What programmatic can and cannot measure
Four systems see a programmatic campaign, and each sees a different part of it. The DSP records the bid requests it received, the auctions it won, the impressions it served and the price it paid, and in a Google stack it also reports, through Active View, whether each impression was measurable and viewable. The ad server, usually CM360 when the stack is Google's, counts the impressions and clicks that reached its own tags across every buy, direct or programmatic, and records conversions through Floodlight. Verification vendors measure viewability, brand safety and invalid traffic on the impressions they were able to observe. Analytics records what happened on the site after a click or a view: sessions, behaviour and outcomes, under its own attribution rules and usually without any sight of the impression.
Between them they answer most delivery questions well: how many impressions ran, on which sites and apps, at what cost per thousand (CPMCost per 1,000 impressions.), how many were in view, how many people were reached and how often each saw the ad. What they cannot do, individually or together, is say whether anyone noticed the ad, whether a conversion would have happened without it, or who a person was once they moved between a browser, an app and a walled garden. Those questions need attention measurement, incrementality tests and modelling respectively, and each has its own treatment in this section. A measurement plan written before launch, naming the question each number answers and the system that owns it, prevents most of the arguments that follow a flight.
Delivery and viewability
Delivery is the first thing to check and the easiest to misread. A served impression is not a seen one: the ad may have rendered below the fold, in a background tab, inside a hidden frame or to a bot. The Media Rating Council's viewable impression guidelines define the minimum conditions under which an impression counts as having had an opportunity to be seen, in terms of the share of the ad's pixels inside the viewable area and the continuous time they remain there, with a longer requirement for video. Google's Active View applies that standard inside DV360 and CM360 reporting, and third-party vendors apply it independently with their own measurement footprint.
The reports then split impressions three ways: measurable, where the script could decide; viewable, where it decided yes; and unmeasurable, where it could not run or could not see. Viewable rate is normally calculated on measurable impressions, which means two campaigns with the same headline rate can carry very different shares of impressions nobody could assess. The viewability and attention guide explains the definition, the three buckets, time-in-view, the attention vendors that go beyond the binary standard and how far any of it predicts an outcome. The controls that act on these numbers before the bid, such as viewability targeting, verification segments and inventory exclusions, belong to the brand safety and viewability guide and are not repeated here.
Reach and frequency
Impressions count exposures. ReachUnique people who saw an ad. counts people, and FrequencyAverage times a person saw an ad. is how many exposures each person received on average. Every awareness plan is written in those terms, and every programmatic report struggles to deliver them, because counting people means recognising the same person across page views, devices and platforms, and the identifiers that once did that job are disappearing unevenly. DV360 and CM360 report unique reach using a mix of signed-in data, cookies where they still exist and statistical modelling; the help centres document which dimensions and date ranges those metrics support and why they cannot be summed across rows or periods.
Frequency is controlled through caps set at several levels: the DSP campaign, insertion order and line item, the advertiser's ad server, the publisher's own server and each walled garden separately. A cap is only as good as the identifier it counts against, so a person who is anonymous to the system is a new person every time. The cross-platform reach and frequency guide covers identity, what unique reach reports assume, capping at each level, overlap with walled gardens and how to plan a cap for a campaign. The reach and frequency calculator turns impressions and reach into frequency, and back again, for the planning stage.
Conversions, attribution and incrementality
In a Google stack, programmatic conversions are counted in the ad server rather than the DSP. Floodlight activities in CM360 fire on the pages or events that matter and are joined to the impressions and clicks the ad server has already logged, which allows a conversion to be credited to an exposure the person never clicked. The Floodlight tags guide covers activity design, counting methods and deployment; this hub only needs the consequence, which is that programmatic reporting rests on view-through as much as click-through, and the two must always be read apart.
