How "Signal" Became Marketing's Password

Every category eventually finds a word that means “trust us.” For a while recently it was “AI-powered.” Before that, “data-driven.” Right now, in performance marketing, it’s “signal.”

Open any pitch deck in this space and the word is doing the same job everywhere: it’s standing in for a mechanism the vendor doesn’t want to explain. A signal, technically, is just a single observed data point: someone searched a term, someone’s location pinged near a store, someone’s card was charged at a merchant category code. That’s a precise, narrow, useful idea.

It says nothing on its own about whether that data point predicts anything, whether it’s been validated against an actual outcome, or whether the “audience” built from a thousand of them stacked together means what the deck says it means.

Somewhere in the past two years, it seems like the word has stopped explicitly meaning “a data point” and is increasingly being used to suggest “proof.”

A platform doesn’t say it built an audience from a purchased list anymore ; it “surfaces high-intent signal.” A number doesn’t come from a campaign with an unknown baseline; it comes from “the graph,” as if naming the infrastructure were the same as showing how it was validated. The word has absorbed the credibility that used to require showing your work, and it’s being used, across a whole category of vendors, in place of showing it. (For someone with a background in direct marketing, where we always have to show our work and validate our approach with results, this is like social media marketing all over again...)

Three things get smuggled in whenever “signal” replaces a specific claim:

First, scale disguised as precision. A platform citing six-figure signal counts or trillions of underlying data points sounds impressive and much more rigorous than one with fifty, but volume of inputs says nothing about whether any given input is predictive. Most of those signals are unused for any specific audience; citing the size of the haystack isn’t evidence about the efficacy of the needle.

Second, correlation dressed as discovery. When a vendor says a segment “over-indexes” some multiple on a given trait, that’s a real, checkable number - but it describes how the existing customer base looks, not how a new prospect will behave. Index and lift are different questions, and the word “signal” gets used for both without distinguishing them.

Third, a black box treated as a feature rather than a gap. Ask a lookalike-modeling vendor which traits actually drove a match, and you’ll often get some version of “the algorithm found it” - not because they’re hiding something, but because in a lot of these systems, nobody, including the vendor, fully knows. That’s true of the newest AI-native data platforms and it’s just as true of Meta’s own audience expansion tools (such as Meta's Advantage+ Audience model), which now treat your own input as a mere suggestion the system can override. “Signal” is the word that lets a vendor say that without it sounding like an admission.

None of this means the underlying data or models are useless - that's not what I'm saying. But It means the word "signal" has stopped functioning as a description and started functioning as a permission slip - permission to skip the question of how a number was produced, because the vocabulary already implies it was produced rigorously. And in many cases we just don't know if that is true.

The rest of this series really consists of trying to address just that question, asked repeatedly and in various forms, across every place the word shows up: what’s the actual mechanism, what was it tested against, and what does the marketer on the other end of the pitch still have to bring themselves?

𝙉𝙚𝙭𝙩 in this series: 𝙩𝙝𝙚 𝙝𝙖𝙣𝙙𝙛𝙪𝙡 𝙤𝙛 𝙜𝙚𝙣𝙪𝙞𝙣𝙚𝙡𝙮 𝙙𝙞𝙛𝙛𝙚𝙧𝙚𝙣𝙩 𝙖𝙧𝙘𝙝𝙞𝙩𝙚𝙘𝙩𝙪𝙧𝙚𝙨 𝙝𝙞𝙙𝙞𝙣𝙜 𝙪𝙣𝙙𝙚𝙧 𝙩𝙝𝙞𝙨 𝙤𝙣𝙚 𝙬𝙤𝙧𝙙, 𝙖𝙣𝙙 𝙬𝙝𝙮 𝙢𝙤𝙨𝙩 𝙤𝙛 𝙬𝙝𝙖𝙩’𝙨 𝙢𝙖𝙧𝙠𝙚𝙩𝙚𝙙 𝙖𝙨 “𝘼𝙄-𝙘𝙤𝙢𝙥𝙤𝙨𝙚𝙙 𝙖𝙪𝙙𝙞𝙚𝙣𝙘𝙚𝙨” 𝙞𝙨 𝙨𝙩𝙞𝙡𝙡, 𝙪𝙣𝙙𝙚𝙧𝙣𝙚𝙖𝙩𝙝, 𝙖𝙣 𝙤𝙡𝙙 𝙞𝙙𝙚𝙖 𝙬𝙞𝙩𝙝 𝙖 𝙣𝙚𝙬 𝙣𝙖𝙢𝙚.

Agree or disagree? What are your thoughts on this? Start a discussion in the comments below.