Count-on-download is a method of counting advertising impressions that registers one when the creative has begun arriving on the user's device, rather than when the ad server decided to send it. The trigger sits on the client: something in the browser or the application fires an HTTP request back to a measurement endpoint, and that request, not the server's own delivery log, becomes the impression of record.

The method exists because those two moments can be far apart. An ad server can select a creative, write a response and log a delivery for a page the reader has already closed, for a slot below a fold that never scrolled into existence, or for a connection that dropped mid-transfer. Counting on download removed the most obvious source of inflation without demanding proof that anything appeared on a screen. That halfway position defined display measurement until the standards bodies moved the line again.

How the count is triggered

The mechanics come from the Interactive Advertising Bureau and Media Rating Council guidelines published in September 2004 as version 6.0b. The label is industry shorthand; the documents themselves speak of client-initiated counting. An impression may be counted only when an ad counter receives and answers an HTTP request for a tracking asset sent by a client, and the count has to happen after retrieval of the underlying page content has started. The document defines the metric as one "recorded at a point as late as possible" in the delivery of creative material to the browser.

Three implementation techniques are named: requests generated by an image tag, an inline frame, or a script source. The counter's response can be a beacon, most often a 1x1 pixel image though the guidelines set no restriction on content type, a 302 redirect, or the ad content itself. One asset may register several advertisements at once, a practice called compound tracking, provided the asset is placed so that every grouped ad loads before it fires. Cache-busting is mandatory: expiry, cache-control and pragma headers, plus a random string or timestamp appended to each call so that a repeat serve from local cache still produces a countable event.

Google Ad Manager shows the method inside a live ad server. Its Data Transfer export splits the transaction into three files. NetworkRequests logs every request the server received. NetworkCodeServes logs every response it sent. NetworkImpressions logs the ones that came back downloaded, matched to the code serve by a shared ImpressionId string. A record appearing in the first two files but not the third is a served impression that never arrived. Ad Manager's documentation works through an example in which 95 per cent of served impressions are also downloaded impressions, giving an impression rate multiplier of 0.95.

What the industry counted before

The alternative was server-initiated counting, known as count on decision or count on insertion, in which the impression is logged at the same moment the surrounding page content is served. The 2004 guidelines ruled it out, on the reasoning that it sits furthest from the user actually seeing the advertisement.

The evidence behind that ruling came from a two-phase project the IAB began in May 2001, facilitated by the MRC with the Advertising Research Foundation and audited by PricewaterhouseCoopers. Ten companies took part, among them AOL, DoubleClick, MSN and Yahoo, together accounting for close to two thirds of industry advertising revenue. The second phase, run through 2003 and 2004, compared the numeric difference between client-initiated and server-initiated counts directly, and produced the rule.

Video moved ahead of display: the digital video impression guidelines revised in December 2009 required counting after buffering, once the advertisement had started to play, the approach that still underpins quartile reporting. Mobile followed in July 2013, when updated mobile web and in-application guidelines prohibited server-initiated measurement outright.

Why the choice of moment became contentious

Buyer and seller almost never count the same impressions, because their systems count at different instants in the same delivery. The commercial machinery for that gap predates the argument: the 4A's and IAB standard terms and conditions finalised in February 2010 nominate a controlling measurement and set a 10 per cent reconciliation threshold for invoice disputes. Where one side counts on download and the other counts on serve, discrepancies land inside that threshold on good inventory and well outside it on slow-loading pages.

Publishers spent 2016 reporting exactly that. Joseph Pamboris, then chief product officer at Flashtalking, told AdMonsters that the discrepancy was starkest on small screens, describing it as disconcerting for a marketer to see that "more than 20% of mobile display impressions didn't even get to the count-on-download stage". A publisher billing on served impressions sells delivery it cannot evidence; one billing on downloaded impressions absorbs the cost of slow-loading advertiser creative.

The standard that replaced it

The move away from count-on-download began at the AdMonsters Publisher Forum in Austin in August 2016, where operations staff raised rising discrepancies between demand-side and sell-side servers. DoubleClick for Publishers announced downloaded impression metrics that November, the first stage of a three-part plan.

The IAB, MRC and Mobile Marketing Association then proposed something stricter. Their revision moved desktop display, mobile web and in-app measurement to count-on-begin-to-render, which requires the creative to be "loaded and at minimum begin to render" before an impression counts. MRC analysis found little difference between counting at the start of rendering and counting at its completion, but a significant difference between rendering and downloading. The choice of the start of rendering rather than its end was deliberate: no party in the chain reliably observes an end-state event, and monitoring one would mean JavaScript polling that consumes processor time and battery. Public comment closed on 1 February 2017 and the final guidelines were issued that October, with counts recorded before rendering begins relabelled as tracked ads rather than impressions.

Google Ad Manager switched to downloaded impressions on 2 October 2017. AdSense followed in 2018, when Google warned publishers to expect impression counts to fall and impression-based rates such as revenue per mille and click-through rate to rise, with earnings unaffected by the change. Sources disagree on when the advertiser-side products moved: PPC Land reported at the time that impression counting was updated across DoubleClick products in January 2018, while Google's own publisher documentation dates the Ad Manager transition to October 2017.

