Think Bayesian without the equation
Start with what you already know, ask how well new evidence fits each explanation, and adjust your confidence. One brand search question shows the whole habit.
By Jason Karpeles Senior Principal, Data Science 3 min read
The question
Does advertising on your own brand name generate additional sales? One explanation is that customers already want your brand and would reach you through an unpaid result anyway. Another is that competitors could intercept them, so your ad protects purchases you'd otherwise lose. Both can be true for different customers.
Start with a belief, then move it
The useful question is: if our ad disappeared, where would those customers go? A purchase protected from a competitor counts as an incremental sale, even though the customer's interest existed before the ad.
Illustration: if the brand ad is removed, customers who would click the unpaid listing add little, while customers a competitor would win show what the ad protects
Here's the habit with rough numbers. They're an illustration, not data.
- Your starting belief. Your brand is well known, so you're about 70% confident most of these customers would find you anyway.
- New evidence. You learn that competitors' ads appear on most searches for your brand name.
- How well it fits. That's about three times as likely to happen if customers are being intercepted than if they aren't.
- Your updated belief. Your confidence that customers would find you anyway drops from 70% to about 44%.
Figure 1
One piece of evidence moves the belief
Confidence that most customers would find you without the brand ad
Illustration: confidence starts at 70%; evidence three times as likely if customers are being intercepted moves it to about 44%
You didn't need the equation. You needed a starting point, and an honest answer to "how much more likely is this evidence under one explanation than the other?"
Research shows why circumstances matter. Experiments on Bing found that branded ads generated relatively little additional traffic when competitors were absent, and could protect traffic when competitors were present. The study measured clicks, so it doesn't establish how much revenue a particular advertiser would protect.
A large field experiment at eBay pointed the same way for a very well-known brand: its brand-keyword ads showed no measurable short-term benefit (Econometrica). Your brand may not be eBay, which is the point. The evidence tells you where to start, and your own situation tells you how far to move.
Give more dependable estimates more influence
Now suppose you test it twice. Again, illustrative numbers.
Figure 2
Less uncertainty, more weight
Estimated lift from brand ads in two tests, each bar showing the range the test leaves open
Illustration, not data
Illustration: Test A estimates 4% lift plus or minus 1 point and gets 92% of the weight; Test B estimates 10% plus or minus 3.4 points and gets 8%; combined, about 4.5%
Test A is far more precise, so it gets most of the weight, and the combined answer sits close to it. This is weighting by the inverse of uncertainty: less uncertainty, more weight. Under suitable assumptions it gives the same estimate as a Bayesian calculation.
Precision isn't the only test. If Test A ran in a market with no competitors bidding on your name and Test B ran in one like yours, Test B may deserve more influence than its range suggests.
Check what your forecasts share
Say you want to forecast sales if you cut brand search spend, and you build three models. Combining them is an ensemble, an idea that goes back to Bates and Granger in 1969, and you can give more influence to the ones the evidence supports.
But look at what they share. If all three assume competitors won't change their bidding when you pull back, you have one assumption repeated three times, not three independent reasons for confidence.
Three forecasts built on different assumptions give independent reasons for confidence; three forecasts built on one shared assumption give one reason, repeated
Formal Bayesian model averaging follows specific probability rules for assigning weights. A simple ensemble doesn't automatically meet that standard.
Try it: a brand search holdout
- Split matched markets into a group where brand ads keep running and a group where they pause.
- Track total brand-term results, paid and unpaid together, so you see what's lost and not only what moved.
- Record how often competitors appear on your brand terms in each group.
- Run it long enough to get a range, not just a number, and write down your starting belief before you look.
Where Alli fits
Alli Scenario Planner applies the same reasoning to budget planning. Every forecast blends a media response model built on historical sales with a brand equity model built on current signals, and it shows a minimum and maximum for every channel, so teams can see how much the model knows before money moves.
Next: Why your MMM's range matters more than its number.
Frequently asked questions
What is Bayesian thinking in marketing?
It's the habit of starting with what you already know, judging how well new evidence fits each possible explanation, and adjusting your confidence to match. You can apply it to a test result or a forecast without writing down Bayes' equation.
Does bidding on your own brand name drive incremental sales?
It depends on where customers would go without the ad. If they'd click your unpaid listing, the ad may add little. If a competitor would win them, the ad protects sales. Research on Bing found branded ads added little traffic when competitors were absent and protected traffic when they were present.
How should you weigh two tests that disagree?
If both are credible and measure comparable conditions, give the more precise one more influence. This is inverse-variance weighting, which matches a Bayesian calculation under suitable assumptions. A test that matches your situation more closely may still deserve more weight.
Is a forecast ensemble the same as Bayesian model averaging?
Not automatically. An ensemble lets several forecasts contribute. Formal Bayesian model averaging assigns those weights by specific probability rules.
Sources
- Marketing Science, Competition and Crowd-Out for Brand Keywords in Sponsored Search, Andrey Simonov, Chris Nosko, and Justin M. Rao, 2018, 37(2), 200 to 215.
- Econometrica, Consumer Heterogeneity and Paid Search Effectiveness: A Large-Scale Field Experiment, Thomas Blake, Chris Nosko, and Steven Tadelis, January 2015, 83(1), 155 to 174.
- Operational Research Quarterly, The Combination of Forecasts, J. M. Bates and C. W. J. Granger, 1969, 20, 451 to 468.
- Statistical Science, Bayesian Model Averaging: A Tutorial, Jennifer A. Hoeting, David Madigan, Adrian E. Raftery, and Chris T. Volinsky, 1999, 14(4), 382 to 417.
Last updated November 5, 2026.