Wasted
Adspend

Internet advertising is priced in impressions, and a machine can now produce an impression for almost nothing. This report follows a dollar of ad spend through what that has done to it.

PREPARED FOR PANTERA CAPITAL
EARNOS RESEARCH · 7 SEPTEMBER 2026
The trend

Machines have sent more requests for web pages than people have since 2024.

The chart shows the share of requests for web pages, which is the surface advertising gets served into. Imperva measured the first three points on its own network and Cloudflare measured the last on a different one, and for 2025 the two came within a third of a percentage point of each other.

40%45% 50% 55%60% 20232024 2025JUN 2026 CROSSOVER MACHINES 57.5% PEOPLE 42.5%

Machines took the majority at some point during 2024, and the line has steepened since then, because the eighteen months to June 2026 added more share than the two years before them. Cloudflare's chief executive had forecast the crossing for late 2027 and said it arrived around eighteen months early.

Which basket

These figures cover requests for HTML, the pages that carry advertising. When Cloudflare counts every content type on its network, which brings in video, images, scripts and API calls, its Radar dashboard puts bots at 35.4% for August 2026, up from 30.3% a year earlier. Both baskets move the same way. The level depends on which one is counted, so the basket needs naming whenever the figure is used.

The response

Meta has been showing more advertising to an audience that has almost stopped growing.

Meta reports these figures in its quarterly filings. Across the six quarters shown, impression volume grew about four times faster than the audience it was delivered to.

Q1 25Q2 25Q3 25Q4 25Q1 26Q2 26
Ad impressions delivered+5%+11%+14%+18%+19%+14%
Average price per ad+10%+9%+10%+6%+12%+12%
Daily active people+8%+7%+3%
ALL FIGURES YEAR OVER YEAR, FROM META'S QUARTERLY FILINGS. BLANK CELLS WERE NOT DISCLOSED IN THAT QUARTER'S RELEASE.

Audience growth slowed from 8% to 3% over the period. Impression growth ran between 14% and 19% for the last four quarters, price growth sat at 12% for the last two, and so the same people were absorbing more advertising at a higher price.

The divergence

The scorecard improved through a year of worse results.

The industry spent 2025 and 2026 cleaning up its programmatic supply chain, and by its own measures the effort succeeded. The columns below put those measures beside what advertisers were dealing with over the same months.

Industry quality metrics
Share reaching quality impressions37%45.1%
Detected invalid traffic0.3%0.1%
Spend on non-viewable inventory13.3%10.1%
Record highs on the ANA benchmark, with invalid traffic at its lowest recorded level.
Market conditions
Machine share of web page requests51%57.5%
Click-through lost to an AI summary34.5%58%
Meta price per ad, year over year+9%+12%
Made-for-advertising share of spend0.6%1.1%
Gambling install fraud, Android49%59%
Owned media fraud rate3.4%11%
Machine-written pages now outscore real ones on all three metrics in the left column.
The quality controls are being satisfied by things they were never built to detect.

When a measurement keeps improving as the thing it measures gets worse, the two have come apart. Every control in the left column still returns a passing grade, and that grade has stopped indicating that a person saw an advertisement.

The mechanism

Advertising is priced in a unit that machines can now manufacture for free.

An impression stands in for a moment of human attention, and for thirty years the substitution held because producing one needed a person sitting in front of a screen with a page loaded. Generative models removed that requirement inside eighteen months.

A machine can now write the page, load it, render the advertisement, hold it on screen for the required second, click through, install the app and fire a purchase event, and each of those steps passes the control that was built to catch it. AppsFlyer describes fraud operators running on-device AI agents that add items to carts and trigger small purchases. TAG found machine-generated inventory grading as premium more than 70% of the time, at a higher price than the real thing.

Cost per install makes fraudulent supply paths look efficient, so the optimiser keeps buying them.

The damage compounds because the bidding system learns from what it buys. A model treats an install as a success and moves budget toward whichever supply path produced it. When InMobi analysed more than ten million supply paths it found roughly 5% of them generating over 65% of all detected fraudulent installs, and those same paths contributed almost no real conversions. Cost per install makes them look efficient, and the bidding model has no way to see past that.

Buying more carefully does little here, because generative models have been lowering the cost of a convincing fake impression faster than detection has been raising it, in every year since 2023.

Q2 2026

Detected invalid traffic in measured programmatic reached 0.1% in the second quarter of 2026, its lowest recorded level. In that same quarter AI-generated inventory scored above real inventory on every quality control the industry sells.

Where it goes

Under seven cents of a programmatic dollar reaches a person who might be interested.

