Opening hook
Three phrases. One market. Your AI would get them all wrong.
Thrift store. Vinyl records. SNAP benefits.
If you asked a frontier language model to sort these into demographic audiences — who uses each phrase, who would be alienated by it, who would be activated — the model would give you three confident answers, each one subtly incorrect in a way that costs you money.
"Thrift store" would be tagged as lower-income-resonant and recommended for working-class-targeting campaigns. "Vinyl records" would be tagged as mid-market hobbyist and recommended for millennial male targeting. "SNAP benefits" would be tagged as lower-income and poverty-adjacent, recommended for welfare-policy audiences.
Each of these reads is a semantic read — the read a model performs by consulting the dictionary-adjacent meaning of the phrase and the contexts in which it most commonly appears in its training corpus. Each of these reads is systematically wrong about actual audience behavior, and the gap between "semantic meaning" and "audience reality" is where most AI-driven marketing quietly loses its most persuadable customers.
This essay is about why.
The single-signal problem
Every frontier language model shipped in 2025 and 2026 reads language with one primary signal: the co-occurrence patterns of words in its pretraining corpus, filtered through the preferences of a small rater pool. When that model generates ad copy, segments an audience, or recommends messaging, it is reading the world through that single signal.
The problem is that language is not mono-signal. A phrase like "thrift store" means one thing in the dictionary-adjacent sense (a store that sells used goods at low prices) and another thing in the actually-how-it-behaves sense (a marker of a specific lifestyle cluster that, empirically, skews toward urban upper-middle-class sustainability-driven consumers).
When you act on the first signal without the second, you make decisions that are locally sensible and globally wrong.
Three worked examples:
Example 1: Thrift store
The dictionary read: "thrift store" is a lower-cost alternative to new retail. The word "thrift" literally means careful economy. The phrase lives in semantic space near "discount," "bargain," "used," "resale." An AI asked "is this phrase appropriate for upscale audiences?" returns no.
The behavioral read: thrift store traffic, geotagged and cohort-analyzed, concentrates in neighborhoods of higher median income than the national average. The specific stores that generate the highest search volume for "thrift store" queries are aesthetic-driven, urban, often in wealthy gentrifying neighborhoods — Housing Works in Manhattan, Beacon's Closet in Brooklyn, Crossroads Trading in LA and SF, Wasteland on Melrose. The modal buyer is upper-middle-urban, professionally creative, often in their late 20s to mid 40s, and motivated by a combination of sustainability values, aesthetic curation, and in-group signaling.
A progressive campaign targeting that specific demographic and using "thrift store" language in its copy is activating an authenticity signal the audience recognizes. A campaign removing "thrift store" because an AI flagged it as down-register is disabling that authenticity signal.
The semantic read and the behavioral read point in opposite directions. The AI that relies only on the semantic read makes the wrong call consistently.
Example 2: Vinyl records
The dictionary read: vinyl records are a retro music-playback medium. The phrase sits in semantic space near "retro," "hipster," "indie," "collector." An AI asked "who is the audience for this phrase?" returns millennial men, 25-40, music hobbyist, mid-market income.
The behavioral read: the vinyl-record market in 2025 is not a mid-market hobby. The average annual vinyl purchase volume for engaged buyers is $400-800; a sizable minority of the market consists of buyers whose annual vinyl spend is over $2,000. The specific consumption pattern — new-release vinyl of current artists, purchased on release day, played once and shelved — is an upper-middle-class aesthetic signal, not a hobbyist behavior. The pattern is closer to luxury-watch collecting or limited-edition sneaker buying than to stamp collecting. The modal heavy vinyl buyer owns a home, has a discretionary hobby budget in four figures, and is largely urban or inner-suburban.
A mid-market consumer campaign that uses "vinyl records" imagery on the assumption that it signals "accessible hobbyist" is communicating something entirely different: aspirational-luxury-adjacent, specifically inaccessible to mid-market buyers. The audience hears "this is not for you" and turns away. The AI read ("accessible hobbyist") and the behavioral read ("aspirational-luxury-adjacent") are nearly opposite, and the AI is consistently on the wrong side.
