How Digital Advertising Actually Works
April 2026
Every time you see an ad online, money has already changed hands. The transaction completed before the page finished loading, before you noticed the ad, before you had any chance of acting on it. Understanding what was actually purchased — and why — makes the economics of the internet make more sense.
The Impression
The unit of trade in digital advertising is the impression: one ad, shown once, to one person. When a brand runs a campaign, they are buying a quantity of impressions. A million-impression campaign means their ad was displayed a million times, across however many people and pages it took to accumulate that.
The standard pricing unit is CPM — cost per mille, where mille is Latin for thousand. A CPM of $5 means the advertiser pays $5 for every 1,000 impressions. The advertiser pays per exposure, not per result.
Why Pay for Something That Might Not Work
This seems counterintuitive until you consider that advertising has always worked this way.
A television spot costs the same whether the viewer buys the product or changes the channel. A billboard costs the same whether drivers notice it or drive past without registering it. A full-page magazine ad costs the same whether anyone reads it. The advertiser is not buying a guaranteed outcome. They are buying access to an audience, and they make a calculated bet that enough of those exposures will eventually produce purchases to justify the spend.
The math behind that bet is conversion rate: the percentage of people who see an ad and eventually buy. If an advertiser shows their ad to 100,000 people and 500 of them purchase, the conversion rate is 0.5%. The advertiser knows their average order value and their margin, and they work backwards to figure out the maximum they can afford to pay per impression while remaining profitable. If that number is higher than the platform's CPM, the campaign makes financial sense.
What Targeting Changes
Without targeting, those 100,000 impressions are distributed roughly at random across the platform's audience. Some will reach people who are genuinely interested in the product. Many will reach people who have no use for it, already bought something similar, are the wrong age or geography, or have no purchasing intent at that moment.
With targeting, the platform filters the audience. The advertiser specifies: show this ad to people who have been researching this product category recently, who match our typical buyer profile, who are in the right geography, who have behavioral signals suggesting they are close to a purchase decision.
If the untargeted conversion rate is 0.5% and the targeted conversion rate is 5%, the targeted impression is worth ten times more to the advertiser. They will pay a corresponding premium. The platform captures that premium. This is why behavioral data — the profiles built from everything you do on the platform — is not incidental to the business. It is the product that justifies the pricing.
How Advertisers Measure Whether It Worked
CTR (click-through rate): The percentage of people who saw the ad and clicked on it. A useful early signal, but clicking is not buying.
Conversion rate: Of the people who clicked, what percentage completed a purchase. Together with CTR this gives a full-funnel view from impression to sale.
ROAS (return on ad spend): Revenue generated divided by ad spend. A ROAS of 4 means for every $1 spent, $4 in revenue came back. Advertisers have a minimum target ROAS below which a campaign is not worth running.
CPM vs CPC vs CPA: Different campaigns buy different things. CPM buys impressions. CPC (cost per click) buys clicks — the advertiser pays only when someone clicks. CPA (cost per acquisition) buys completed purchases — the advertiser pays only when a sale happens. Each model shifts the financial risk differently. CPA is most favorable to the advertiser but hardest for platforms to offer reliably at scale.
The Attribution Problem
Here is the part that keeps entire teams employed: how do you know which ad caused the sale?
A typical consumer sees multiple ads for a product before buying. They might see a display ad on a news site Monday, a YouTube pre-roll Wednesday, a sponsored search result Friday, an Instagram post Saturday — and buy Sunday. Which ad gets credit? How you answer that question changes how you allocate budget across channels, which changes which platforms get money.
The industry uses attribution models to make this judgment. Last-click attribution gives all credit to the ad clicked immediately before purchase. First-click gives credit to the first ad seen. Multi-touch models distribute credit across all touchpoints proportionally. Each model produces different budget recommendations, and none of them is provably correct.
The attribution problem is genuinely unsolved at the industry level. Billions of dollars of ad spend are allocated based on models that are approximations. This is one reason why retailer data, covered in the third post in this series, commands such a premium — retailers can observe purchases directly and close the measurement loop in a way that most platforms cannot.
Part of a series. Previously: The 150ms Ad Auction and Engineered to Keep You There. Next: Ad Inventory and the Attention Economy.