Why reach and frequency are hard across platforms
Impressions add up; people do not. If a campaign serves a person an ad on a news site, then in an app, then on YouTube and then in a social feed, the DSP has delivered four impressions, and whether its report shows one person at a frequency of four or four people at a frequency of one depends entirely on whether it could tell that the four exposures belonged to the same person. In practice it often cannot. The browser gives it one identifier, the app another, the video platform a third that it may or may not be allowed to join, and the social platform nothing at all because the buy happened elsewhere.
Three obstacles stack on top of each other. Identity is fragmented, so a person is many identifiers. Platforms are bounded, so each reports its own population and none sees the others. And everything is modelled, so each reach figure is an estimate built on assumptions that differ between vendors. The consequence is that ReachUnique people who saw an ad. summed across platforms is overstated and FrequencyAverage times a person saw an ad. is understated, often badly, and the person who saw the ad on every platform is the one the plan least wanted to overexpose. The measurement hub treats reach and frequency as the second delivery question after viewability; this guide is about why the answer is always an estimate and how to make it a useful one. Within the media planning framework, reach and frequency are planning figures first, and the measurement that follows should be read as a check on the plan rather than a restatement of it. Deduplicated reach matters most to advertisers planning for broad reach among category buyers, such as the supermarket brands described in the FMCG marketing guide. The same overlap question arises when linear and connected television join a plan, and the guide to planning a first television campaign shows how a first-time advertiser combines them with online video.
For the difference between the two counts in the first place, and why impression totals can be summed where reach cannot, see the reach versus impressions comparison. How reach and frequency accumulate in a booked plan, line by line and across offline and online media, is set out in the media planning page on how reach builds across a media schedule.
Identity: cookies, logged-in users, modelling
Four kinds of identity are available, and each covers a part of the picture. Third-party cookies were the historical basis for recognising a browser across sites, and they still work in some browsers and not in others; browser restrictions have changed more than once and vary by vendor, so the share of traffic without a usable cookie is something to measure in a campaign's own reports rather than assume. A cookie is per browser, not per person, so one person on a laptop and a phone is at least two cookies even where cookies work. Mobile apps use a device advertising identifier instead, subject to the consent framework of the operating system, and it is per device and reset at the person's choice.
Logged-in identity is the strongest kind and the most bounded. Google can recognise a signed-in person across the devices they are signed into on Google properties, and DV360Google's enterprise DSP. uses that signal, in combination with statistical modelling, to estimate unique reach across devices; the help centre documents that the metric blends signed-in data with modelled estimates rather than counting every person directly. Meta, Amazon and the other walled gardens have the same strength within their own walls and share none of it outside. Publisher first-party identifiers, hashed email-based identifiers and publisher-provided identifiers each cover their own slice of inventory, and a DSP that receives several of them on one bid request still has to decide which to trust.
Modelling fills the gaps. Where a signed-in population exists, its behaviour is used to estimate how many of the unrecognised impressions belong to people already counted; where it does not, assumptions about device sharing and cookie churn do the work. The direction of error matters. A missing identifier makes a person look new every time, so the systems that cannot recognise people overstate reach and understate frequency, and the people least recognisable to the system, those who decline consent or use restrictive browsers, are precisely the ones whose frequency nobody is managing.
Unique reach reports and what they assume
DV360Google's enterprise DSP. reports unique reach metrics, including total reach and average impression frequency, as estimates rather than counts. The help centre documents the conditions attached: the data arrives with a lag, the metrics are only available with certain dimensions, they cannot be summed across rows or date ranges, and small rows may be withheld or rounded under documented thresholds. All of those conditions follow from the method. A modelled estimate for a line item is not a subset of the estimate for its insertion order, because the two were estimated separately, and adding them double counts the people both reached.
CM360Ad server and measurement platform. has its own reach reporting, built on the identifiers the ad server sees across every buy that passes through its tags, which is the ad server's particular strength: it can see a direct buy and a programmatic buy side by side and report how many identifiers both touched. The CM360 reporting guide covers the report types and their documented restrictions on dimensions and lookback. The caveat is the identifier. An ad server reach report counts what its cookie or identifier could recognise, and the modelled cross-device estimate and the cookie-based count are different metrics that should not be presented as one.
Every unique reach report rests on assumptions that are worth writing down. It assumes the recognised population behaves like the unrecognised one. It assumes the cross-device join is right. It is specific to the date range, the dimensions and the platform that produced it. It includes walled-garden inventory only when that inventory was bought through the same platform, so YouTube bought through DV360 can be included in DV360's unique reach while a Meta campaign never can be. Used for relative comparisons between line items, formats or periods on the same platform, these reports are informative. Used as proof that the plan's reach was delivered to the person, they claim more than they can know.
Frequency capping at each level
A frequency cap limits how many times one identifier is served an ad in a window. Caps exist at several levels, each counts against a different identifier, and each is blind to everything outside its own system. The tightest applicable cap wins within one system; across systems, caps do not combine at all.
| Level | Where it is set | What it counts against | What it cannot see |
|---|---|---|---|
| Walled-garden campaign (YouTube, Meta, retail media) | The platform's own campaign or ad set settings | The platform's logged-in user | Anything bought outside that platform |
| DSP campaign (DV360) | Campaign-level settings, applying to every insertion order beneath | The identifiers DV360 can see, with signed-in signals where available | Other DSPs, direct buys, walled gardens bought elsewhere |
| DSP insertion order and line item | Each insertion order or line item | The same identifiers, scoped to that object | Sibling objects unless a campaign cap also applies |
| Advertiser ad server (CM360) | Ad-level frequency settings | The ad server's identifier across every buy using its tags | Impressions not served through its tags; identifier-less traffic |
| Publisher ad server (Ad Manager) | Line item or label frequency caps on the publisher's side | The publisher's identifier for its own properties | Every other publisher |
| Verification vendor | Not a cap; reports the frequency distribution it measured | Its own measurement footprint | Everything it did not measure |
The DSP settings themselves, including where in DV360 each cap lives and how caps interact with pacing, are covered in the bidding and pacing guide. The measurement point is simpler: a cap is a ceiling per identifier per window, and the frequency a person experiences is the sum of what every system served to every identifier they carry. A tight cap in one system does nothing about exposures from another, and a cap set where the identifier is weakest, such as in browsers without cookies, is a cap in name only. When people report seeing the ad far more often than the setting allows, the frequency cap troubleshooting page starts from this table and works out which system served the exposures the cap could not count.
