Paid Search and SEO Are Not Independent
Pause branded paid search and most of those conversions come back through organic. The paid channel's cost per acquisition describes a transfer between two of your own channels, not a purchase of new demand.
Someone types your brand name into a search engine. They have already decided who they are looking for. At the top of the results is your ad, and just beneath it is your organic listing.
You are paying for that click. The question nobody asks in the quarterly review is what would have happened if the ad had not been there.
Channel reports cannot answer this
A channel report tells you what happened in that channel. Paid search conversions, paid search cost, paid search return. Every number is correct.
The trouble is that the interesting question spans two channels. If paid and organic compete for the same searcher, then a conversion appearing in paid may be a conversion that would have appeared in organic. That is a transfer between two of your own lines, and it looks identical in the paid report to a conversion you created.
The design that answers it
Turn the channel off in some places and watch total conversions.
Three arms across 60 markets, twenty in each: one continues as normal, one pauses branded paid search, one pauses non-brand paid search. 8 weeks of pre-period, 6 weeks of test.
Two arms rather than one because the answer is not the same for both, and a test that only pauses branded search invites the conclusion that all paid search behaves that way.
The measured outcome is total conversions from search, paid and organic together. Paid conversions and organic conversions are also tracked, but as diagnostics rather than as the answer.
What happened
- Paid branded
- Organic
- Total
Show the data
| Period | Paid branded | Organic | Total |
|---|---|---|---|
| W-8 | 94.5 | 100.2 | 96.7 |
| W-7 | 107.0 | 98.4 | 100.4 |
| W-6 | 104.0 | 104.5 | 103.3 |
| W-5 | 103.5 | 105.3 | 104.6 |
| W-4 | 103.2 | 98.5 | 100.5 |
| W-3 | 88.8 | 101.3 | 97.6 |
| W-2 | 97.3 | 98.8 | 99.3 |
| W-1 | 101.7 | 93.0 | 97.6 |
| W1 | 0.0 | 130.8 | 93.4 |
| W2 | 0.0 | 123.5 | 88.7 |
| W3 | 0.0 | 127.5 | 88.8 |
| W4 | 0.0 | 136.7 | 98.3 |
| W5 | 0.0 | 128.3 | 92.3 |
| W6 | 0.0 | 138.5 | 100.8 |
The coral line falling to zero is what the paid search report shows, and it is what would be escalated. The teal line is what happened to the business.
The numbers
Branded search. Pausing it cost 2,288 paid conversions. Organic picked up 2,042 of them, or 89%. Both of those are measured well: the paid collapse and the organic rise are large against the noise. The number the budget question needs is the difference between them, and it is not. Net loss to the business: 306 conversions, with a 95% interval running from a loss of 687 to a gain of 75.
That interval crosses zero. Over six weeks and twenty markets, this test cannot rule out that pausing branded paid search cost nothing, and it cannot rule out a loss of nearly 700.
As a share of what the channel was credited with, the point estimate is 13% incremental and the interval runs from about -3% to 30%. Reported cost per acquisition was $55. At the point estimate the cost per conversion the business kept is $412; at the pessimistic end of the interval it is about $184; at the optimistic end it is undefined, because a channel that lost nothing has no cost per conversion. What the test does settle is a ceiling. Even at the far end of the interval, at least 70% of the conversions credited to branded search would have arrived anyway.
Non-brand search. The same test, run on the same business, in the same weeks, gives a different answer. Pausing it cost 2,593 paid conversions and organic recovered only 24% of them. Net loss: 1,896 conversions, comfortably clear of zero. Reported CPA $67, true CPA $92, and 73% of the credit was real.
Why the branded arm did not resolve
The net loss is a small difference between two large, well-measured numbers, and its noise is the noise of both. The design had ample power for the gross effect, which was never in doubt, and about a tenth of the power it needed for the net one. Detecting an effect one tenth the size at the same precision takes roughly a hundred times the market-weeks. The practical route is more markets or a longer window; until then the interval is the finding, and a ceiling of 30% incrementality is a usable one.
Why they differ
Someone searching your brand name has already chosen you. The ad and the organic result are two routes to the same destination, and removing one leaves the other. The ad is competing with your own listing.
Someone searching for a product category has not chosen you. Your organic listing may be on the second page or absent entirely. Remove the ad and the route disappears, so a much larger share of those conversions depended on it.
