CPM is not a price. It is a division: what you spent, over what was counted, times a thousand. Almost everything written on the subject treats it as a rate you negotiate, when in most campaigns you never bid on it at all. You bid on a click or a conversion, and the CPM is the quotient you find in the report afterwards.

Two things explain why that quotient moves so much between platforms, and neither is the list of factors you usually get. The denominator is not the same thing everywhere, and most of the numerator never reaches the publisher.

Four different things are called an impression

The measurement standard does not define an impression. It defines a chain, and each rung holds fewer units than the one above it.

A served impression is a response from an ad server to a request. A rendered impression adds a condition: the ad must have loaded and at minimum begun to render. A measured impression is one whose viewable status could actually be determined. A viewable impression has crossed a threshold, and the threshold is written down:

“Pixel Requirement: Greater than or equal to 50% of the pixels in the advertisement were on an in-focus browser tab on the viewable space of the browser page, and Time Requirement: The time the pixel requirement is met was greater than or equal to one continuous second, post ad render.”

Thirty percent is enough for formats of 242,500 pixels or larger. Video requires two continuous seconds rather than one. Audio has its own standard built on audibility instead of pixels, with the same two-second floor and an explicit exclusion of ads served while the player is muted.

Two consequences follow, and together they account for a good share of the gaps you see.

The definition changed, and it changed late. The 2004 standard explicitly allowed counting at the moment the server decided to serve the ad, before anything appeared. The rule requiring an ad to begin rendering arrived in October 2017 for display and June 2018 for video. That is thirteen years during which an unknown share of billed impressions never reached a screen. The revision was explicit about rejecting the old approach:

“server-initiated ad counting methods (the configuration in which impressions are counted at the same time the underlying page content is served also known as count on decision or count on insertion) are not acceptable for counting rendered impressions because they are the furthest away from the user actually seeing the ad.”

And there is still not one definition of a viewable impression. Microsoft’s auction platform documents this plainly:

“The vCPM payment type supports the ability to use a custom definition for viewable impressions. We currently support two definitions, the IAB standard and the Group M standard. The IAB standard definition is used by default.”

Two competing standards, natively implemented in the same auction infrastructure. And inside the official standard, measuring 50% of the ad’s pixels or 50% of the video player’s pixels both remain acceptable, provided the choice is disclosed.

The arithmetic consequence is that two platforms dividing the same spend by structurally different denominators are not reporting the same metric. One honest caveat: no standards body has ever published that sentence. It follows from the structure of the guidelines and it is logically airtight, but it is a deduction, not a citation. What does exist on the record is the industry body’s own acknowledgement of measurement disparity between vendors scoring the same campaigns.

The four rungs of the advertising impression counting chain and the associated viewability thresholdsThe four rungs of the advertising impression counting chain defined by the industry measurement standards, and the viewability thresholds associated with them. The first rung is the served impression, defined as a response from an ad delivery system to an ad request from the user’s browser. The second rung is the rendered impression, which adds the condition that the advertisement must have loaded and at minimum begun to render; this condition was only imposed in October two thousand seventeen for display and June two thousand eighteen for video, the earlier standard of two thousand four having explicitly permitted counting at the moment the server decided to serve the advertisement, a method known as count on decision or count on insertion, which the revised standard describes as not acceptable because it is the furthest away from the user actually seeing the ad. The third rung is the measured impression, comprising those impressions whose viewable status could be determined, whether viewable or not viewable, and excluding those whose status could not be determined. The fourth rung is the viewable impression, which for display requires that at least fifty percent of the advertisement’s pixels be on an in focus browser tab in the viewable space of the browser page for a period greater than or equal to one continuous second after the ad renders, with that threshold reduced to thirty percent for formats of two hundred and forty two thousand five hundred pixels or larger, and raised to two continuous seconds for video. Audio has a separate standard founded on audibility rather than pixels, excluding advertisements served during a known player muted state and those audible for less than two continuous seconds. Two competing definitions of the viewable impression coexist in practice: one major auction platform’s official documentation states that it supports two definitions, that of the industry standards body and that of an agency group, with the former applied by default. Within the official standard itself, measuring fifty percent of the advertisement’s pixels or fifty percent of the video player’s pixels both remain permissible provided the choice is disclosed and supported by evidence that the impact is immaterial.Four rungs, four possible denominatorsServed impressionThe ad server answered a request.Rendered impressionThe ad began to render. Required only since 2017.Measured impressionIts viewable status could actually be determined.Viewable impression50% of pixels, 1 second. 2 seconds for video.And two competing definitions of “viewable” run natively inside the same auction platform.
Each rung holds fewer units than the one above. A CPM computed at the top is not the same metric as one computed at the bottom. Source : MRC, Viewable Ad Impression Measurement Guidelines v2.0, 18 August 2015; IAB Tech Lab and MRC, Desktop Display Impression Measurement Guidelines v1.1, October 2017; Microsoft Xandr documentation, vCPM payment type, 21 October 2025 (2015)