AttributionAssigning conversion credit to touchpoints. then decides which exposure receives credit when several precede the conversion. The ad server applies its own model and lookback windows, the DSP applies another, and analytics applies a third and typically cannot see impressions at all. None of these models observes what would have happened without the ad. That is a question of IncrementalityConversions caused by marketing., which only a controlled test answers: a holdout audience, a geographic split or a platform lift study, as the incrementality testing guide sets out. For display and video, where view-through credit can be generous, the gap between attributed and incremental conversions is often the most important number a programmatic team does not know. The attribution-models framework and the measurement plan should agree in advance which model is used for which decision, so that the number used to optimise line items is never confused with the number used to set the budget.
DSP numbers vs analytics numbers
The DSP, the ad server and analytics report different totals for the same campaign, and most of the difference comes down to five things: the unit counted, which is impressions and clicks on one side and sessions and users on the other; the attribution window and model; what consent choices and ad blockers removed from each system; modelled data filling gaps in one system but not another; and the time zone and reporting lagThe time between an ad interaction and the conversion it leads to, which leaves the most recent reporting days incomplete. each applies. CM360 counts an impression when its tag is called, the publisher counts it earlier in the request, and analytics never counts it at all. A click the DSP logs becomes a session only if the landing page loads and the analytics tag fires before the person leaves.
The platform versus analytics discrepancies guide gives a general reconciliation method, and this section applies it: treat the ad server as the system of record for delivery, analytics for on-site behaviour and an experiment for causation; document the expected gap between each pair; and investigate only when a gap changes. A reconciliation note written once saves a weekly argument.
Three systems, three purchase counts
A furniture retailer runs a four-week programmatic display and online video campaign through DV360 with CM360 as the ad server. At the end of the flight the DSP reports 1,420 purchases, CM360 reports 1,180 and GA4 attributes 310 to the campaign. The team documents the gaps rather than arguing about them. CM360's count is lower than the DSP's because the DSP applies a longer view-through window and keeps conversions that the ad server has already credited to a search click. GA4's count is lower still because it only sees sessions that began with a click and applies a last-click model across every channel. A geographic holdout run over the same period measures roughly 420 incremental purchases, and that is the number the budget decision uses; the 1,180 is used to compare line items against each other, and the 310 is used to check that landing pages behaved.
Common measurement problems
Each symptom below is a measurement question before it is a buying one, and each has a home in one of the guides. Where Floodlight records nothing or counts twice, the viewable rate falls well short of plan or a cap does not hold, the programmatic troubleshooting section gives the symptom its own diagnosis, with a check, fix and prevention habit for every cause.
| Symptom | Likely cause | Where to look |
|---|---|---|
| Viewable rate looks healthy but a large share of impressions is unmeasurable | Rate calculated on measurable impressions only; inventory where the script cannot run | Viewability and attention |
| Unique reach changes when the date range changes | Reach is modelled and not additive; each range is estimated separately | Reach and frequency |
| A frequency cap is set but people report seeing the ad constantly | The cap counts per identifier, not per person; cross-device and walled-garden exposures sit outside it | Reach and frequency |
| DSP conversions far exceed ad server conversions | Different windows; view-through credit; no cross-channel deduplication in the DSP | Floodlight tags, discrepancies guide, Floodlight conversions counted twice |
| Analytics shows almost no conversions from display | Analytics cannot see impressions; last-click credits search and direct | Discrepancies guide, incrementality testing |
| Attributed conversions rise after launch but sales do not | View-through crediting purchases that would have happened anyway | Incrementality testing, Floodlight conversions counted twice |
| Active View and a third-party vendor disagree on viewability | Different measurement footprints and handling of unmeasurable impressions | Viewability and attention, DV360 reporting |
| Publisher and ad server impression counts differ | Counted at different points in the ad request | CM360 ad serving basics |
The programmatic guide places this section within the wider ecosystem of exchanges, DSPs and ad servers, and the author's media planning and buying expertise page describes how these measurements are used to plan a buy and judge it afterwards. Measurement is the part of programmatic that survives every change in targeting and identity: the questions stay the same even as the answers get harder to obtain.
