19 Apr 2025
Google Ads Shitshow Report 2024
previous: Winners don’t click search ads
Google ads are full of crime and most web users should block them. If you don’t believe the FBI, or Malwarebytes, believe Google. Their 2024 Ads Safety Report is out (Search Engine Land covered it) and things do not look good. The report is an excellent example of some of the techniques that big companies use to misrepresent an ongoing disaster as somehow improving, so I might as well list them. If I had to do a corporate misinformation training session, I’d save this PDF for a reading assignment.
release bad news when other news is happening This was a big news week for Google, which made it the best time to release this embarrassing report. Editors aren’t going to put their Google reporter to work on an ad safety story when there’s big news from the Federal courthouse.
count meaningless numbers Somehow our culture teaches us to love to count, so Google gives us a meaningless number when the meaningful numbers would look crappy.
Last year, we continued to invest heavily in making our LLMs more advanced than ever, launching over 50 enhancements to our models which enabled more efficient and precise enforcement at scale.
The claim is that Google continued to invest heavily
and
that’s the kind of statement that’s relatively easy to back up with a
number that has meaningful units attached. Currency
units, head count, time units, even lines
of code. Instead, the count is enhancements
which could be
almost anything. Rebuild an existing package with different compiler
optimizations? Feed an additional data file to some ML system? What this
looks like from the outside is that the meaningful numbers are going in
the wrong direction (maybe some of the people
who would have made them go up aren’t there any more?) so they
decided to put out a meaningless number instead.
control the denominator to juice the ratio Only takes elementary school math to spot this, but might be easy to miss if you’re skimming.
Our AI-powered models contributed to the detection and enforcement of 97% of the pages we took action on last year.
Wow, 97%, that’s a big number. But it’s out of pages we took
action on
which is totally under Google’s control. There are a bunch
of possible meaningful ratios to report here, like
(AI-flagged ads)/(total ads)
(ads removed)/(AI-flagged ads)
(bad ad impressions)/(total ad impressions)
and those could have been reported as a percentage, but it looks like they wanted to go for the big number.
pretend something that’s not working is working The
AI models contributed
to 97% of the actions, but
contributed
isn’t defined. Does it count as contributed
if, say, human reviewers flagged 1,000 ads, the AI
flagged
100,000 ads, and 970 ads were flagged by both? If AI were flagging ads
that had been missed by other methods, this would have been the place to
put it.
The newsworthy claim that’s missing is the count of bad ads first
detected by AI before getting caught by a human reviewer. Contributed
to the detection
could be a lot of things. (If this were a report on
a free trial of an AI-based abuse detection service, contributed
wouldn’t get me to upgrade to the paid plan.)
report the number caught, not the number that get through Numbers of abusers caught is always the easiest number to juice. The simplest version is to go home at lunch hour, code up the world’s weakest bot, start it running from a non-work IP address, then go back to work and report some impressive numbers.
To put this into perspective: we suspended over 39.2 million accounts in total, the vast majority of which were suspended before they ever served an ad.
Are any employees given target numbers of suspensions to issue? Can anyone nail their OKRs by raising the number of accounts suspended? If this number is unreliable enough that a company wouldn’t use it for management, it’s not reliable enough to pay attention to. They’re also reporting the number of accounts, not individuals or companies. If some noob wannabe scammer writes a script to POST the new account form a million times, do they count for a million?
don’t compare to last year Here’s the graph of bad ads caught by Google in 2024.
And here’s the same graph from the 2023 report.
The total number isn’t as interesting as the individual, really
problematic categories. The number caught for enabling dishonest
behavior
went down from about 20 million in 2023 to under 9 million
in 2024.
Did the number of attempts at dishonest behavior with Google ads
really go down by more than half in a single year? Or did Google catch
fewer of them? From the outside, it’s fairly easy to tell that Google
Ads is understaffed and the remaining employees are in the weeds,
but it’s hard to quantify the problem. What’s really compelling about
this report is that the staffing situation has gotten bad enough that
it’s even showing up in Google’s own hand-picked numbers. In general
when a report doesn’t include how a number has changed since the last
report, the number went in the wrong direction and there’s no good
explanation for why. And the number of ads blocked or removed for
misinformation
went from 30 million in 2023 to (checks notes)
zero in 2024. Yes, misinformation
has friends in high places now, but did all of the sites worldwide
that run Google ads just go from not wanting to run misinformation to
being fine with it?
report detection, not consequences Those numbers on
bad ads are interesting, but pay attention to the text. These are
numbers for ads blocked or removed, and repeat
offenders drive the bulk of tech support scams via Google Ads. Does
an advertiser caught doing misrepresentation
in one ad get to
keep going with different ads?
don’t compare to last year, part 2 The previous two graphs showed Google’s bad ads/good site problem, so here’s how they’re doing on their good ad/bad site problem. Here’s 2024:
And 2023:
Ad-supported AI slop is on the way up everywhere, making problem pages easier to create at scale, but Google somehow caught 800 million fewer pages than in 2023. How many pages they took action against isn’t even a good metric (and I would be surprised if anyone is incentivized based on it). Some more useful numbers would be stuff like
What percentage of advertisers had their ad run on a page that later had action taken against it?
