A large online marketplace switched off its brand keyword advertising and kept 99.5% of the clicks. The traffic simply arrived through organic results instead.
That experiment is the most cited evidence in this debate, and it is usually quoted to prove that paid search does not work. A second experiment, on a different design, found that companies which stop bidding on their own brand while competitors keep bidding lose between 18% and 42% of their clicks.
Both results are correct. This page reconciles them, and gives what the published experiments actually establish about the choice.
What happened when eBay switched paid search off
This is the study everyone half-remembers. The details matter more than the headline.
The brand experiment. In March 2012, eBay stopped bidding on all terms containing “eBay” on two search engines while continuing to pay on a third, which served as the control. The design is a difference-in-differences across platforms, which is why it can support a causal claim.
The result. “Almost all (99.5 percent) of the forgone click traffic from turning off brand keyword paid search was immediately captured by natural search traffic.”
Read that as a business fact rather than a statistic. The company was paying for clicks it would have received for nothing, because it already occupied the organic result those searchers were going to click anyway.
The non-brand experiment, which is larger and less quoted. Using geographic bidding controls across 210 media markets, eBay switched off all non-brand paid search for 60 days across roughly 30% of US traffic: 68 test markets against 142 controls.
The result there. Attributed sales fell 72% in the test markets, and the overall effect on actual sales was “very small and statistically insignificant”. The paper’s summary: “The entire regime of paid search adds only 0.66 percent to sales.”
The number that should change how you read any advertising report. Standard observational methods on the same data produced a return on investment “of over 4,100% without time and geographic controls, and a ROI of over 1,400% with such controls”. The experimental estimate was minus 63%, with a 95% confidence interval of minus 124% to minus 3%.
The limit the authors state themselves, and which is routinely dropped. “Our results show that for a well-known brand like eBay, the efficacy of SEM is limited at best.” And explicitly: “This may not be true for small and new entities that have no brand recognition.”
The counter-experiment, and why it does not contradict the first
If the story ended above, the answer would be simple and wrong. A second experiment tested the case eBay’s design could not.
The design. Four experiments on a search engine over nine days in January 2014, randomised at user level, capping the number of ads shown at zero, one, two or three against four in the control. The analysis covers 2,517 companies and 824 brands.
The finding when nobody else is bidding. “When no competitors are present, we find a positive, statistically significant impact of brand ads of 1%-4%, with larger brands having a smaller causal effect.”
Read that alongside eBay. eBay is about as large a brand as exists in its category, which puts it at the bottom of that range. A smaller company sits at the top of it. The two studies agree.
The finding that changes the decision. “For a set of brands that face competition on their brand search but choose not to advertise, competitors ‘steal’ 18%-42% of clicks.”
And when you do advertise, but competitors sit below you. “Competitors in paid positions 2-4 can ‘steal’ 1%-5% of the focal brand’s clicks and raise its costs.”
The conclusion the authors draw. “Under such position effects, we find the return on investment on defensive advertising to be strongly positive.”
How to reconcile the two studies in one sentence. Bidding on your own brand is close to worthless when nobody contests it, and clearly worth it when somebody does. The question is not whether to bid on your brand. It is whether anyone else is.
How to answer that for your own name, today. Search it, in an incognito window, from your market. If competitors appear above your organic result, you have your answer. If nobody does, you are probably paying for clicks you would get free.
Why almost nobody can measure this on their own account
Before you decide to settle the question with your own test, it is worth knowing what settling it actually requires.
The study. A review of 25 online field experiments run with large US retailers and brokerages, representing $2.8 million of advertising spend, each campaign covering more than 500,000 unique users and most of them over a million.
The finding, stated plainly by the authors. “A concise statistical argument shows that the required sample size for an experiment to generate sufficiently informative confidence intervals is typically in excess of ten million person-weeks.”
Why the number is so large. Individual purchasing is enormously variable relative to the per-person cost of a campaign. The paper reports an R-squared of advertising exposure on sales of about 0.0000054 for a campaign that was in fact highly profitable. The signal is real and it is minuscule next to the noise.
What that did to the 25 experiments themselves. The median standard error on return on investment was 51%. Nine of the 25 could not rule out a return of minus 100%. Only three could distinguish a return of plus 50% from zero, and the median campaign would have needed to be nine times larger.
