tells you about that site.

Single-client case studies are not benchmarks. Do not quote one as the typical result of pausing branded search campaigns. Treat each as a hypothesis about the conditions it came from.

Think about where the click would have gone if the ad were not there. Branded search means ads on searches that include your own brand name. Cannibalisation is paid ads taking credit for sales that would have arrived by another route. Here, the other route is your organic listing.

Our view, not a research finding: these are the conditions to check first.

Treat this as a checklist of what to look at, not a verdict. The agency case above found a lift with no competitors on the brand, which this table would not predict². The only way to know is to measure it on your own account.

When someone shows you a very high brand ROAS, be careful. Platform ROAS counts conversions the platform can link to an ad. A marketing mix modelling vendor writes that branded search can capture demand created elsewhere, so platform ROAS can sit well above incremental ROAS (iROAS)⁴. The same vendor says a lower incremental ROAS is not, on its own, a reason to switch branded search off⁴. Branded clicks come from people who already know your name, so a high figure does not show that the ad created the sale.

All three are forms of holdout test. Each trades cost against clarity.

• Pause. Switch off branded campaigns for a set period and watch total sales and organic clicks. It is cheap and quick. It is also easy to misread, because seasonality, promotions and other channels all move at once.
• Geo split. Pause branded ads in some regions and keep them on in matched ones. A geo test controls for time effects because both groups run at the same time. It needs enough orders per region to read. See how to run a geo holdout test.
• Schedule. Alternate on and off periods, for example by day or week. It gives several comparisons rather than one. It also gets noisy if demand swings through the week.

Whichever you pick, compare organic clicks and total orders, not just the paid line. The question is whether total sales dropped, not whether paid clicks dropped. Of course paid clicks drop when you switch the ads off.

A pause test is cheap, but it needs a reading rule agreed in advance. Write the rule before you switch anything off.

1. Define the outcome. Total orders and revenue, from your own shop data, plus organic branded clicks.
2. Set the threshold. What fall in total sales would make you keep the ads? What fall would you accept as noise?
3. Fix the window. Choose the length and dates before you start, and do not stop early because it looks good or bad.
4. Note what else changes. Promotions, new campaigns, stock-outs. Anything that moves sales in the window muddies the read.
5. Accept "can't tell". If the interval is wide, the honest result is "no clear difference". That is a proper outcome.

Example (illustrative): Branded search spends €10,000 a month and the platform reports €100,000 in revenue, a platform ROAS of 10. You switch it off for a month in half of your regions, matched against the other half. That half would have cost €5,000. Total sales there fall by only €2,500 against the control regions. That is an iROAS of 0.5. The real return should then be set against margin.

Decide on incremental margin, not on brand ROAS. ROAS is attributed revenue divided by ad spend. It measures revenue, not profit, and it counts sales that would have arrived anyway. iROAS is incremental revenue divided by ad spend. That is the better yardstick, and you only get it from a test.

Three practical outcomes:

• Little incremental effect. Cut or cap the spend, and put the money into something that adds sales.
• Clear incremental effect. Keep it. Branded search that defends against competitors can earn its place.
• Mixed. Keep a smaller budget, defend on the terms where rivals bid, and retest when conditions change.

Branded search also sits inside automated campaign types, which makes it hard to isolate. We cover that in Performance Max black box.

Our view: a branded budget that nobody has tested is a guess. A single test is cheap next to a year of paying for clicks you would have had free.