What does "AI-powered" actually mean when a programmatic DOOH platform says it, and how do you tell whether it helps your campaign? In practice it usually means machine learning models that forecast audiences, set bids and pacing, or shift budget toward better-performing screens, and the way to judge it is to ask what data the model learns from, what it is optimising for, and how its results compare with a baseline that did not use it.
This guide sets out where machine learning fits in the pDOOH chain, what it needs to work, the questions that separate real optimisation from marketing language, and why campaign size limits what any algorithm can learn.
Where machine learning fits in pDOOH
| Task | What a model does | What it learns from | What to check |
|---|---|---|---|
| Audience forecasting | Estimates how many people, and what kind, will be near a screen at a given time | Footfall counts, mobile location panels, sensors, past audience data | Data source, update frequency, how estimates are validated |
| Bidding and pacing | Decides which plays to bid on and at what price to hit goals within budget | Past auctions, prices, delivery and your campaign goal | Which goal it optimises: impressions, cost, or an outcome |
| Budget reallocation | Moves spend between screens, hours and networks during a flight | Delivery and response data from the live campaign | Whether you can see and override its choices |
| Creative selection | Shows the version predicted to do best in each context | Response data by version, time and place | How much data each version needs before a decision |
| Anomaly detection | Flags screens or hours behaving unusually, such as sudden drops in plays | Delivery logs and proof of play | Whether alerts reach someone who acts on them |
None of these is new in online advertising. What is different in out-of-home is the data. A screen has no clicks and no logged-in users, so models depend on audience estimates and indirect response signals, which are noisier than online data. That is the main reason to be careful with bold claims.
Rules versus learning
Many "smart" features are rules you set yourself: run this creative when it rains, bid only between 7:00 and 10:00. Our guide to dynamic creative based on weather and time covers that approach, and it works well when you already know which conditions matter.
Machine learning is different in that the system finds patterns you did not specify, such as which screens or hours deliver the most response for your brand, and adjusts automatically. Both have a place. Rules are transparent and predictable. Learning can find gains you would miss, but only if it has enough data and a clear goal. Some products marketed as AI are rules underneath, which is not a problem in itself, as long as you know which you are buying.
Ask for the baseline
The most useful single question is: compared with what? The Trade Desk, for example, describes in a May 2023 guide to optimising DOOH campaigns an audience reach measure that it reports against a baseline: the score the campaign would have delivered without the platform's decisioning, by targeting all screens in the deal. Whatever platform you use, ask for that kind of comparison. An optimised result with no baseline tells you little.
Illustrative example: is there enough data to learn?
Illustrative example. The response rates below are hypothetical, chosen to show the arithmetic.
Suppose a campaign tests two creatives, each with a QR code, and the true scan rates are 0.10% and 0.12% of impressions. A standard sample-size calculation for comparing two rates, at 95% confidence and 80% power, gives:
n per version ≈ (1.96 + 0.84)² × [p₁(1 − p₁) + p₂(1 − p₂)] ÷ (p₁ − p₂)²
- With p₁ = 0.0010 and p₂ = 0.0012: (2.80)² × (0.000999 + 0.001199) ÷ (0.0002)² ≈ 431,000 impressions per version.
- For two versions, that is about 862,000 impressions before the difference can be detected reliably.
- A campaign of 300,000 impressions split across both versions would produce roughly 150 to 180 scans per version, far too few to tell them apart.
No algorithm can get around this. If a platform claims to have found a winning creative on a small campaign, the "winner" may be noise. Bigger differences need less data, and outcome signals such as store visits or sales may be more plentiful than scans, but the principle holds: ask how much data a decision was based on.
Questions to ask a platform about its AI
- Which tasks does the model handle: forecasting, bidding, pacing, budget shifts or creative choice?
- What goal does it optimise for, and can you choose it?
- What data does it learn from, and does it include data from other advertisers' campaigns?
- How are audience estimates produced and checked against real counts?
- Can you see why it moved budget, and can you override it?
- What is the baseline it reports against?
- How much data does it need before changing course?
- If cameras or sensors are involved, what is captured, what is stored, and how does it comply with privacy law where the screens are?
Privacy and audience data
Audience estimates in pDOOH come from aggregated sources such as mobile location panels, footfall counters and, on some screens, cameras that count people passing. Ask how anonymous each source is, what is retained, and what consent applies. Location data from phones has also become harder to collect: Apple's App Tracking Transparency rules require apps on iOS 14.5 and later to get permission before tracking users or accessing the advertising identifier, which shrinks the data some models rely on. Our guide to how location intelligence platforms measure footfall explains how these panels are built.
Consistent data also matters. Models compare campaigns, so labels for placements, formats and media types need to mean the same thing everywhere. The IAB's proposed Campaign Data Standards 1.0, released for public comment in May 2026, aim to standardise those classifications across the industry.
What this means for brands and media owners
- Brands should treat AI features as tools to evaluate, not reasons to buy. Set a clear goal, keep a comparison group, and judge results against a baseline. Our guide to measuring pDOOH success covers the setup.
- Media owners benefit most from accurate audience forecasting and pricing, which depend on good underlying data. Moving Walls' article on growing in-mall revenue with programmatic DOOH makes the case that standardised inventory and trusted reporting come first.
- Everyone should keep humans in charge of the goal, the budget and the creative idea, and let models handle the repetitive decisions within those limits.
For the basics behind all of this, see Moving Walls' explainer on programmatic DOOH and our own guide to how programmatic DOOH buying works. Planning for a specific audience, such as reaching Gen Z and Millennials, still starts with choosing the right moments, whatever the platform automates afterwards. The programmatic DOOH glossary defines the terms used here.