Thirty-six cents on the dollar

The American advertiser association ran the arithmetic in 2023, across 21 advertisers, $123 million of spend and 35.5 billion impressions, with data from three demand-side platforms, six supply-side platforms and three verification vendors. Its conclusion:

“After accounting for both transaction costs and loss of media productivity costs, only 36 cents of every ad dollar that enters a DSP effectively reaches the consumer. That is TrueAdSpend.”

The breakdown, on $1,000 entering a demand-side platform: $160 to the DSP in platform, feature and data fees, $130 to the supply-side platform, leaving $710 of seller revenue, of which $95 is lost to non-viewable impressions, $5 to invalid traffic, $150 to impressions that could not be measured at all, and $100 to spend on made-for-advertising sites.

Then the caveat that makes the number worse rather than better:

“Again, agency fees (a transaction cost) and brand safety (a loss of productivity cost) were both beyond the scope of this project. Depending upon those costs, less than 36 cents of every dollar would effectively reach the consumer.”

A British study three years earlier reached a compatible conclusion by a different route, tracking 267 million impressions and finding 51% of advertiser spend reaching the publisher, with 15% unattributable. Its authors were candid about the limits of their own visibility, listing data gaps, invisible fees, post-auction bid shading, financing arrangements, currency translation and inventory reselling as possible explanations. They were equally candid about their sample: it covered “the most premium parts of programmatic”, and they expected the long tail to look worse.

The finding that should change how you read a campaign report comes from the same American study:

“Notably, media CPMs paid on MFA websites are 25 percent lower than those paid on non-MFA websites. All this makes MFA websites attractive to DSP bidding algorithms.”

Made-for-advertising sites, the ones stacking ad slots around templated content, sell for a quarter less. They score well on surface metrics, which is exactly why bidding algorithms favor them, while ads on them are “at least 50 percent less likely to be attributed with driving a sale”. Cheap inventory is often cheap for a reason.

One negative result worth carrying with you, because it tells you how solid all of this is: inside the American study, two measurement platforms processed the same exercise and produced an untraceable share of 0% for one and 17% for the other. How much money you cannot follow depends on the tool you use to follow it.