Walled gardens and overlap
YouTube, Meta, Amazon, TikTok and the retail media networks share a design: logged-in identity inside, reach and frequency reported from that identity, and no user-level data released outside. Each report is credible on its own terms and useless for deduplication against anyone else's, because the join that would reveal overlap is exactly what the platform does not permit. Audiences overlap heavily in practice, since the same people use several platforms every day, so adding platform reach figures together produces an upper bound that is nearly always far above the truth.
There are five ways to do better, none of them free. Where a walled garden can be bought through the DSP, as YouTube can through DV360, the DSP's own cross-property reach estimate covers both. Clean rooms, such as the one Google offers for its own advertising data, allow an advertiser's data to be matched and analysed under restrictions without the user-level data leaving; they answer overlap questions for the platform that runs them, not across rivals. Measurement partners with panels observe the same people across platforms and model the overlap from the panel, which is the closest thing to an independent view and inherits every limitation of panel size. A modelled overlap, in which an assumed duplication between platforms is applied to the separate reach figures, is cheap and transparent and only as good as the assumption. And an incrementality test sidesteps reach entirely by measuring whether the combined plan changed an outcome. The Meta side of the problem, where frequency rises inside one platform as spend holds and the audience does not grow, is the subject of the frequency and creative fatigue guide.
Summed reach against deduplicated reach
A beverage brand runs a six-week campaign with DV360 display, YouTube bought through DV360, and a Meta campaign bought separately. DV360 reports 4,100,000 unique people across display and YouTube combined at an average frequency of 3.2; Meta reports 3,600,000 people at an average frequency of 4.5. The first draft of the wrap-up adds them to 7,700,000 reached. The planner instead applies a modelled overlap: audience research for the category suggests that around 55% of the Meta audience also uses YouTube or the news sites on the display plan, so roughly 1,980,000 of Meta's people were likely already counted. The deduplicated estimate is about 5,700,000, and the combined frequency for people reached on both sides is closer to 7 than to either platform's average, which is above the cap the creative was written for. The brand records the overlap assumption in the plan, lowers the Meta cap for the next flight, and commissions a panel-based measurement partner for the following campaign to replace the assumption with an observation.
Planning a frequency cap
A cap should follow from the campaign, not from a number inherited from the last one. Six inputs decide it. The objective: an awareness message needs enough exposures to be remembered and no more, while a response campaign is bounded by the point at which further exposures stop producing actions. The creative: a single message with one execution wears out sooner than a sequence with several, and a sequenced plan needs the cap to allow the sequence to complete. The purchase cycle: a product bought weekly tolerates a different rhythm from one bought once a year. The flight length, which decides whether the cap should be expressed per day, per week or per flight. The audience size against the budget: a small audience with a large budget will hit any cap quickly, after which the money either goes unspent or buys exposures above the cap through identifiers the cap cannot see. And identity coverage, because a cap counts identifiers, and a campaign running mostly in environments without them should expect real frequency to exceed the setting.
The practical sequence is to set a campaign-level cap as the ceiling, use line item caps to shape frequency by format where video and display play different roles, choose a window that matches the flight, and decide in advance what happens when the cap is reached, since the DSP will move spend to new identifiers and pacing will slow if there are none left. Then read the frequency distribution rather than the average: an average frequency near the cap can hide a long tail of identifiers far above it, and the share of impressions served above the intended frequency is the key performance indicator (KPIA metric tied directly to an objective.) that tells whether the cap is working. Once a channel's audience is saturated, the next unit of spend buys frequency rather than reach, which is where the marginal logic of the budget allocation guide takes over and money should move to a channel that still has people to reach. The planner below shows how impressions, reach and frequency constrain each other for a given audience, and the reach and frequency calculator does the same arithmetic for a single set of figures.
Reach and frequency planner
For how deduplicated reach sits alongside lift studies and branded search in a campaign's reporting, see the measurement plan for awareness campaigns.
For the console settings behind these ideas, the DV360 setup for brand awareness sets a campaign-level cap first, bids on viewable impressions and paces evenly so reach accumulates.
For a single platform's share of an awareness plan, the awareness campaign walk-through for the reach and frequency calculator shows where extra impressions stop finding new people.
Common mistakes
- Adding reach across platforms and presenting the sum as people reached.
- Adding reach across date ranges or rows from a modelled report that documents it cannot be summed.
- Treating a frequency cap as a promise to the person when it is a ceiling per identifier.
- Setting the cap in one system and forgetting the others, so the ad server, the DSP and the walled gardens each enforce a different ceiling on different identifiers.
- Reading average frequency instead of the distribution. The average hides the people seeing the ad far too often.
- Presenting cookie-based reach and modelled cross-device reach as the same metric.
- Ignoring consent and identifier coverage. The people the system cannot recognise are the ones whose frequency nobody is managing, and they are not a random sample.
- Inheriting last year's cap. The creative, the flight and the audience changed; the cap should have been re-derived from them.
