- Google blocked or removed 8.3 billion ads and suspended 24.9 million accounts in 2025, and says over 99 percent of policy violating ads were stopped before they ran.
- The same report says incorrect advertiser suspensions fell by 80 percent, which is an admission that wrongly suspending real businesses was happening at a rate worth fixing.
- Under the EU's Digital Services Act, platforms must declare whether a moderation decision was automated. Google Shopping declares fully automated decisions and files more statements of reasons than any other platform.
- Your landing page is reviewed alongside the ad. A rejection often points at the destination rather than the words you wrote.
- Appeals have two distinct routes, disputing the decision and declaring a fix, and choosing the wrong one wastes the round.
You wrote three headlines, uploaded an image, set a budget of forty a day and went to bed. In the morning the campaign says Disapproved, the named policy violation is one you have never read, and there is no phone number. Somewhere in that loop a decision was made about your business in less than a second, by something that has never seen your shop.
Disapproved ads, and the account suspension that sometimes follows them, are now an ordinary condition of selling online, and it is worth understanding as machinery rather than as bad luck. The systems are documented, the volumes are published, and the appeal routes are real. What follows is what the platforms themselves say about how the decision gets made.
How much of this is automated?
Nearly all of it at the first pass. Google describes a review that starts automatically the moment you create or edit an ad, covering the headline, description, keywords, destination and any images or video, with most reviews finishing inside one business day.
The scale explains the automation. Google's 2025 report states it blocked or removed 8.3 billion ads and suspended 24.9 million accounts, with more than 99 percent of policy violating ads stopped before they ran. No human workforce reviews 8.3 billion anything. The report credits Gemini powered systems analysing signals including account age, behavioural cues and campaign patterns rather than matching keywords.
There is a line in that report worth reading twice, because it is the one that concerns you rather than the fraudsters. Google says incorrect advertiser suspensions were reduced by 80 percent through better separation of legitimate content from deceptive content. Reducing a number by 80 percent means it was previously five times higher, and that the people on the wrong end of it were running real businesses.
Do platforms have to admit when a machine decided?
In the European Union, yes, and the disclosures are public. The Digital Services Act requires platforms to file a statement of reasons for each moderation decision, and the database records whether automated means were used for detecting the content and for deciding it was infringing, with each declared as fully automated, partially automated or not automated.
That field turns a suspicion into a public fact, and the fact is stark for anyone selling products. An academic analysis of 131 million statements filed in a single ten day window in late 2023 found that Google Shopping alone accounted for 52.2 percent of all submissions, ahead of TikTok at 17.1 percent, Amazon at 10.8 percent and Facebook at 9.8 percent.
The same analysis grouped platforms by how they decide. Google Shopping, Google Maps, AliExpress and TikTok were reported as making fully automated decisions for most or all cases. Facebook, Instagram, Snapchat and YouTube automate some. X, LinkedIn and Booking.com were reported as not automating the decision at all.
Read as a merchant rather than as a policy researcher, that ranking is a map of where your listings face a machine and where they face a person, and it lines up with where sellers report the most baffling removals.
Why did an ad that ran for months suddenly stop?
Usually because something moved that was not the ad. This is the single most common confusion, and it sends people rewriting perfectly good headlines for a week.
The destination is reviewed alongside the creative, so a change on your own site can disapprove an untouched ad. A product page that now redirects, a policy page that disappeared during a redesign, a checkout that throws a certificate warning, a page that loads a different offer than the ad promised: all of these are landing page problems wearing an ad rejection's clothes.
Policies also change under running campaigns, and enforcement changes without the policy text moving at all. When a platform retrains the classifier that reads your category, the boundary shifts even though the written rule reads the same as last year. Google's own framing of moving toward intent based detection describes exactly that kind of shift.
The third cause is your own account history. When the signals include account age and behavioural patterns, a burst of edits, a new payment method or a sudden budget increase are inputs to a risk model, not neutral admin. Doing all three on the same afternoon is a worse idea than it looks.
Before you rewrite anything, open the ad's destination in a private browser window on a phone, from scratch. A surprising share of rejections resolve there: an expired banner, a broken redirect, a cookie wall the reviewer cannot pass, or a page that no longer mentions the product the ad is selling.
What is the right way to appeal?
Pick the correct reason, because the two options do different things. Google's help documentation on fixing a disapproved ad or appealing a policy decision offers two paths: dispute the decision, or declare that you made changes to comply.
Disputing says the classifier was wrong and nothing has changed. Declaring a fix says the classifier was right and the problem is gone. Choosing "made changes" when you changed nothing is a wasted round, and disputing when you did quietly edit the landing page invites a second look at an ad that would have passed.
Appeals can be submitted in batches across ad groups or campaigns, which matters when one policy call has knocked out forty variants. Fix the underlying cause once, then appeal the batch, rather than editing forty ads individually and restarting forty review clocks.
Timing has published thresholds, which saves a lot of refreshing. Google's guidance is that most reviews finish within one business day, that anything still under review after two full business days is worth checking on, and that a review running beyond a week is a reason to contact support. Those are the numbers to hold, rather than the folklore about resubmitting at particular hours.
One nuance in the transparency data is worth carrying into an appeal. The database's own definitions treat detection and decision as separate questions, and the automation field refers only to the decision taken after content was identified, not to how it was spotted. A partially automated decision means a model and a person both touched it. So the honest expectation when you appeal is not that a machine will be replaced by a human, but that a case which was decided at speed may be looked at by someone with more context, if it clears the bar for that.