Where one thousand dollars entering a demand side platform actually goesWhere one thousand dollars entering a demand side platform actually goes, according to a study by an American advertiser association covering twenty one advertisers, one hundred and twenty three million dollars of spend and thirty five point five billion impressions between September two thousand twenty two and January two thousand twenty three, with participation from three demand side platforms, six supply side platforms and three advertising verification vendors. Transaction costs account for twenty nine percent of the advertising dollar: eighty dollars in demand side platform fees, twenty dollars in additional platform tool and feature fees, sixty dollars in data fees, and one hundred and thirty dollars in supply side platform fees, leaving seven hundred and ten dollars of seller revenue. Loss of media productivity then accounts for thirty five percent of the advertising dollar, calculated against that seller revenue: ninety five dollars lost to non viewable impressions, five dollars lost to invalid traffic, one hundred and fifty dollars lost to impressions that could not be measured for viewability at all, and one hundred dollars spent on made for advertising websites. The result is that only three hundred and sixty dollars of every thousand, or thirty six cents of every dollar, effectively reaches the consumer. The study states explicitly that agency fees, which are a transaction cost, and brand safety, which is a productivity cost, were both beyond its scope, so that depending on those costs less than thirty six cents of every dollar would effectively reach the consumer. A separate British study published three years earlier, tracking two hundred and sixty seven million impressions of which thirty one million could be matched between the buy side and the sell side, found that fifty one percent of advertiser spend reached the publisher and that fifteen percent of the spend could not be attributed at all, its authors listing data limitations, fees invisible in the study data, post auction bid shading, post auction financing arrangements, foreign exchange translations and inventory reselling between technology vendors as possible explanations, and noting that their sample represented the most premium parts of programmatic such that the long tail would presumably reinforce their findings.$1,000 in, $360 outTransaction costsDemand-side platform, fees and data$160Supply-side platform$130Lost to quality, out of $710 of seller revenueNever measurable for viewability$150Made-for-advertising sites$100Non-viewable$95Invalid traffic$5$360 reaches the consumer. And agency fees were outside the study’s scope.
Fees take 29%. Quality problems take most of the rest. Agency fees were outside the study's scope. Source : ANA, Programmatic Media Supply Chain Transparency Study, December 2023; ISBA and PwC, Programmatic Supply Chain Transparency Study, May 2020 (2023)

You do not bid on it, and you do not pay your bid

On Meta, the impression is a billing event, not the object of the auction. The developer documentation keeps the two apart and says so directly:

“You can also set objective and billing_event but neither directly impacts bid_amount or your effective bid.”

What the system evaluates is your bid combined with the estimated probability of the outcome you asked for, plus a quality assessment. Meta is explicit that this can beat money: “an ad that’s more relevant to a person could win an auction against ads with higher bids.” It also caps what you pay, without publishing the formula: “such auction adjustments will not cause us to charge you more than your bid to show your ad.”

On Google, target CPM bidding exists only for video campaigns. Everywhere else the billing rule is stated plainly:

“You’ll often pay less than your maximum bid because you’ll only pay what’s minimally required to hold your Ad Rank.”

Three mechanisms then move the observed price with no change in advertiser behavior at all.

The auction type. In March 2019 Google moved Ad Manager display and video inventory to a first-price auction, where “the buyer that wins the auction pays the price they bid”, instead of paying only what was needed to beat the next bidder. A different quantity now settles, so the average observed CPM shifts even at identical demand. Worth flagging: Google published no numerical projection of the price effect. Every figure in circulation about that transition comes from the trade press.

Relevance. All three major platforms document that a more relevant ad pays less. LinkedIn puts it in one line: “The more relevant your ad, the lower the price you pay.” Two advertisers bidding the same amount on the same audience therefore do not pay the same price. One nuance almost nobody reproduces correctly, from Google’s own help center:

“Quality Score is not an input in the ad auction. It’s a diagnostic tool to identify how ads that show for certain keywords affect the user experience.”

Floors. They are no longer fixed barriers. Google Ad Manager “dynamically sets the floor prices on individual bid requests”, with a machine learning model trading fill rate against inventory value. Your CPM moves partly because the publisher on the other side is optimizing theirs.

And a systematic absence worth naming: across Meta, Google, LinkedIn, Amazon and TikTok, none publishes an average CPM, a floor price in currency, or a fill rate.