Example 3: SNAP benefits
The dictionary read: SNAP is a federal food assistance program. The phrase sits squarely in welfare-policy semantic space and maps to lower-income, policy-aware audiences. An AI asked "who searches for this phrase?" returns low-income households, policy researchers, social workers.
The behavioral read: the phrase "SNAP benefits" is used by two very different populations, in equal volume, across completely different contexts. One population is the user population — low-income households who interact with SNAP as a resource. The other is the policy population — upper-middle-class professionals who work in, legislate, report on, or advocate around SNAP. The two populations are demographically, geographically, and psychographically opposite. A single phrase, two audiences.
A campaign that uses "SNAP benefits" language assuming one audience (policy) will reach the other audience (users), and vice versa. Neither message converts. The AI's single-signal read assumes a unified audience that does not exist in the behavioral data. Only a signal that reads phrases in context of use — who is using the phrase, in what channel, with what correlated behaviors — can segment these two populations cleanly.
Why "just use better prompts" doesn't solve it
The instinct of most marketing-technology teams, when confronted with the above, is to respond: "We'll tell the AI to think harder. Better prompts will disambiguate."
They will not.
The reason is that the AI is not failing to disambiguate; it is failing because its underlying signal is one-dimensional. You cannot prompt a system to return a second signal it does not have. You can ask it to consider both readings, and it will confidently produce both readings — but it will produce them by hallucinating a behavioral read from the same semantic signal. The output looks like two signals. It is one signal, labeled twice. The same bias, expressed with more words.
The only durable fix is to pair the semantic signal with an out-of-pipeline behavioral signal — one whose provenance is the actual behavior of actual audiences, not the co-occurrence patterns of the training corpus. When both signals are present and can be compared, three useful things happen:
- Agreement cases (most phrases) confirm fast. When the semantic read and the behavioral read both say "Bentley = upscale" or "food stamps = lower-tier," you ship with high confidence and move on.
- Strong-disagreement cases (like "thrift store") become segmentation decisions. The phrase is cross-class. You don't pick one read; you segment your audience and use the phrase only where the behavioral read is positive.
- Mixed-magnitude cases become weighting decisions. Some phrases are weakly positive semantically but strongly negative behaviorally, or vice versa. You can now weight your copy decisions by the magnitude of each signal, not just the direction.
Single-signal AI cannot do any of this. It simply returns its one read and calls it an answer.
The operational implication
For a marketing leader reading this piece, the practical implication is a shift in how to evaluate AI copy tools. The right question to ask a vendor is not "how good is your AI?" The right questions are:
- How many signals does your system return per phrase? If the answer is "one," you have a single-signal system by another name — and it will miss all the cross-class phenomena described above.
- Where does your second signal come from? If the answer is "a second LLM pass," you have one signal labeled twice. If the answer is "behavioral data, audience telemetry, revealed preference" — you have a second signal, and the pairing is worth paying for.
- How do you surface disagreement between the signals? The most valuable output of a multi-signal system is not the consensus; it is the disagreement. A dashboard that shows you which phrases in your copy are flagged by one signal and not the other is a dashboard telling you which phrases need a segmentation decision. A dashboard that shows you only the consensus read is hiding the signal that matters most.
The macro point
The industry has spent three years treating "AI-driven marketing" as a throughput problem — how do we generate more variants, faster, at lower cost. The throughput gains are real. They are also consumed almost entirely by a quality decline that nobody has been measuring: the single-signal AI gets more phrases subtly-wrong than the human copywriter it replaced, because the human copywriter had a second signal (life experience, cultural fluency, the intuition of "this word means something different in this neighborhood") that the AI does not.
You cannot replicate human cultural fluency at scale by making the AI smarter. The AI is already near-peak on its single-signal performance. You replicate human cultural fluency at scale by pairing the AI with a second signal — one whose provenance is audience behavior, not preference-label training — and by building workflows that surface the disagreements between the two signals as the most valuable output of the system.
Thrift stores, vinyl records, SNAP benefits. Every loaded phrase in your next campaign has one dictionary reading and one behavioral reading. The campaigns that ship both are the campaigns that convert at the ceiling. The campaigns that ship only the dictionary reading — the ones every single-signal AI produces — are the campaigns that stall.
Which kind is yours?