AI in DOOH: signal vs noise

“AI-powered” is on every DOOH vendor deck, which makes it hard to tell what’s genuinely deployed from what’s a relabelled spreadsheet. The honest summary: the technology is real and operating at scale, but the impact numbers attached to “AI” are almost entirely self-reported by the people selling it. This analysis separates the signal from the noise — what AI actually does in DOOH today, what’s mislabeled, and what’s still a showcase.

What’s genuinely real

Strip away the labelling and several AI/ML capabilities are genuinely operating in DOOH at scale:

Anonymous video analytics (computer vision). Sensors detect presence, dwell and inferred attributes to feed audience impression estimates — and this is operational, not theoretical: one vendor’s CV data is accepted inside a major programmatic exchange, the strongest proof it’s deployed (Quividi, AdMobilize — primary for capability). (The privacy mechanics are in Privacy & in-venue measurement.)
Attention modelling as a planning layer. Attention-prediction products (e.g. an “attention unit” score) are real; one had its methodology externally reviewed and found reproducible — though the review validated the method, not business outcomes, and published no lift numbers (Adelaide / MediaSense review — primary).

These are the signal: real technology, doing a real job, today.

What’s mislabeled

A lot of “AI in DOOH” is real functionality wearing an AI label it doesn’t earn:

“AI yield optimization.” Programmatic mediation that picks the winning bid is mostly auction and price logic, not machine learning. It’s genuine and valuable — it’s just not “AI” in the sense the deck implies.
“AI-powered triggers.” Weather-, traffic- or time-triggered creative (a temperature-responsive campaign, say) is rule-based automation — “if humidity > X, show anti-frizz” — not an AI model. Real and useful; not intelligence.
Demographic/mood inference. This is ML, but it’s the soft kind: peer-reviewed work finds automated age/gender inference is systematically biased — less accurate for women and some ethnic groups, degraded by makeup, lighting and angle (academic — primary). In a beauty venue, where makeup and dramatic lighting are the norm, that error is structural. So footfall counting is trustworthy; the demographic read on top is a biased estimate.

Calling auction logic and if-then rules “AI” isn’t fraud — it’s marketing inflation. But it makes the category hard to evaluate, and it’s worth deflating when you’re assessing a vendor.

The lift numbers don’t survive scrutiny

Here’s the firmest finding: no trustworthy, independent number for AI-attributable lift in DOOH exists. Every circulating figure is vendor- or agency-self-reported and attribution-broken:

Omnichannel campaign results (a sales lift across Meta, TikTok, YouTube and DOOH) can’t isolate DOOH, let alone the “AI” within it.
Award-case figures (“+60% visits,” “+8% sales”) are self-reported by the party that won the award.

This doesn’t mean AI does nothing — it means the quantified claims are unverified. The honest statement for any plan: treat any ”% uplift from AI” as marketing until it’s independently measured, the same discipline as the no-fabricated-benchmark rule everywhere else.

Generative AI creative: real but showcase

Generative AI for DOOH creative is genuinely emerging, but mostly as flagship demonstrations rather than everyday infrastructure. The best-documented case generated thousands of hyperlocal headlines on moving LED trucks using a large language model — a real, award-recognised deployment (PODS × Google Gemini — primary that it ran). What that proves: generative copy at scale is demonstrated. What remains unproven publicly: generative imagery, live at scale, with attributable revenue. So gen-AI in DOOH is real signal at the showcase end and noise when pitched as standard production capability today.

Signal vs noise — the scorecard

Claim	Read
Anonymous footfall/impression counting	Signal — deployed at scale
Programmatic automation	Signal — real (but it’s auctions, not “AI”)
Rule-based data triggers	Signal — real (but rules, not ML)
Attention modelling (planning layer)	Signal — real method; outcomes unproven
Demographic/mood inference	Grey — real ML, systematically biased
”AI yield optimization”	Noise — mostly auction logic relabeled
”AI drove +X%” lift figures	Noise — self-reported, attribution-broken
Gen-AI creative as everyday infra	Noise — still showcase, not standard

The takeaway

AI in DOOH is neither vapour nor magic. The technology is real and working — CV counting, automation, attention modelling — and worth using. But three disciplines keep you honest: deflate the label (much “AI” is auctions and rules), distrust the demographic read (biased, especially in beauty), and reject the lift numbers until someone independent measures them. Cite the mechanisms confidently; flag every ”% from AI” as vendor-reported. That’s the same posture this field demands everywhere — the technology is ahead of the evidence, and the credible operator says so.

Methodology & caveats

Verified (primary): anonymous computer-vision analytics being operational (and accepted into a major programmatic exchange) — Quividi, AdMobilize; the systematic bias of automated age/gender inference from peer-reviewed computer-vision research; an attention-prediction method receiving an external (vendor-commissioned) methodology review that found it reproducible — Adelaide / MediaSense (validated method, not outcomes; no lift numbers published); a generative-AI DOOH deployment generating thousands of headlines via an LLM — PODS × Google Gemini (that it ran; performance figures directional). Directional / not asserted: every “AI-attributable lift” figure (omnichannel or award self-reports — attribution-broken; no independent AI-lift number for DOOH exists); vendor CV accuracy claims (90–94%) as real-world; “AI” in mediation/yield (it’s auction logic); “first to use generative AI” superlatives. Programmatic automation and trigger systems are real but largely rule/auction-based, not ML — the “AI” label is often marketing.