Advertising Robot Impression Measurement
At a glance: The advertising robot business case turns on a number nobody can audit, which is why so many pilots stall at the second meeting. A brand will not pay for impressions it cannot verify, and the difference between a defensible impression count and a vanity figure is a written counting rule, a timestamped log the retailer owns, and a reconciliation against the store's own footfall. This guide sets out the counting rules, the audit procedure and the contract terms that turn a robot's screen time into sellable retail media inventory.

Why Impression Measurement Decides Whether the Business Case Exists
An advertising robot in a supermarket has three plausible revenue lines: a per-impression or per-campaign fee from brands, a service fee from the retailer for the platform, and the residual value of the robot as a delivery or guidance asset when it is not advertising. The first line is the one that scales, and it is the one that fails first, because it requires a number that a brand's media agency will accept against its own measurement standards.
Those standards are not generous. A media agency buying digital out-of-home inventory expects a currency-grade measure: a defined audience, a defined exposure event, a stated counting rule, third-party validation or at minimum an auditable log, and a reconciliation against an independent footfall source. A pilot that reports "our robot generated 180,000 impressions last month" without any of that is a claim, not inventory, and the agency's planner cannot put it in a plan.
The consequence is a specific failure pattern that repeats across retail robot pilots: the technology works, the screens are bright, shoppers do stop, and the campaign does not renew. The reason is almost never the hardware. It is that the impression figure could not be defended in a post-campaign review, so the second purchase never got signed.
This guide fixes that layer. The commercial structure of a retail media programme is covered in the retail media business case for supermarket robots, and the payback arithmetic in the buy versus lease payback model; both assume an impression figure, and this article is about making that figure stand up.

Four Counting Rules, and Why Only One Is Sellable
Every impression claim rests on a counting rule, and there are four in common use. They produce numbers that differ by an order of magnitude, which is why the rule has to be written down before the first campaign rather than after the first invoice.
| Rule | What is counted | Typical multiplier | Sellable? |
|---|---|---|---|
| Screen-on time | Minutes the display was lit | Highest, no audience link | No |
| Pass-by count | Everyone detected within a fixed radius | High, includes backs turned | Rarely |
| Opportunity to see | Persons facing the screen within a viewing cone for a minimum duration | Medium | Yes, as gross |
| Verified impression | Opportunity-to-see events deduplicated per person per campaign window, logged and auditable | Lowest | Yes, as the traded unit |
Screen-on time is a proxy for nothing except electricity consumption. Pass-by counting inflates badly in a supermarket aisle where the same shopper passes four times in a single trip. Opportunity to see is the standard digital out-of-home unit and is defensible if the viewing cone and the minimum dwell are stated. Verified impressions add deduplication and a log, which is what an agency needs to reconcile a claim against footfall.
The right commercial structure is usually to sell on verified impressions and to report gross opportunity-to-see alongside, so the agency sees both the currency figure and the reach story. Quoting only the opportunity-to-see number invites the challenge that the audience was double counted, and losing that argument once costs the contract.

What the Sensor Can and Cannot Tell You
A robot carrying a screen already has the sensors to count people: the navigation lidar, the depth cameras used for obstacle avoidance, and often a forward RGB camera. The temptation is to treat the obstacle-detection pipeline as an audience measurement system. It is not, and understanding why keeps the claim honest.
Lidar detects legs and bodies as geometry, which is excellent for counting presences in a zone and poor for direction of gaze. Depth cameras add pose estimation, enough to distinguish a person facing the robot from a person walking past at an angle, but not enough to prove attention to the screen. An RGB camera with a gaze model can estimate attention, but its accuracy varies with lighting, with face coverings, and with the demographic mix, and it raises a data protection question that must be settled before the first deployment rather than after.
The practical consequence is that the honest measurement is a two-stage estimate: a geometric count of persons within a stated viewing cone and distance, multiplied by a measured attention rate for that placement, where the attention rate comes from a sampled observation study rather than from a model claim. A sample of 300 observed passers, observed by a person with a counter for two hours at the same time of day, produces an attention rate with a defensible confidence interval at negligible cost. Documenting that method is what makes the number credible; a vendor quoting a model accuracy figure with no sample behind it is quoting a hope.
| Quantity | Source | How to state it |
|---|---|---|
| Persons in viewing cone | Lidar and depth fusion, radius and angle stated | Count per 15-minute bin, per location |
| Dwell time | Time within the cone, per person track | Median and 75th percentile, plus share above 2 seconds |
| Attention rate | Sampled human observation, minimum 300 passes | Percentage with confidence interval and observation protocol |
| Verified impressions | Dwell-qualified tracks, deduplicated per window | Single currency figure per campaign |

Dwell Time Is the Real Product
Impressions get the invoice signed, but dwell time is what the brand is actually buying, because a two-second glance at a moving robot is worth less than a twelve-second stop where the shopper reads the offer. Reporting both is what separates a serious retail media offer from a screens-on-wheels pitch.
Dwell distributions are heavily skewed and the mean misleads. A placement where the median dwell is 1.8 seconds and 12 percent of passes exceed 6 seconds is a different product from one where the median is 2.6 seconds and only 4 percent exceed 6 seconds, even if the average looks similar. The useful report gives the median, the share above a 2-second and a 6-second threshold, and the count of verified impressions at each threshold. Brands buying a call to action care about the long tail; brands buying awareness care about the total.
Placement drives dwell more than content does. A robot stationed at the end of an aisle where shoppers slow down to turn produces materially longer dwell than one in a circulation route where foot traffic is continuous and fast. That is an operational finding with a commercial implication: the retailer can price placements differently, and the robot operator can demonstrate why the premium position costs more. Reporting dwell by placement, rather than a store-wide average, is what makes that segmentation possible. The peak hour scheduling guide covers the timing half of the same problem, and the wayfinding deployment guide covers how a robot's guidance function changes its dwell profile.