The conclusion that follows for a small advertiser. “The weak informational feedback means most firms cannot even approach profit maximization.” If experiments at that scale struggle, a company running a few thousand euros a month is not going to resolve the question with a clean test.
What to do with that, rather than despair. Stop trying to prove incrementality and start using the cheap structural checks instead. The brand search test above costs nothing. Comparing paid conversions against total business over a long window costs nothing. Both beat an underpowered experiment.
And yet advertising does work
The literature that shows measurement is hard is often quoted as showing advertising is useless. That is a misreading, and the same authors provide the counterexample.
The experiment. A national US retailer matched its customer database against a large web platform, producing 1,577,256 individuals randomised into 81% treatment and 19% control. A 14-day campaign delivered 32.3 million impressions to 814,052 exposed users.
The result. “The advertising profitably increases purchases by 5%.”
The two details that matter more than the 5%. “93% of the increase occurs in brick-and-mortar stores” and “78% of the increase derives from consumers who never click the ads.”
The economics. The authors estimate “the retailer’s incremental revenues were more than seven times the cost of the ads.”
What this establishes, and its limit. Online advertising can produce large, real, profitable effects that are almost entirely invisible to click-based measurement. The population here was the retailer’s existing customers, not the general public, so it is evidence about reactivating a known audience rather than about acquiring strangers.
The uncomfortable synthesis. Advertising works, most of its effect does not produce a click, and the instruments most companies use to judge it measure clicks. That is the actual state of the field, and it explains a great deal of the disagreement in this debate.
What organic actually costs, measured in time
Organic is described as free, which is wrong, and as slow, which is right but rarely quantified. There is one study with a declared sample.
The design. A million random URLs seen by a crawler in September 2023, tracked for a year, plus two million URLs created in October 2023, and 1.3 million random US keywords for the age analysis.
How many new pages make it. “Only 1.74% of newly published pages rank in the top 10 within a year (down from 5.7% in 2017).” On the filtered English-language subset, 6.11% made it within a year.
How fast, for those that do. “40.82% of pages that ranked in the top 10 did so within 1 month.” The distribution is bimodal: either quickly or never.
How old the incumbents are. “72.9% of pages in Google’s top 10 are more than 3 years old (up from 59% in 2017)”, and “the average #1 ranking page is 5 years old (up from 2 years old in 2017)”. Pages under a year old make up 13.7% of top-ten results, down from 22%.
The wider context from the same publisher. “96.55% of all pages in our index get zero traffic from Google, and 1.94% get between one and ten monthly visits.” The publisher flags its own sample as “somewhat biased towards the quality side of the web”, which makes the figure conservative rather than alarmist.
What this means for the choice. Organic is a two to three year investment whose failure mode is silent. Paid is an expense whose failure mode is immediate and visible. Those are different risk profiles, and a company that needs pipeline this quarter cannot substitute one for the other.
The cost figure nobody has. No institution publishes a unit cost for search optimisation work. What circulates comes from self-selected practitioner polls with no sampling plan. If you see a monthly rate presented as a market average, that is what it is.
Where the traffic actually comes from
One study publishes a channel mix with a declared method, and its shape is more useful than its decimals.
The method. Clickstream data across more than 50,000 sites, 17 industries, mobile and desktop, worldwide, January to December 2025, with “channel shares of traffic calculated as medians across tracked domains within each industry”.
The hierarchy. Direct dominates, followed by organic search, then referral, with organic social and paid search accounting for small single-digit or sub-one-percent shares of the median site’s traffic.
Where the growth is. Paid search grew fastest in relative terms over 2025, from a very small base, while organic search grew slightly and organic social declined.
Two internal inconsistencies you should know about before quoting it. The article states an annual organic figure that cannot be reconciled with its own total, and the channel shares sum to about 92% rather than 100%, which is expected for sectoral medians but means they are not a partition.
How to use it responsibly. Quote the hierarchy, not the decimals. Direct and organic carry most traffic on a typical site; paid is a small share of volume growing quickly. That statement is safe and it is enough to reason with.