The Association of National Advertisers reconciles impression-level log data from member marketers with TAG TrustNet and Fiducia every quarter. Its Q1 2026 sample covered 86 marketers, 20.9 billion impressions and $160 million of spend, which makes the first four steps below the closest thing to an audit that programmatic has.

$1.00
Budget committed
$0.728
Less 27.2¢ to platform, data and verification fees. This number rose last quarter.
$0.553
Less 17.5¢ to unmeasurable inventory, made-for-advertising pages and AI slop.
$0.451
Less 10.1¢ rendered off-screen and 0.1¢ of caught invalid traffic. This is TrueAdSpend, an all-time high.
$0.136
Less the 70% of viewable ads that eye-tracking shows nobody looks at.
$0.066
Less targeting error, using the 51% inaccuracy Truthset measures in third-party data.
STEPS 1 TO 3 ARE MEASURED BY THE ANA. STEPS 4 AND 5 JOIN SEPARATE STUDIES ONTO THAT BASE, SO THE BOTTOM TWO BARS ARE OUR MODEL RATHER THAN AN OBSERVATION.

Nothing in that chart accounts for whether the advertising changed anyone's behaviour. Gordon and Zettelmeyer ran fifteen large randomised trials on Facebook and compared the results against the observational methods in common industry use. In half the studies the observational estimate of purchase lift was wrong by a factor of three.

Your budget

Put a number in.

This applies the same steps to a figure of your choosing. Type an annual programmatic budget or move the slider.

$
From $100k to $100m
Platform, data and verification fees-$2,720,000
Unmeasurable, made-for-advertising and AI slop-$1,750,000
Rendered off screen-$1,010,000
Invalid traffic the industry caught-$10,000
Reaches a quality impression$4,510,000
Rendered, viewable, never looked at-$3,157,000
Shown to the wrong person-$690,030
Reaches an interested human$662,970
$663kProgrammatic, after every step above
$7.50mSame budget through EarnOS Missions, paid to verified humans and creators
11.3×Difference
Steps one to four are measured by the ANA. Steps five and six join Lumen and Truthset onto that base and are a model. The EarnOS figure is the 75% of a Missions budget paid to the person completing the action and to the creator or agency.
The complete map

Seven kinds of waste, measured separately.

Each category has its own vendors and its own denominator, so the industry rarely adds them together. The seven lines below give the headline figure for each one, and the rows beneath hold the evidence for anyone who wants to check it. Gold marks show where generative AI moved a number in the past eighteen months.