How much money was paid out to sites that were later removed for violating the law or Google policy?
But as in the previous graph, the big problem is in one of the
categories. Google caught fewer pages for malicious or unwanted
software
in 2024 than in 2023. Is there a good explanation for why
Google is taking less action on malicious or unwanted software
in
2024 than in 2023? As far as I know, nobody is claiming that developers
are writing less of this kind of software, or promoting it less. (icymi:
Researcher
uncovers dozens of sketchy Chrome extensions with 4 million
installs) Google management is just so mad about the union situation
that they’re willing to make the users suffer in order to keep
threatening workers with more layoffs. And did the amount of
dangerous or derogatory content
really go down from 104 million
pages to 24.8 million pages in a year? Or did something happen on the
Google side? (icymi: A
trillion-dollar problem: how a broken digital ad industry is fracturing
society – and how we can fix it)
A real Ad Safety Report
would help an advertiser answer
questions about how likely they are to sponsor illegal content when they
buy Google ads. And it would help a publisher understand how likely they
are to have an ad for malware show up on their pages. No help from this
report. Even though from the outside we can see that Google runs a bunch of
ads on copyright-infringing sites, not only does Google not report
the most meaningful numbers, they’re doing worse than before on the less
meaningful numbers they do choose to report.
Google employees, (yes, both FTEs
and TVCs
) are doing a
lot of good work trying to do the right thing on the whole ads/crime
problem, but management just isn’t staffing and funding the ad safety
stuff at the level it needs. A company with real competition would have
had to straighten this situation out by now, but that’s not
the case for Google. Google’s services like Search are both free and
overpriced—users don’t pay in money, but in over-exposure to fraud and
malware risks that would be lower in a competitive market. If a future
Google
breakup works, one of the best indicators of success will be more
meaningful, and more improved, metrics in future ad safety reports.
More: Another ad safety* report (2025)
just in case anyone wants to release better numbers: How to leak to a journalist by Laura Hazard Owen
Bonus links
Flaming
Fame. by George Tannenbaum. We don’t see shitty work and say
that’s shitty.
It’s worse than that. We simply don’t see it at
all.
LG
TVs’ integrated ads get more personal with tech that analyzes viewer
emotions by Scharon Harding. The company plans to incorporate a
partner company’s AI tech into its TV software in order to interpret
psychological factors impacting a viewer, such as personal interests,
personality traits, and lifestyle choices.
(What happens when you do
a Right to Know for the family TV?)
Former
Substack creators say they’re earning more on new platforms that offer
larger shares of subscription revenue by Alexander Lee. Since
leaving Substack, some writers’ subscriber counts have plateaued over
the past year, while others have risen — but in both cases, creators
said that their share of revenue has increased because Ghost and Beehiiv
charge creators flat monthly rates that scale based on their subscriber
counts, rather than Substack’s 10 percent cut of all transaction
fees.
The
Mediocrity of Modern Google by Om Malik. What’s particularly
ironic is that today’s Google has become exactly what its founders
warned against in their 1998 paper: an advertising company whose
business model fundamentally conflicts with serving users’
needs.
With
Support of Check My Ads Institute’s Advocacy, Senator Warner Urges FTC
and DOJ to Investigate Ad Fraud Affecting U.S. Government Agencies
Senator Warner’s letters to FTC Chairman Andrew Ferguson and DOJ
Attorney General Pam Bondi cite new research by cybersecurity and
digital forensics firm Adalytics, exposing how major adtech vendors have
failed to deliver the “real-time bot detection” that they promised. As a
result, advertisements intended for human audiences instead were shown,
for at least five years, to easily-identifiable bots operated from data
centers, including bots on industry group bot lists.
(Adalytics:
The Ad Industry’s Bot Problem Is Worse Than We Thought)
Git
turns 20: A Q&A with Linus Torvalds by Taylor Blau. So I was
like, okay, I’ll do something that works for me, and I won’t care about
anybody else. And really that showed in the first few months and
years—people were complaining that it was kind of hard to use, not
intuitive enough. And then something happened, like there was a switch
that was thrown.
I’m not an expert on electric cars, so I don’t know enough to criticize some of the hard parts of the design of a Tesla. But when they get obvious stuff like getting out without power wrong, that’s a pretty good sign to stay away.
How
the U.S. Became A Science Superpower by Steve Blank. Post war, it
meant Britain’s early lead was ephemeral while the U.S. built the
foundation for a science and technology innovation ecosystem that led
the world—until now.