Why direct dominating matters for this debate. A large share of what analytics calls direct is unattributed organic and word of mouth. It is the accumulated result of everything you did before, which is precisely the asset paid advertising does not build.
How to decide for your own case
The literature does not produce a winner. It produces a small number of questions whose answers decide, and you can answer all of them this week.
Does demand already exist? Look at search volume for the problem you solve, in the words your buyers use. If people are searching, paid harvests it immediately and organic harvests it in two years. If nobody is searching, neither channel creates demand, and the work is upstream in positioning and category education.
Is anyone bidding on your name? Search it privately from your market. Competitors above your organic result means defensive bidding pays. Nobody there means you are buying clicks you already had.
What is your time horizon, honestly? If you need pipeline this quarter, organic cannot deliver it and saying so is not defeatism. If you have two years and a content capability, organic compounds where paid does not.
Can you tell the difference between a click and a customer? If not, fix that before spending on either. Every experiment above turns on knowing what happened after the click, and companies that cannot answer it will misread both channels indefinitely.
What happens if you stop? Paid stops the day you stop paying. Organic decays slowly and keeps working for years, which is the whole argument for it and the reason its cost is systematically underestimated.
The framing that avoids the false choice. Paid buys attention you rent. Organic builds attention you own. Renting is right when you need the room tonight, owning is right when you will be there for a decade, and most companies need both in a ratio their cash position decides rather than their marketing philosophy.
Three checks that cost nothing and settle most of it
You cannot run an incrementality experiment. You can run these, and between them they answer most of the practical question.
The brand search check. Search your own name in a private window, from your market, on a phone and on a desktop. Count the paid results above your organic listing. Zero means your brand bidding is probably buying free clicks. Two or three competitors means it is defending revenue.
The stop-and-watch check on brand terms only. Pause brand keyword advertising for two to three weeks and watch total sessions and total enquiries, not paid ones. If the totals hold while the paid line collapses, organic absorbed the traffic, exactly as it did in the marketplace experiment. This is the single most informative test available to a small advertiser, it costs nothing, and it is reversible in an hour.
The long-window comparison. Take twelve months of total spend and twelve months of total new customers, ignoring attribution entirely. Compare that ratio across years or across periods when spend changed materially. It is crude, it has no confidence interval, and it is immune to every attribution problem in this article.
What none of these tells you. Whether non-brand paid search is incremental. That genuinely requires an experiment at a scale you do not have, and the honest position is to say so rather than to pretend a dashboard answered it.
How to hold the uncertainty without being paralysed. Treat non-brand paid as a cost of accessing demand you can otherwise reach only slowly, and judge it on cost per qualified enquiry against your margin. That is a decision you can make. Proving incrementality is not.
Where to go next
You want the buying mechanics. Media buying explained.
You are structuring the paid search side. Google Ads account structure for B2B.
You are choosing between platforms. LinkedIn Ads vs Google Ads, and how to split budget between Google and Meta.
You are splitting brand and performance budget. Brand vs performance in B2B.
Your organic traffic is falling and positions are stable. Click-through rate benchmarks.
You want to know which metric to run on. ROAS vs MER vs CAC vs LTV.
In short
- A marketplace switched off brand keyword ads and kept 99.5% of the clicks, because it already held the organic result those searchers were going to click.
- On the same account, the whole paid search regime added 0.66% to sales, and the experimental return on investment was minus 63% where regression said over 4,100%.
- The authors state their own limit: this “may not be true for small and new entities that have no brand recognition”.
- A second experiment found the opposite where competitors bid: brands that do not defend their own name lose 18% to 42% of clicks to those competitors.
- Both results are right. The question is not whether to bid on your brand, it is whether anyone else is.
- Measuring incrementality yourself is usually impossible: the sample required “typically exceeds ten million person-weeks”, and nine of 25 large experiments could not rule out a return of minus 100%.
- Advertising still works. One experiment on 1.58 million people found a 5% lift, 93% of it in physical stores and 78% among people who never clicked.
- Organic is slow and silent: 1.74% of new pages reach the top ten within a year, and the average first-ranking page is five years old.
Paid buys attention you rent. Organic builds attention you own. Book a diagnostic, or see how we approach B2B paid acquisition.