For at least a decade, companies have built personalization systems on the promise that first party data and machine learning could combine to anticipate customer needs and drive retention, engagement and revenue.

In zero interest rates, consumer businesses built data science and engineering teams to develop proprietary machinery they could sell to shareholders as fixed cost and OpEx investments for differentiated software assets toward e.g., Bill Gurley’s Perfect Business Model where each marginal use

• visit
• purchase
• engagement

delivered data that could in turn be transformed into machine learning systems that induce higher consumer switching costs.

The investment in these teams and infrastructure was all paid up front on the bet that these algorithms would pay off later in customer (and shareholder) value.

This worked when cost of capital was low, but companies are evaluating these bets with more scrutiny today – does personalization actually drive metrics that matter to customers and business?

There are a few reasons to carefully consider this today: “CAC is too damn high.” Privacy changes like those of iOS 14 and death of cookies and ID bridging are making advertising less effective and more expensive, with reports suggesting a 60% increase in CAC over the last 5 years.

When you acquire a customer, you want to keep them, and creating personalized experiences that lead customers to return is a clear way to do this. In January we declared 2024 is the year of Loyalty and Rewards and in May the Logged-In Web – both related tactics to keep customers engaged. Judging by TripAdvisor and Target Q2 earnings reported this month, we’d say we were right. TripAdvisor explained

The formula is simple: when we keep travelers engaged on our platform, we have more opportunities to monetize, not just through clicks but through higher value transactions, as well.

Generative AI is another reason to revisit personalization investments. Generative AI is poised to unlock the personalization we’ve long been sold – that which knows us (to the extent we wish to be known) and anticipates our needs. The stakes here are high

brands thrive or die based on conversion and retention.

explained Scott Belsky in May of last year. With generative AI Scott expects

retention for most brands will be dramatically stronger in just a few years.

That said, enterprise customers are taking a hard look at ROI on their AI investments. Lux Capital’s Josh Wolfe explained on Redpoint’s Logan Bartlett podcast:

People are looking to see: ‘Do we have price premium translation?’ So if you were Adobe and you were charging $29.99 for creative cloud .. are you getting another 10-20% premium on top of it that justifies pricing increase or are you also getting cost savings?

I think there’s a massive overreaction of expectations on the pricing power that people are going to get – that it’s just not going to translate – and the cost savings that people are going to get – that it’s not going to translate.

The story for (generative AI-powered) personalization is the same. Do brands get a lift from generative AI spend that can justify its costs?

No surprise, we are supremely excited about the promise of generative AI for personalization. We see it delivering better customer experiences with better cost structure (and better privacy properties!) than past personalization regimes.

In this blog, we’ll unpack how and why.

‍

It takes money to make money

The business of personalization is a simpler form of the business of advertising.

Following Google economist Hal Varian, the business of personalization is a one-sided “yenta”, a traditional Yiddish word for matchmaker.

Brands match people who want to buy things to things that the brand sells.

The need for personalization today is more pressing than when Hal Varian wrote his “The Economics of Internet Search” when there were 100 million web servers.

Last quarter Walmart reported having over 420 million unique SKUs for sale with the number of US marketplace sellers growing 36% on the quarter. Reducing CAC means relying more on product to engage loyal customers and become your own yenta: move personalization in-house than pay the search or social-ads toll-road.

Consumers are growing overwhelmed with options, and AI is poised as an incredible solution. AI is the ultimate scalable yenta.

The primary source of revenue for personalization is marginal revenue growth or reduction in customer acquisition costs.

But today great personalization is a scale-intensive business.

Building personalization systems is an expensive enterprise. They face high fixed costs of headcount and infrastructure.  Technologies from these efforts also don’t readily scale: one personalization application (e.g., predicting what movies a user will watch next) doesn’t costlessly apply to another (e.g., predicting what offer will prevent a user from churning).

Each instance of personalization may yield a small marginal increase in likelihood to buy. Since the marginal probability of incremental revenue from any instance of personalization is low, personalization must be applied at scale to have any hope of covering the initial fixed costs of headcount or infrastructure.

