Relevance Beats Money: How Google Rebuilt the Ad Auction

In 1998, two Stanford graduate students published a paper describing a search engine they’d built called Google. Most of it is technical — PageRank, crawling, indexing. But buried at the back, in Appendix A, is a short section titled “Advertising and Mixed Motives,” and it reads less like computer science than like a warning.

Sergey Brin and Larry Page argued that ad-funded search engines are structurally corrupt. Their words: “the goals of the advertising business model do not always correspond to providing quality search to users.” They concluded that the mixed incentives were serious enough that it was “crucial to have a competitive search engine that is transparent and in the academic realm.”

Four years later, the same two men were running the most profitable ad-funded search engine ever built.

That’s not hypocrisy, exactly. It’s the more interesting thing: they found a mechanism that partly resolved the conflict they’d identified. That mechanism is why paid search works the way it does today, and why a two-truck HVAC company can still outrank a national franchise on the same keyword.

The first version didn’t work

Google launched AdWords on October 23, 2000, with roughly 350 advertisers. It was self-serve — you signed up with a credit card, no salesperson required — which was genuinely novel.

The problem was the pricing. AdWords sold on CPM. You paid for impressions, the same way you’d buy a banner or a magazine page. Google was handling around 20 million searches a day, and it was selling that inventory by the thousand.

Meanwhile, GoTo.com — covered in Part 3 of this series — had spent two years proving that advertisers preferred to pay per click. Paying for a click means paying for a response. Paying for an impression means paying for a chance. Direct-response advertisers had understood that distinction since the mail-order era, and they voted with their budgets.

Google also ran a second, older product called Premium Sponsorships: the ads at the top of the page, priced on CPM and sold by human salespeople. So by 2001 Google had two ad products, neither of which was built on the model the market actually wanted, and a real cash problem.

February 2002: the fix

AdWords Select launched in February 2002. Two changes, one of which mattered enormously.

The first change was expected: Google switched to cost-per-click. Advertisers now paid only when someone clicked, in a live auction, exactly as GoTo had done.

The second change was the one nobody else had made. Instead of ranking ads purely by bid, Google ranked them by bid multiplied by click-through rate.

Ad Rank = Bid × CTR

Read that formula slowly, because it’s the entire article. Under GoTo’s model, the highest bidder won the top slot, full stop. Under Google’s model, an advertiser with a lower bid and a better ad could beat one with a higher bid and a worse ad.

The internal work was done by a small group. Salar Kamangar — Google’s ninth employee, a 22-year-old Stanford biology graduate handling business planning — pushed for the auction and the self-serve model. Eric Veach, a Stanford PhD who’d done a stint at Pixar, worked out the math and led the technical side; he and Kamangar are named co-inventors on the resulting ad-ordering patent, with a provisional filed in September 2001. Hal Varian, the Berkeley economist Google hired as chief economist, became its public explainer.

Sheryl Sandberg, who ran ad operations, later described Eric Schmidt walking past her desk and asking, more or less constantly, how many advertisers Google had. This was not a confident company optimizing at the margin; it was one that needed the ad product to work.

It worked. Within months, the auction-priced Select ads on the right side of the page were regularly out-earning the human-sold Premium ads at the top — because Google’s system could tell which ad would actually make more money per impression, and the salespeople couldn’t. Google’s revenue went from roughly $19 million in 2000 to about $347 million in 2002 to roughly $1.5 billion in 2003, with about $350 million in operating income. The AOL deal that took effect in May 2002, reportedly guaranteeing around $100 million, ended the cash crisis outright.

Why the formula resolved the 1998 problem

Here’s the elegant part, worth understanding as mechanism rather than history.

CTR is a proxy for user satisfaction: if people click an ad, it was probably relevant to what they searched for. By multiplying bid times CTR, Google made its own revenue a function of that satisfaction. A huge bid times a tiny click rate is a small number, so Google made more money by showing better ads.

The 1998 warning had been that advertising and search quality pull in opposite directions. Bid × CTR partially stapled them back together — imperfectly, since Google still decides what “quality” means and has never published the full formula, but enough to change the industry’s economics permanently.

The pricing had a second layer. Google, following GoTo’s lead, used what economists call a generalized second-price auction: you don’t pay your maximum bid, you pay roughly the minimum needed to hold your position over the advertiser below you. This is why your Average CPC is almost always lower than your Max CPC, twenty-four years later.

One common misconception is worth flagging. People often call it a Vickrey auction, the kind where honest bidding is your best strategy, but it isn’t one. Benjamin Edelman, Michael Ostrovsky, and Michael Schwarz published the definitive analysis in the American Economic Review in March 2007, finding that generalized second-price auctions “look similar” to the textbook version but behave very differently: there’s no dominant strategy, and truth-telling is not an equilibrium. Google’s engineers built it in the wild rather than from a textbook, and the practical consequence is that a naive advertiser who bids their true maximum can end up overpaying.

That’s still true in your account today.

One clarification on Quality Score

The relevance weighting arrived in February 2002. The product called “Quality Score” arrived in mid-2005, with quality-based minimum bids, and landing page quality was folded in that December.

A lot of write-ups collapse these into a single event. They’re separate: 2002 is the concept, 2005 is the named feature with a number attached to it. The distinction matters for understanding why the modern system has the components it does — expected CTR came first, ad relevance and landing page experience were bolted on afterward as Google got better at measuring them.

What this means for your account

Three things carry directly forward.

You are not just buying clicks, you’re buying a quality assessment. Two advertisers bidding the same amount on the same keyword will pay different prices and get different placement. That gap is where the work is. Tighter keyword-to-ad-to-landing-page alignment lowers your cost per click without you touching the bid.

Outbidding is the expensive route to the same position. Improving relevance and outbidding are two paths to the same Ad Rank, but the pricing formula is asymmetric: quality improvements move your rank and lower what you pay, while raising your bid only moves your rank. When a competitor with a bigger budget shows up in your service area, the answer usually isn’t a bid increase.

Your max bid is a ceiling, not a price. Because of the second-price mechanism, setting a max bid at $45 does not mean paying $45. It means you’re willing to. Advertisers who don’t understand this either bid too timidly and never enter the auction, or bid without a ceiling and find out what the auction thinks they’re worth.

Everything since 2002 — Smart Bidding, Performance Max, AI Max — has been layers of automation on top of this same core. The auction underneath hasn’t changed shape.