Relationship between the advertiser bid the price actually paid and the observed cost per thousandRelationship between the advertiser’s bid, the price actually paid, and the observed cost per thousand impressions, according to official platform documentation. In most campaigns the advertiser bids on an optimization goal such as a click or a conversion rather than on the impression. One major platform’s developer documentation classifies the impression as a billing event distinct from the optimization goal, and states that setting the objective and the billing event does not directly impact the bid amount or the effective bid. That platform also states that the winner of the auction is the ad with the highest total value, combining the bid, the estimated action rate and ad quality, that an ad more relevant to a person could win an auction against ads with higher bids, and that auction adjustments will not cause it to charge more than the advertiser’s bid, without publishing the formula determining what is charged below that ceiling. On another platform, target cost per thousand bidding is available only for video campaigns, and the documentation states that advertisers will often pay less than their maximum bid because they pay only what is minimally required to hold their ad rank. The cost per thousand is therefore a quotient observed afterwards, equal to spend divided by counted impressions and multiplied by one thousand. Three mechanisms move this quotient without any change in advertiser behaviour. The first is the auction type: in March two thousand nineteen one platform transitioned its display and video inventory to a first price auction in which the buyer that wins pays the price they bid, replacing a second price mechanism, which shifts the average observed price even at identical demand; that platform published no numerical projection of the price effect. The second is relevance: all three major platforms document that a more relevant advertisement pays a lower price, so two advertisers bidding identical amounts on the same audience do not pay the same price, with the nuance that the quality score displayed in one platform’s interface is described by that platform as a diagnostic tool and not an input in the ad auction. The third is the price floor, which is no longer a fixed barrier: the documentation states that floor prices are set dynamically on individual bid requests using a machine learning model trading off fill rate against long term inventory value. None of the five major platforms examined publishes an average cost per thousand, a floor price in currency, or a fill rate.You bid here, the CPM shows up thereYou bid on a click or a conversion, not on the impressionThe system combines your bid, the estimated action rate and ad qualityYou pay “what’s minimally required to hold your Ad Rank”, never more than your bidThe CPM is the division you read afterwardsThree things move it while you change nothingThe auction typeYour ad’s relevanceThe publisher’s floor
The CPM sits at the end of the chain. Three mechanisms move it without you touching anything. Source : Meta, Marketing API Bidding overview and About ad auctions; Google Ads Help, Choose your bid and budget and What matters when it comes to ads quality; Google Ad Manager blog, 6 March 2019; Google Ad Manager Help, Understand target CPM; LinkedIn Help, advertising cost and pricing (2026)

What genuinely moves the price

Seasonality, the only factor a listed company quantifies. Meta’s annual report states it with three years of arithmetic:

“our total revenue increased 17%, 19%, and 17% between the third and fourth quarters of 2025, 2024, and 2023, respectively, while total revenue for the first quarters of 2025, 2024, and 2023 declined 13%, 9%, and 11% compared to the fourth quarters”

Pinterest and Snap describe the same year-end peak in their own filings. Alphabet names seasonality as a risk factor and never quantifies it. On US elections, routinely blamed for CPM spikes: none of these filings connects elections to advertising prices. The 15% to 50% increases you read about come from media agencies.

Geography. Through late 2023, Meta published average revenue per user by region. In the fourth quarter of 2023: $68.44 in the United States and Canada against $5.52 in Asia-Pacific, a factor of 12.4, and stable since 2021. Meta attributes it to “the size and maturity of those online and mobile advertising markets”. The necessary caveat: Meta stopped publishing the regional breakdown from 2024, so the series ends there.

Audience scarcity, but not in the direction you would assume. Intuition says narrow targeting costs more. The peer-reviewed literature says something more precise: finer targeting reduces the number of bidders competing for a given impression. A review in Information Systems Research puts it this way:

“providing more information decreases the number of participating bidders and creates thin markets, as fewer advertisers are interested in a given impression with highly differentiated attributes”

Two forces pull against each other: thinner competition pushes the price down, a more valuable audience pushes it up. The net effect is not settled empirically, and the authors flag it as an open research question. What platforms do document are thresholds: LinkedIn requires 300 member accounts minimum and recommends 50,000, and states that “the cost required to win the auction depends on the desirability of your target audience”.

Available signal, also against the prevailing story. The common claim is that privacy restrictions push CPMs up. A listed company says the opposite in an audited filing. Snap, on Apple’s changes:

“This has resulted in, and in the future is likely to continue to result in, reduced demand and pricing for our advertising products and could seriously harm our business.”

Less signal means less value per impression, so lower prices, not higher. And since a great deal of published material is out of date on this: third-party cookies were not removed from Chrome. That decision was reversed in April 2025, and most of the replacement technologies were retired in October 2025.