The Audit Procedure a Media Buyer Will Run
An agency auditing a retail media claim follows a fixed sequence, and preparing for it in advance is far cheaper than surviving it. The sequence is: does the log exist at the required granularity, does the counting rule match what was sold, does the total reconcile against independent footfall, and are there anomalies that suggest gaming.
Granularity first. The log must carry a timestamp per event to the second, the location or zone, the dwell duration, and a session or track identifier that lets duplicate events be spotted. Aggregated hourly totals in a spreadsheet do not survive this step, because they cannot demonstrate that 40,000 impressions were 40,000 distinct qualifying events rather than 8,000 events counted five ways.
Reconciliation second. The store has at least one independent footfall measure, whether that is a door counter, a till transaction count, or a WiFi or camera-based people counter. The robot's person-count in its own zone should sit in a plausible relationship to the store's total entry count: a robot near the entrance might see 30 to 60 percent of entries, a robot deep in an aisle far less. A robot claiming impressions exceeding the store's footfall is the fastest way to lose an audit, and it usually indicates the counting rule counted the same person repeatedly.
| Audit step | What is checked | Failure signal |
|---|---|---|
| Log granularity | Event-level rows with timestamps and track IDs | Only hourly aggregates available |
| Counting rule match | Rule stated in the insertion order equals rule implemented | Rule was never written down |
| Footfall reconciliation | Zone count plausible against store totals | Impressions exceed footfall |
| Anomaly review | Distribution of dwell, diurnal shape, zero-traffic gaps | Flat 24-hour profile, impossible dwell values |
| Change control | Rule or model version changes are logged | Numbers shift with no recorded change |
Anomaly review is where a serious audit ends. Real data has a diurnal shape, a quiet hour, and days where the store was shut. A campaign report with a perfectly flat hourly profile, or with traffic during a closed period, tells the auditor the data pipeline is synthesising rather than measuring. Preparing a report that shows the shape of a real trading day, including the boring gaps, is more persuasive than any accuracy claim.
Brand Safety, Privacy and the Rules That Cannot Be Broken
Retail media inventory carries obligations that a normal screen deployment does not. The three that matter most for an advertising robot are audience measurement privacy, content adjacency, and the prohibition on profiling.
On privacy: counting persons in a zone is generally treated differently from identifying them, and the design decision that keeps the two apart is to process detection on the device and store only counts and durations. Retaining images of shoppers for impression verification is attractive forensically and carries a data protection burden that varies by jurisdiction, and in some markets is not permissible at all without a specific lawful basis. The safe architecture stores the count and the dwell, plus a low-resolution thumbnail only where the site's own policy permits, and states the retention period explicitly in the media contract.
On content adjacency: a robot moving through a store cannot control what is beside it. A brand that has bought exclusivity in a category will object to its campaign running beside a competitor's gondola end. The practical control is placement zoning, where the robot's route is restricted to agreed positions for a given campaign, and the log records the zone per impression so adjacency is demonstrable rather than assumed. That single field, zone per event, is also what lets the retailer sell adjacent inventory separately.
On profiling: inferring demographics or repeat visits from a robot's sensors is where retail media programmes get into difficulty. The commercial temptation is real, because a brand will pay more for a known audience. The defensible position is aggregate measurement only, with no individual-level record retained, and the media contract should say so plainly, because the agency will ask and an evasive answer ends the conversation.
Contract Terms That Make Impressions Tradable
Turning an impression figure into inventory requires four terms in the media agreement, all of which are cheap to write and expensive to omit. They are the points at which an unsellable pilot becomes a renewing campaign.
| Term | What it specifies |
|---|---|
| Traded unit | Verified impressions, with the counting rule, viewing cone, minimum dwell and deduplication window written into the schedule |
| Reporting deliverable | Event-level log delivered to the brand or agency, not a summary, with the fields in this article |
| Reconciliation | Monthly comparison against the store footfall source, with a stated tolerance and a remedy if the log is unavailable |
| Make-good and audit rights | Shortfall credited on verified impressions; the buyer may audit the log on reasonable notice |
Audit rights are the term that does the most work, because their existence changes how the operator builds the log in the first place. An operator who knows the buyer may inspect event-level data designs for inspectability from day one, and that discipline is what makes the second campaign easy to sell.
The rest of the programme is ordinary operations. A robot that advertises also needs to be charged and scheduled, which the dock scheduling guide covers, and its route and placement need to be planned against shopper flow rather than left to default paths. The electronics retail advertising analysis shows how the measurement changes in a store with fewer, longer-dwell visits, and the payback model shows what happens to the business case once the impression figure is real rather than optimistic. That last comparison is usually sobering and always useful: it is the difference between a fleet that gets funded on a defensible number and one that gets funded on a number that evaporates at the first audit.
