The microtargeting pitch, and why more data about your users usually isn't the unlock

There's a pitch I've heard in a dozen forms over the years, and it always starts the same way: we know more about your users than you do. Some new tracking technique, some new model, some new signal nobody else is using yet. It sounds compelling on paper. After actually building a few of these systems, I'm a lot less convinced.

The client isn't bound by how much you know

The classic version of the pitch: your product has some unprecedented level of detail about a client's users. What they click, what they like, what they'd respond to. Here's the problem — the client isn't really bound by how much detail you have about their users. They're bound by the content they actually have to serve, and the capabilities of whatever system is serving it.

Will knowing more about a user really tell you more than what that user has already clicked on? More data can make a recommendation more precise. But precise about what, exactly, if the underlying options haven't changed?

Put another way: two competitors in the same space usually aren't separated by whose recommendation algorithm is smarter. They're separated by who has better content, better product, better distribution. The algorithm is fighting over the last few percentage points once everything else is already decided.

Even the companies with the most data don't need it that badly

Google and Facebook have more detail about you than anyone. And yet the ads are still frequently irrelevant, because they're serving whatever content advertisers are actually paying to show — the depth of the targeting data caps out against the shallowness of the inventory. Netflix has a genuinely enormous catalog, and uses microtargeting to suggest what you watch next. But notice what's actually being mapped there: detail about the content, joined against a comparatively thin signal about you. It's not that user data doesn't matter. It's that it's rarely the bottleneck people pitch it as.

The math on who can even use this

Say the clever algorithm is real and it works. There's still a finite number of organizations sophisticated enough to act on it. Some of them already have their own version. For a handful of exceptionally high-stakes, high-budget operations — a national political campaign, a company where a single conversion-rate point is worth real money — a sharper targeting layer is worth paying for. For most mid-sized companies, a 10-20 point bump in conversion is noticeable, not transformative. What actually changes their business is volume through the funnel in the first place, which a targeting model doesn't create.

What it costs to build, versus what it costs to buy

A generic version of this product has to scrape from everywhere (the audience could be anywhere), integrate with every client's internal data in every possible shape, and still ship an interface simple enough that someone acts on the output. A custom, one-off version for a single client scrapes one or two networks, integrates one or two internal datasets — maybe a single SQL query — and gets explained in person instead of through a dashboard. The custom version is usually cheaper to build and more likely to actually get used. That should tell you something about which one is the real product.

Same pitch, new decade

I wrote most of this years ago about "Big Data" personalization startups. The branding has changed — it's an AI-powered personalization layer now instead of a big-data one — but the shape of the pitch, and the shape of the skepticism, hasn't moved at all. More signal about a user is not automatically more value for the business serving them. If nothing else about the product changed, don't expect the targeting to carry it.