The general rule: cannibalization rises with how well the searcher already knows who you are. Brand terms sit at one end, generic category terms at the other, and competitor terms somewhere else again because there is usually no organic listing to fall back on at all.
What to do about it
Do not read paid search efficiency from the paid search report. That report answers a question nobody is asking.
Split branded from non-brand everywhere. In reporting, in budgeting, and in how the team talks about the channel. Blending them produces an average that describes neither, and it is the blended number that usually justifies the budget.
Test rather than assume. Published cannibalization estimates for branded search range from about half to nearly all, which is a wide enough range that adopting someone else’s figure is only marginally better than guessing. The design above is cheap: pausing one channel in a third of markets for six weeks.
Do not cut it to zero on this evidence. The test bounds branded incrementality below about 30% and does not separate 13% from zero. It also measures the effect of removing the channel entirely, which is not the effect of bidding less. The right question is usually where to set the bid, and that needs a spend-level test rather than an on-off test. The marginal return lesson is the argument for why.
Watch the competitive case. The standard objection is that a competitor will bid on your brand term if you stop. That is a real risk and this design measures it automatically, because if it happens during the test the organic recovery will be lower and the total loss larger. Six weeks may be too short for a competitor to react, which is a genuine limitation of the result rather than a reason to dismiss it.
Context
Best suited for
Any program spending meaningfully on branded search, especially where reported CPA is used to justify the budget.
Data requirements
- Conversions attributable to a geography, and to paid or organic within it
- Ability to pause paid search in selected markets without affecting others
- At least eight weeks of pre-period at market grain
- Roughly 20 markets per arm before the interval gets wider than the decision
What changes by context
- B2C ecommerce
- The clearest case, and where the brand term volume is highest. Watch for markets that share a media footprint, since a paused market receiving national brand advertising still generates the searches.
- B2C subscription
- Brand searches often come from existing subscribers looking for the login page, which inflates branded conversion volume with people who were never going to buy anything. Separate new from returning before measuring.
- B2B SaaS
- Long consideration cycles mean the recovery shows up outside a six-week window, so the test understates organic recovery and overstates the loss. Run longer or use an earlier proxy outcome.
- Marketing-led
- The finding usually lands on whoever owns paid search, whose targets are set on the blended reported CPA. Fixing the measurement without fixing the incentive rarely changes behavior.
When it breaks down
- Markets that share a media footprint, where national brand advertising keeps generating searches in a paused market.
- Very low brand search volume, where the interval will be wider than any effect worth acting on.
- Categories with aggressive competitor bidding, where six weeks is too short to observe the competitive response.
- Businesses whose organic listing is weak, since the whole recovery mechanism depends on there being something to fall back on.
What this does not establish
- That branded search should be switched off. It measures the effect of removing the channel entirely, which is not the decision most teams face. Where to set the bid needs a spend-level test.
- That branded search is 13% incremental. That is the point estimate; the interval runs from about −3% to 30%, and what the test establishes is that at least 70% of the credited conversions would have arrived anyway. Recovery depends on organic listing strength, brand recognition, and competitive bidding, so neither number transfers to another business.
- That the competitive response is accounted for. Six weeks may be too short for a competitor to notice and start bidding, which would make the true loss larger than measured.
- Anything about non-search channels. This is a substitution effect between two routes to the same searcher, and it does not generalize to display or video.
- That the recovery rate is stable over time. It was measured during one six-week window and is a behavioral quantity, not a constant.
What to test next
If this is new ground, the branded search pause is one of the highest-return tests available: large spend, a result that is usually dramatic, and a design that takes six weeks.
After it, the useful follow-up is not another on-off test. It is a spend-level test on the same channel, running two arms at different bids rather than one at zero, which estimates where the return falls off instead of just confirming that the floor is low.
How the numbers were produced Code
Three arms of 20 markets, 8 pre-period weeks and 6 test weeks, with market-level size variation and week-to-week noise. Effects are difference-in-differences against the control arm, computed on per-market weekly means so arms of slightly different total size stay comparable, with standard errors taken from the spread across markets because that is the level treatment was assigned at.
Recovery rates are an assumption of the simulation: 85% for branded and 30% for non-brand. The estimator recovered 89% and 24%. So this example demonstrates that the design measures the quantity it claims to, and does not constitute evidence about what any real cannibalization rate is. That number has to come from your own test.
Regenerated from seed 20260828 by the script named
under the figure.