Keep the appeal factual. The reviewer, human or otherwise, is matching your case against a policy definition, and a paragraph about how much you spend last quarter does not enter that comparison. What helps is naming the specific policy, stating what the ad and page now contain, and pointing at the evidence.
What if the appeal fails?
In Europe there is a further step almost no small seller uses, and it is written into the same law that produced the transparency database. Losing an in platform appeal is not the end of the road.
The Digital Services Act obliges platforms to run an internal complaint handling system and to tell you, when they reject your complaint, that out of court dispute settlement exists. The Commission's own explanation of certified out of court dispute settlement bodies sets out the mechanism: you may select any certified body to resolve a dispute about a platform's decision, including complaints the internal system did not resolve, and the body must decide within 90 calendar days, extendable by a further 90 in highly complex cases.
That is not a court and it does not bind the platform in the way a judgment would, but it is a route to a human outside the company that made the call, at a cost designed to be bearable. For a seller whose account suspension has removed their main channel, a documented independent process is worth more than another round of the same form.
Two practical notes. Keep every notice you receive, because a statement of reasons is the document that starts this process and platforms do not always resurface it later. And act on it promptly rather than after months of appeals, since the value of the route decays with the age of the dispute.
What makes it worse?
Three reflexes, all of them understandable, all of them costly. Knowing them is worth more than any appeal technique.
The first is opening a second account. After a suspension this feels like a workaround and it is treated as circumvention, which converts a recoverable policy violation into a permanent one and can pull your other properties down with it. Whatever the temptation, do not do this.
The second is deleting and recreating the rejected ad rather than appealing it. That destroys the record the appeal would have run against, restarts the clock, and teaches you nothing about the cause, so the replacement usually gets rejected too.
The third is changing many things at once. If you rewrite the copy, swap the image, move the landing page and raise the budget in one session, an approval tells you nothing about which change mattered and a rejection tells you even less. It is the same reason a creative test that moves several levers at once cannot be reused. Change one surface, resubmit, note the result.
Which categories get hit hardest?
The ones where a legitimate product shares vocabulary with a regulated or prohibited one, and the classifier cannot see the difference from text alone. This is where small honest sellers pick up rejections that feel personal.
| What you sell | What it collides with | The usual trigger |
|---|---|---|
| Supplements and food | Health claims rules | Wording that implies treating a condition |
| Skincare and cosmetics | Before and after imagery rules | Body imagery, or a promised result |
| Knives, tools, airguns | Weapons policy | The product image alone |
| CBD, vape hardware, alcohol accessories | Restricted goods | Category by default, often needing certification |
| Financial or coaching services | Financial products policy | Any figure that reads as a return |
| Second hand and refurbished | Counterfeit and misrepresentation | Brand names in the title without condition context |
If your catalogue sits in one of those rows, treat rejection as a running cost rather than an incident, and build the workaround into how you write listings. Certification exists for several of them, and completing it once converts a permanent low grade problem into an administrative step you do annually. Sellers in these categories who never certify tend to conclude the platform dislikes them, when what it dislikes is an uncertified listing in a gated category. The structured data that keeps a listing out of trouble is the same data that keeps a feed from bouncing, which we worked through in why two thousand rows produced one feed rejection.
What about the claims in the ad itself?
Claims are the part you fully control and the part most often written carelessly. A classifier trained on health, financial and performance claims is looking for a specific grammatical shape: a promise of an outcome, a comparison without a basis, a number without a source.
This has teeth beyond the ad platform. Consumer protection regulators enforce the same territory with rather more consequence than a disapproval, and the rules about what you may say tighten further when the claim is about artificial intelligence itself, which we set out in what you may claim about AI in your own marketing. An ad that passes review is not automatically an ad that is legal.
The practical drafting rule is boring and effective. Describe what the product is and what it contains, attribute any outcome to a named source, and let the customer draw the conclusion. Ads written that way get rejected far less often, and they survive the review that matters more.
Can you build a shop that gets rejected less?
To a useful degree, and most of the work happens before you write a single ad. The reviewable surface is bigger than the campaign, and the parts you own are the parts you can make boring.
Landing pages carry most of the avoidable risk. A destination that loads fast, works on a phone, shows the product the ad names, states the price and links to your policies removes an entire class of rejection reasons. Sites that get flagged tend to share the same handful of faults, and none of them are exotic. This is one reason we treat the storefront and its policy pages as one thing rather than two in the storefronts built with our ecommerce website builder: an ad reviewer, a shopping feed and an AI assistant all read the same pages.
The second lever is account hygiene. One business, one account, verified, with a payment method that matches the business name and a website whose registration matches too. Mismatches between these are exactly the signals a risk model weighs, and every one you remove is a false positive that does not happen.
The third is documentation. Certifications where your category requires them, and evidence for claims kept somewhere you can find it in five minutes rather than an afternoon.
What this means day to day
Rejections are not a judgement on your business and they are not a conversation. They are the output of a classifier operating at a volume where the error rate, however small in percentage terms, is enormous in absolute terms. Eight billion decisions with a one in a thousand error rate is eight million wrong calls, and one of them being yours says nothing about you.
What changes outcomes is treating the process as a process. Read the named policy rather than guessing. Check the destination before the copy. Choose the right appeal reason. Fix once and appeal in batches. Keep the account boring.
And keep a second route to your customers that no classifier gates. Sellers who took a hard automated hit and survived it were, almost without exception, the ones with an email list and a shop of their own. The platforms are a channel worth using and a poor thing to depend on entirely, which is the same argument that runs through what happens when you connect your ad account to a platform assistant. Owning the destination is the part nobody can disapprove, and it is the only asset in this whole system whose rules you write yourself.