‍

Surplus turned profit

Language models transform the economics of personalization.

Whereas traditional personalization has high fixed costs, personalization with language models do not. With language models, you can leave the intelligence to the model, and only pay model marginal costs.

Training proprietary models is expensive and have unclear returns or durability.  Even well-capitalized data-focused companies like Bloomberg who’ve attempted to train their own model are seeing their specialist models overrun by more potent generalists.

UPenn’s Ethan Mollick explained

Remember BloombergGPT, which was a specially trained finance LLM, drawing on all of Bloomberg's data? You may not have seen that GPT-4 … beat it on almost all finance tasks.

It is part of a pattern - the smartest generalist frontier models beat specialized models in specialized topics. Your special proprietary data may be less useful than you think in the world of LLMs...

This means that the economics of personalization may reduce to simply the marginal costs of using or hosting them.

‍

AI marginal costs are going to zero

While many have worried about marginal costs of serving generative AI at scale versus other computing workloads, LM prices are putting personalization workload costs into more familiar territory.

In 2011, Boston-tech OG Bill Warner estimated it cost Google 1.19 cents to serve a Google search.  Taking this as an early benchmark and comparing it to Gemini 1.5 Flash’s published prices

• $0.075 / 1 million input tokens
• $0.3 / 1 million output tokens

for the same price as a single Google search you can, for instance, input half the tokens of JRR Tolkien's The Hobbit and output a fifth of the book. Setting 1.19 cents as the budget and graphing the input and output budget curve, you have incredible flexibility to use input and output tokens to ship rich personalization at the same price as Google once served a single search.

And that’s just today’s model token prices. In the 1.5 years since GPT3.5 launched, its price has fallen 40x. The cost of GPT4 level 1mm tokens has fallen 240x. A16Z’s Martin Casado expanded

Very clearly a supercycle where marginal costs of a foundational resource are going to zero.

Gemini 1.5 Flash pricing is attractive relative to other "yenta" benchmarks as well.

Consider purpose-built personalization technologies like Algolia and GCP Recommendations. For conservatively chosen LM-powered personalization token workloads, the market price of Gemini 1.5 Flash is cheaper than both Algolia and GCP Recommendations.

‍

Calculating personalization profit

All this makes optimal personalization easy to pencil out.

Personalization in a given session costs the number of input and output tokens. Its value to the business can be taken as the marginal increase in spend, which can be calculated via A/B tests. Assuming for simplicity input and output tokens cost the same, profit from personalization is easy to calculate

v(tokens) - cost(tokens)

where v is the marginal lift in revenue. You can solve for the tokens you should use following conventional price theory of setting marginal value equal to marginal cost.

And unlike traditional personalization, developing new types of personalization given identical infrastructure scales endlessly: all you have to do is change the prompt or context.

This has profound impacts for both consumers and businesses.

‍

New personalization economics

For Josh Wolfe’s economic concerns of AI, language models for personalization both provide premium translation and cost savings in the form of superior cost structure and lower marginal costs.  

Personalization with LMs have great privacy properties, but they also reduce the personalization exercise to an activation of relevant context and well-chosen prompts.

Removing the high fixed costs of personalization while enabling activation of more consumer context than ever with Crosshatch, we envision a path to a personalized internet that’s 10x better than today’s web but also cheaper!

How we get 10x better personalization is obvious. As we’ve previously written, today’s personalization runs on historical clicks, scrolls and the occasional conversion. With Crosshatch, brands can activate cross-app context with a prompt and just a few taps by the user.

Personalization on the web today is “shockingly bad”, said Upfront’s Peter Zakin last week. But great and safe personalization should be available to everyone, not just those who pay for it.

The true information superhighway is free and open. That includes personalization.

The web screwed up personalization in the past, but now with Crosshatch we have a real path to great personalization not just in a private network or a trusted hardware device but within any network, where anyone can turn on personalization anywhere, all in just a tap.

If you want to learn more about building futuristic personalization that delights users and drives your business forward, check out our docs and get started!