01The supply chain takes its cut
27.2%of a programmatic dollar goes to fees charged by platforms and by the data and verification vendors around them, before any media is bought. The figure rose in Q2 2026, a quarter in which media quality also improved.ANA Programmatic Transparency Benchmark, Q2 2026
32%is what the weakest cohort of advertisers converts into qualified impressions, against 54% for the strongest. Media quality explains almost all of that gap. Fees differ by 2.4 points between the two cohorts.ANA, Q1 2026
02The inventory is manufactured for adsAI
1.3 to 2.4%of open-web programmatic spend goes to AI slop, meaning pages generated to carry advertising. Slop pages record an invalid traffic rate of 0.05% against 0.32% on clean inventory, reach 77.2% viewability against 74.9%, grade as premium more than 70% of the time, and sell for $7.08 per thousand against $6.15.TAG TrustNet, ANA and Fiducia, July 2026
1.1%goes to made-for-advertising sites, up from 0.6% a quarter earlier. That is the first rise since the ANA crackdown started in 2023, when members were putting 21% of budgets into MFA.ANA, Q1 2026
91%higher conversion rate on quality sites than on the ad-clutter sites IAS classifies, with cost per conversion 25% lower.Integral Ad Science
03The traffic is not humanAI
53%of requests for web pages came from machines by the end of 2025. Cloudflare classified HTML traffic across its network as 47% human, 44% non-AI bots, 4.2% AI bots and 4.5% Googlebot. Imperva measured the same year separately and published 53%.Cloudflare Radar Year in Review and Imperva, 0.3 points apart
57.5%by June 2026, when Cloudflare's chief executive said the crossover had arrived around eighteen months ahead of his own forecast. Cloudflare carries roughly a fifth of all websites and handles 81 million HTTP requests a second.Cloudflare Radar, June 2026
40%of internet traffic is malicious bad bots, up three points. Imperva blocked 17.2 trillion malicious bot requests and saw AI-driven attacks rise twelvefold to 25 million a day.Imperva Bad Bot Report 2026
7,851%year-over-year growth in traffic from AI agents and agentic browsers, which move through pages and complete forms on a person's behalf. Automated traffic overall grows about eight times faster than human traffic.HUMAN Security, March 2026
140%more connected TV fraud schemes in Q1 2026 than a year earlier. Unprotected CTV campaigns run near 9% fraud against under 1% on protected ones, and 21% of advertisers track fraud as a CTV metric at all.DoubleVerify, May 2026
04The conversions are fabricatedAI
275%Real Users Lift on Social Media iOS in Q2 2025, meaning advertisers received one real user for every four they paid for across a full quarter. Gambling on Android reached 175% in Q4.AppsFlyer, 106.4B installs across 246K apps
59%app install fraud rate for Gambling on Android in Q1 2026, up from 49%, peaking at 64%. Finance Android has held at 50 to 53% Real Users Lift for five straight quarters.AppsFlyer State of Fraud, Q1 2026
36×gap between affiliate fraud rates near 40% and self-reporting network rates near 1%. Owned media fraud rose 221% in a year as scrutiny tightened elsewhere.AppsFlyer, 55.3B paid installs analysed
Spoofingwas the fastest-rising technique of 2025. It fabricates the device and every event that device appears to send, so what reaches the measurement layer looks like a clean signal.AppsFlyer State of Fraud, 2026 edition
05Nobody looks at the ad
70%of viewable ads are never looked at, measured on consented eye-tracking panels. The viewability standard asks for half the pixels on screen for one second.Lumen Research
29.5%of internet users run an ad blocker at least sometimes, around 1.77 billion people.GWI, Q2 2025
06It reaches the wrong person
51%of ad targeting data is inaccurate at the top of the range, with provider accuracy spanning 32% to 69%. Accuracy degrades further at each step between the data provider and the media channel.Truthset
the size of the error in half of fifteen randomised trials at Facebook, when observational attribution was compared against the experimental result.Gordon, Zettelmeyer, Bhargava and Chapsky, Marketing Science
07The audience stops arrivingAI
8%of Google searches produce a click on a result when an AI summary appears, against 15% without one. Summaries now trigger on roughly 48% of tracked queries.Pew Research, 68,879 tracked searches
500,000:1peak crawl-to-refer ratio for Anthropic in 2025, counting page fetches against humans sent back to the source. OpenAI reached 3,700:1 and Google ran between 3:1 and 30:1. AI user-action crawling grew 21 times over the year.Cloudflare Radar
+12%average price per ad at Meta in Q2 2026, with impressions up 14% and the audience seeing them up 3%, so the same people absorbed more advertising at a higher price.Meta Form 10-Q, June 2026
The defence

The counter-spend is large and it comes out of the same budget.

DoubleVerify took $748 million of revenue in 2025, guiding to as much as $826 million for 2026, and Novacap paid $1.9 billion to take Integral Ad Science private that December. On the platform side, Google blocked 8.3 billion ads last year, a rise of 63%, though it suspended 36% fewer accounts. Meta removed 159 million scam ads over the same period and has set itself a target of routing 90% of ad revenue through verified advertisers by the end of 2026. Cloudflare, sitting underneath all of it, now sends more than a billion payment-required responses a day to AI crawlers.

Advertisers pay for the verification layer through the 27.2 cents at the front of the dollar, so the cost of the defence sits inside the figure it exists to protect.

Some of the spending carries a conflict. Reuters reported in November 2025, from internal documents, that Meta projected around 10% of 2024 revenue would come from advertising for scams and banned goods, with a 95% certainty threshold before disabling an advertiser and a $135 million cap on revenue it would forgo to clean up. Meta calls that figure rough and overly inclusive.

On sourcing

Every figure here comes from a primary source that is named in the list below. Syndicated market-size estimates were excluded because published 2026 numbers for the fraud detection category span $461 million to $89.9 billion.

A widely repeated claim that Juniper Research forecasts $131 billion of ad fraud by 2030 misreads a February 2025 release about eCommerce fraud. Juniper's advertising series runs from $84 billion in 2023 to $172 billion in 2028.

EarnOS

EarnOS pays for actions a member has proved.

A member of ero connects an account they already hold and use, which might be Spotify, Netflix, Uber, a bank or a retailer. The app then uses zkTLS to produce a cryptographic attestation on the member's own device, proving that a specific action happened inside that account without handing over the password or any of the underlying data.

Brands fund payouts against those proven actions, and because the money attaches to the action itself, impression valuation and viewability thresholds drop out of the process.

66%of a Missions budget goes to the person who completed the action
9%to the creator or agency
25%platform fee to EarnOS

Three quarters of a Missions budget reaches a named human who did what a brand asked and can show it, against under seven cents of a programmatic dollar once attention and targeting error are applied.

A cryptographic proof of an action inside a real account holds its value as generative models improve, because there is no behavioural signal in it for a model to imitate.
Sources