Where the definitions come apart

Count-on-download has no single technical definition, and the platforms that used it describe it differently. Google's help centre now characterises the old standard as counting an impression when "the ad code payload started executing in the browser or app", and its replacement as recording one only once the creative assets have finished downloading. That reading is stricter than the MRC's, which asks for the beginning of rendering and explicitly does not require any part of the advertisement to be visible. Two platforms can claim begin-to-render compliance and count different events.

Other failure modes are structural. Pre-fetch and pre-render requests pull creative that a reader may never navigate to, and the guidelines exclude them, but detection depends on publishers declaring the behaviour. Cache serves produce no fresh request unless busting is implemented correctly. Latency inflates the loss rate on slow connections, which concentrates the shortfall in exactly the markets where inventory is cheapest.

Adjacent terms

A served impression is counted when the ad server dispatches a response, with no client confirmation, and is the metric count-on-download was designed to displace. Begin-to-render is one step later than download, requiring the creative to start painting. A viewable impression is later still, requiring 50 per cent of pixels in an in-focus tab for one continuous second for display, two seconds for video. A podcast download is a different metric entirely despite the shared word: the IAB Tech Lab podcast measurement guidelines at version 2.2, published in February 2024, count one when roughly a minute of file bytes has been served and deduplicate repeat requests inside a 24-hour window.

Recent developments

Google began retiring the method in its buying stack in 2025. An August announcement set September 2025 for the shift from on-download to begin-to-render in Display and Video 360 display reporting. Campaign Manager 360 made the same change on 6 November 2025. Reported impression counts fall under the stricter rule, a decrease reflecting measurement rather than delivery.

Billing lags reporting. As of an August 2026 update to Google's help documentation, Campaign Manager 360 and Display and Video 360 will move billed display impression counting from on-download to begin-to-render in February 2027, citing MRC accreditation and warning of a slight fall in billed volume.

Changes to three MRC-accredited metrics on 14 July 2025 were expected to roughly double reported invalid and general invalid traffic begin-to-render impressions, a detection improvement that reads as deterioration on a dashboard. Accreditation has settled on the newer basis: Amazon's demand-side platform holds MRC accreditation for display begin-to-render impressions and click-throughs, and Integral Ad Science was accredited for third-party measurement on that platform on 13 November 2025.

The direction of travel runs past rendering altogether. Display and Video 360 moved to counting impressions on Open Measurement inventory when a single pixel of creative is on screen, introducing a provisional impressions metric so the effect could be compared against the standard count. Microsoft's Xandr documentation offers publishers a one-pixel-in-view option alongside begin-to-render in mobile applications, noting that it is not an industry standard and creates discrepancies with third-party counts. Each step moves the impression closer to the eye and further from the server, and each one lowers the number.

Timeline

  • May 2001: The IAB begins a two-phase measurement project with the MRC, ARF and PricewaterhouseCoopers, comparing client-initiated and server-initiated impression counts
  • September 2004: Version 6.0b of the Interactive Audience Measurement and Advertising Campaign Reporting and Audit Guidelines requires client-initiated counting and rules out count on decision
  • November 2004: The IAB Measurement Task Force issues the global standard for counting online advertising impressions
  • December 2009: Revised digital video guidelines require counting after buffering, once playback has begun
  • July 2013: Mobile web and in-application measurement guidelines prohibit server-initiated counting
  • June 2014: The MRC releases viewable ad impression guidelines, setting 50 per cent of pixels for one second on display and two seconds on video
  • August 2016: Publishers raise download versus render discrepancies at the AdMonsters Publisher Forum in Austin
  • November 2016: DoubleClick for Publishers introduces downloaded impression metrics
  • 1 February 2017: Public comment closes on the IAB, MRC and MMA proposal to move display counting to begin-to-render
  • 2 October 2017: Google Ad Manager transitions to downloaded impressions
  • October 2017: Final desktop display, mobile web and in-app guidelines formalise begin-to-render and reclassify earlier counts as tracked ads
  • 2018: AdSense switches from served to downloaded impressions
  • May 2024: Amazon's demand-side platform receives MRC accreditation for display begin-to-render impressions
  • 14 July 2025: Changes to three MRC-accredited metrics in Display and Video 360 roughly double reported invalid begin-to-render impressions
  • September 2025: Display and Video 360 display reporting moves from on-download to begin-to-render
  • 6 November 2025: Campaign Manager 360 adopts begin-to-render impression counting
  • February 2027: Billed display impression counting in Campaign Manager 360 and Display and Video 360 is scheduled to move to begin-to-render

Summary

Who. Ad servers and measurement vendors implement the counting method; the IAB, the MRC and the Mobile Marketing Association define it; publishers, agencies and demand-side platforms bill and reconcile against it.

What. Count-on-download records an advertising impression when the creative begins arriving on the user's device, triggered by a client-side request to a tracking asset rather than by the ad server's own delivery log.

When. Formalised by the September 2004 IAB and MRC guidelines, adopted across desktop display through the following decade, superseded by begin-to-render in the October 2017 guideline revision, and retired from Google's buying platforms between September 2025 and a scheduled February 2027 billing change.

Where. Desktop and mobile web display, mobile applications and the ad server reporting built on them, with video and audio operating on separate counting rules.

Why. The moment an impression is counted decides what an advertiser pays for and what a publisher can invoice, so each move down the chain from server to screen lowers reported volume, shifts derived rates, and reopens the discrepancy between buyer and seller counts.