Seasonality and geographic variation in advertising value as reported in audited securities filingsSeasonality and geographic variation in advertising value as reported in audited securities filings. On seasonality, one large platform’s annual report states that its total revenue increased seventeen, nineteen and seventeen percent between the third and fourth quarters of fiscal years two thousand twenty five, two thousand twenty four and two thousand twenty three respectively, and that total revenue for the first quarters of those years declined thirteen, nine and eleven percent compared with the preceding fourth quarters. Two other listed platforms describe the same year end peak in their own filings, one stating that industry advertising spend tends to be strongest in the fourth quarter resulting in higher revenue, and the other that overall advertising spend tends to be strongest in the fourth quarter of the calendar year. A fourth platform names seasonality as a risk factor without ever quantifying it. None of these filings connects elections to advertising prices, so the widely quoted increases of fifteen to fifty percent during election periods rest on media agency estimates rather than audited sources. On geography, the same large platform published average revenue per user by region through the end of two thousand twenty three: in the fourth quarter of that year the figure was sixty eight dollars and forty four cents in the United States and Canada, twenty three dollars and fourteen cents in Europe, five dollars and fifty two cents in Asia Pacific and four dollars and fifty cents in the rest of the world, giving a ratio of twelve point four between the highest and the lowest, a ratio that had been stable since two thousand twenty one. The company attributes the gap to the size and maturity of those online and mobile advertising markets, and cautions that its data on the geographic location of users is estimated from factors such as internet protocol address and self disclosed location which may not always accurately reflect actual location. That regional breakdown has not been published since two thousand twenty four, when the company moved to a single worldwide average revenue per person measure, so the series ends with the fourth quarter of two thousand twenty three.The two factors an audited filing actually quantifiesRevenue change between Q3 and Q4, three years running2025+17%2024+19%2023+17%And Q1 falls back:-13%, -9%, -11%Average revenue per user, Q4 2023, by regionUnited States and Canada$68.44Europe$23.14Asia-Pacific$5.52Rest of world$4.50A factor of 12.4, stable since 2021. The company stopped publishing this breakdown in 2024.
Both come from audited filings. One of the two series stopped being published in 2024. Source : Meta Platforms, annual report for fiscal year 2025 filed 29 January 2026, and annual report for fiscal year 2023; Snap, annual report for fiscal year 2025; Choi, Mela, Balseiro and Leary, Information Systems Research 31(2), 2020 (2026)

The benchmark you are looking for does not exist

No sector report publishes a unit price. The reference report for the US market measures reported revenue, and its own auditor states: “PwC does not audit the information and provides no opinion or other form of assurance with respect to the information.” The telling detail: when that report wants to reference a CPM level, it cites a third-party research firm, because it does not hold the data itself. The European equivalent, covering thirty national markets, contains zero occurrences of the term CPM across its entire text. The French observatory measures publisher revenue, not prices.

One public institution publishes a price index. The Bureau of Labor Statistics maintains a producer price index for internet advertising sales, series WPU366 and WPU366101, based at 100 in December 2022. It is the only official measure with no commercial interest attached. Its limits rule it out as a campaign reference: it starts at the end of 2022, it is aggregated nationally with no breakdown by platform or advertiser industry, it swings ten to fifteen points month to month unadjusted, and it is an index rather than a price in dollars.

Among listed companies, only one still publishes a unit price measure. Meta reports the year-over-year change in its average price per ad, defined as “total advertising revenue divided by the number of ads delivered”. Across the last six published quarters that change has run between +9% and +12%. Meta attaches two warnings: the geographic breakdown is affected by measurement limitations, and formats are not monetized alike, since “our video and Reels products are not currently monetized at the same rate as our Feed or Stories products”. It is an index of mix, not a CPM.

And the most datable fact in this whole file: Alphabet stopped publishing its cost-per-click series. It had run for years and still appeared in the fiscal 2024 annual report, showing paid clicks up 5% and cost-per-click up 7%. It is absent from the fiscal 2025 report filed on 5 February 2026. The only quasi-official price series for the search market has just gone dark.

So where do the industry CPM tables come from? From software vendors publishing statistics about their own paying customers. One of the most reproduced draws on 554 to 726 campaigns in total across fifteen industries, some subcategories resting on two or three campaigns, with “averages” that are actually medians, and a methodology section describing search campaigns inside a report about social ads. Another aggregates customers of an e-commerce analytics tool, all direct-to-consumer brands, whose figures are then republished as sector benchmarks covering finance, B2B and real estate.

