Baseline sales vs incremental sales from advertising: how to tell them apart

Baseline sales are the sales you would have made without the advertising you are evaluating. Incremental sales are the difference between what you actually sold and that baseline. Simple to define. The trouble is that the baseline is a number from a world that didn’t happen, so it has to be estimated, and how you estimate it changes the answer by a factor of five or more.

This guide covers what goes into a baseline, the four common ways to estimate one, a worked example showing how each changes the result, and the usual suspects when advertising takes credit for sales it didn’t cause.

What baseline sales actually include

A baseline is not a flat line, and it isn’t last year’s number. It is everything that drives sales except the advertising you are measuring: the long-run trend, the season, prices, your menu or product range, the number of locations, the weather, local events, competitors, and the lingering effect of advertising you ran in the past.

One subtlety matters a lot. The baseline depends on the question. “What did this spring campaign add?” has a baseline that includes your always-on search and social. “What does all of our advertising add?” has a much lower baseline, one with no ads at all. Mixing those up is how two teams measure the same campaign and get answers that differ by double.

Why the baseline decides your ROI

Incremental sales are a small number found by subtracting two big ones. That makes them fragile. If a market sells $138,000 a week and the campaign’s true effect is $3,000 a week, a baseline that is off by just 2% (about $2,760) nearly doubles or nearly erases the result.

That finding isn’t a reason to give up. It is a reason to be careful about how the baseline is built and to report the result as a range. A careless baseline doesn’t just add noise; it usually adds bias in the advertising’s favour, because campaigns tend to run when sales are already rising.

Four ways to estimate baseline sales

Pre/post: the default, and usually wrong

Compare the campaign weeks with the weeks before. It is what most post-campaign decks do. It works only if nothing else changed, which is never. Campaigns launch in spring, before holidays, alongside menu launches. Pre/post hands all of that to the ads.

Year-over-year: better, still fragile

Adjust the pre-period by how much sales rose in the same weeks last year. This removes typical seasonality. It can’t remove anything new: a warmer spring, a price rise, a competitor that closed in March.

Control markets: the big step up

Use markets where the campaign didn’t run to tell you what happened in the same weeks. They lived through the same season, the same economy and the same national news. This is the logic behind every geo test. The weakness is choosing the controls: a simple average of all other markets includes some that look nothing like the one you are measuring.

Synthetic control: controls chosen by fit

Instead of averaging all steady markets, find the weighted blend of them that tracked the campaign market most closely before the campaign started. If the blend matched week by week for months, it is a credible stand-in for the campaign market afterwards. (Economists call this the synthetic control method. It was built for exactly this case: one market that changed and many that didn’t.)

A worked example: one campaign, four baselines

An illustrative pizza chain runs an eight-week campaign in one metro with four locations, adding $6,000 a week in paid social and streaming audio. In the eight weeks before, the metro averaged $120,000 a week in sales. During the campaign it averaged $138,000. Here is what each baseline says.

Over eight weeks, the synthetic-control read is about $24,000 of incremental sales (range $4,800 to $43,200) for $48,000 of advertising. Even at the top of the range that is $0.90 of sales per dollar, before food cost and labour. The pre/post read said $3.00 and would have had the chain doubling the budget.

Why did the methods disagree so much? Spring. The whole chain was growing. Last year’s spring was a bit softer, so year-over-year undercounted it. The simple average of other markets included slower-growing ones that look nothing like this metro. The weighted blend matched this metro’s pre-campaign pattern and grew 12.5%, leaving a small, honest slice for the ads.

Examples of ads taking credit for baseline sales

Most of the time, overstated lift isn’t anyone’s fault. Something else moved sales during the campaign and the baseline didn’t account for it. Research comparing standard observational ad measurement with randomized experiments at Facebook found the observational methods often failed to match the experimental answer, and more often overstated the effect than understated it (Source: Gordon, Zettelmeyer, Bhargava and Chapsky, Marketing Science, 2019). The usual culprits:

• Season. Campaigns launched into spring patio season, back-to-school or January gym resolutions ride a wave that was coming anyway.
• Promotions and price changes. A two-for-one running in the same weeks moves sales on its own. If the ads promoted the offer, you are measuring a package, not the media.
• New or remodelled locations. A store that opens mid-campaign adds sales to the market whatever the ads do. Leave it out of the read until it settles.
• Competitor changes. A rival closing down the road sends you customers. A rival opening takes them. Neither belongs to your media.
• Branded search and retargeting. They intercept people already on their way, so they look superb in platform reports and add little to the baseline.
• Regression to the mean. Campaigns often launch after a bad month “to fix it.” Bad months tend to be followed by better ones regardless. Pre/post reads that recovery as lift.

Common baseline mistakes

• Using last year as the baseline. Year-on-year growth includes price rises, new stores and the economy. It isn’t what the ads did.
• Letting the platform set the baseline. Platform-reported conversions implicitly assume a baseline of zero for everyone who saw an ad. That is the most flattering baseline possible.
• Choosing controls by geography alone. The next market over isn’t automatically a good comparison. Choose by how closely it tracked the campaign market before the campaign.
• Short pre-periods. Four weeks of history can’t show you whether two markets really move together. Use six months or more where you can.
• Reporting a single number. An incremental estimate without a range invites false confidence, and the range is often the most useful part.

When you can’t estimate a good baseline

Be honest about the cases where a baseline can’t be built well. A national campaign that runs at the same weight in every market leaves no unexposed markets to compare with. A brand with only three or four markets has too few comparisons for a reliable blend. A market where five things changed in the same month can’t be untangled after the fact. In those cases the fix is design, not analysis: plan the next campaign with deliberate variation across markets, as in a geo holdout test, so a clean baseline exists when you need it.

Questions people ask about baseline and incremental sales

Written by The Countlift team at First Byte in The Woodlands, Texas. Questions or corrections? Email hello@countlift.com.