Advertising price sources that actually exist and the nature of what each one measuresAdvertising price sources that actually exist and the nature of what each one measures. The reference report for the American advertising market, published by an industry body with an accounting firm, measures revenue reported directly by companies selling advertising rather than unit prices, and that accounting firm states that it does not audit the information and provides no opinion or other form of assurance with respect to it; when the report wishes to reference a cost per thousand level it cites a third party research firm because it does not hold such data itself. The European equivalent, a meta study covering thirty national markets, contains zero occurrences of the term cost per thousand across its entire text and measures aggregated advertising spend in euros. The French sector observatory measures revenue generated by the sale of owned inventory by advertising sales houses, without publishing unit prices. The only public institution publishing a price index is the American federal statistical agency, which maintains a producer price index for advertising space and time sales covering internet advertising sales, based at one hundred in December two thousand twenty two, published under several series identifiers corresponding to different aggregation levels; that series is aggregated nationally with no breakdown by platform or by advertiser industry, begins only at the end of two thousand twenty two which prevents any earlier comparison, moves ten to fifteen points month to month in non seasonally adjusted terms, and expresses an index rather than a price in currency. Among listed companies only one still publishes a genuine unit price measure, the average price per advertisement, defined as total advertising revenue divided by the number of advertisements delivered, whose year over year change has run between nine and twelve percent across the last six published quarters, accompanied by two warnings from the company itself concerning measurement limitations in user geography and the fact that its video and short form products are not currently monetized at the same rate as its feed and stories products. Another company published for years the year over year change in its cost per click and cost per impression, a series still present in its annual report for fiscal year two thousand twenty four, which showed paid clicks up five percent and cost per click up seven percent, and absent from the annual report for fiscal year two thousand twenty five filed on the fifth of February two thousand twenty six. The industry tables of average cost per thousand by sector come from advertising or analytics software vendors republishing statistics drawn from their own paying customer base, one of the most reproduced resting on five hundred and fifty four to seven hundred and twenty six campaigns in total across fifteen industries, with subcategories based on two or three campaigns and displayed averages that are in fact medians.What price sources genuinely existSourceWhat it measuresUS sector reportrevenue, not pricesEuropean sector reportspend. Zero occurrences of “CPM”French observatorysales house revenueFederal producer price indexan index, base 100 in December 2022Average price per ad, Metaan index of mix, +9% to +12% a yearCost-per-click, Alphabetremoved from the 2025 annual report”Average CPM by industry” tablesone software vendor’s customer baseNo institution publishes an average CPM by industry or platform. Not in the US, not in Europe, not in France.
One official public index, one platform measure still published, and no institutional average CPM anywhere. Source : IAB and PwC, Internet Advertising Revenue Report, April 2025; IAB Europe, AdEx Benchmark 2024; BLS series WPU366 and WPU366101; Meta Platforms, Q2 2026 quarterly report; Alphabet, annual reports for fiscal years 2024 and 2025 (2026)

What to do with this

Stop comparing CPMs across platforms. You are comparing two fractions where neither the numerator nor the denominator is built the same way. The only comparison that carries information is one platform against itself, over time, at constant targeting and format.

Ask what you are billed on. Served impression or viewable impression is the first question to put to a platform or an agency, and it is almost never asked. If the answer does not come in writing, the CPM on your report is not interpretable.

Treat an unusually low CPM as a warning, not a win. The cheapest inventory on the market is measurably the least effective, and it is attractive to bidding algorithms precisely because it is cheap and presentable.

Look at cost per result, not cost per thousand. CPM sits upstream of everything you actually care about. It becomes useful only when divided into what it produced, and that is the one metric your own system can measure without depending on anybody else’s definitions. Reporting on that single number rather than on the ones the platforms volunteer is the discipline our B2B paid acquisition work runs on.

And import no benchmark. There is no institutional one. What you will find online is a software vendor’s customer base, sometimes a handful of campaigns per industry, with an impression definition nobody states. Your own history, on your own campaigns, is the only defensible reference you have.