Why Your AdSense Account Got Banned — And Why Appeals Almost Never Work

The email arrives on a Tuesday. The subject line is bureaucratic and slightly bloodless, and the body explains that your account has been disabled for invalid traffic activity. There is a link to an appeal form. There is no explanation of what was found, when, or on which pages.

For a small publisher this is not a policy notice. It is the end of the business, arriving without notice and without detail.

The figure most often cited by practitioners for appeal success on invalid-traffic cases is under 10%, and Google's own documentation states that permanently disabled accounts are nearly impossible to recover — publishers in that position are also not permitted to participate in AdSense again. That combination is closer to a professional death sentence than to a timeout, and it is worth being clear-eyed about it before spending three weeks writing appeals.

This article is about what actually triggers enforcement, how the escalation works, and what reduces the risk. Not about how to write a better appeal — by the time you are writing one, the useful decisions are behind you.

One framing note before the mechanics. It is tempting to read enforcement as a moral judgement — Google decided you were a cheat. It is more accurate, and more useful, to read it as a risk decision made at enormous scale with incomplete information. Google is protecting advertiser budgets across millions of accounts using automated classification, and it errs toward removing uncertainty rather than resolving it. That is not a defence of the process. It is an explanation of why an appeal that argues about your character tends to go nowhere, and why one that changes the classification's inputs occasionally does.

It is three stages, not one event

Most publishers learn the enforcement ladder in the worst possible way: by arriving at the bottom rung and discovering there were two above it that they did not recognise as warnings.

Stage one — the ad serving limit

This is the one people misread most often. An ad serving limit is applied when Google cannot verify the quality of traffic on your account. Note the wording: not "has determined your traffic is fraudulent", but cannot verify it. Ads are throttled or stopped, revenue drops sharply, and the account otherwise functions normally.

Limits are frequently temporary and often lift on their own within about thirty days once the underlying traffic pattern normalises. This is the stage at which intervention actually works, and it is also the stage most publishers spend chasing support rather than auditing their traffic sources.

The productive response to a limit is diagnostic, not communicative. Pull the last sixty days of traffic by source, country and landing page. Find what changed and when. If something started — a new referral partner, a campaign, a syndication deal, a sudden burst from one country — that onset date is almost always the answer, and stopping it is almost always the fix. Support cannot tell you what was found; your own logs can tell you what changed, and those are the same question from different sides.

Stage two — suspension

A suspension is roughly a thirty-day timeout. Ad serving stops; you can still sign in and see your account. If nothing further is found, it clears automatically at the end of the period.

The trap in this stage is not widely understood: if additional issues are identified during the suspension window, the account can be permanently closed before that window expires. A suspension is not a safe holding pattern where you can carry on and wait it out. It is a period of heightened scrutiny, and behaving as though it is business as usual is how a recoverable situation becomes a terminal one.

Stage three — disabled

Permanent. No ads, and no participation in Google's publisher advertising products afterwards. This is the stage the opening figures apply to, and the honest summary is that the appeal is a formality far more often than it is a remedy.

What an appeal can and cannot do

Since the numbers at the top are bleak, it is worth saying precisely where the remaining percentage comes from. Appeals that succeed tend to share a shape: the publisher identifies a specific cause, can show when it started, has already removed it, and describes the change in concrete terms. Appeals that fail tend to assert good faith, express confusion, or ask what was found.

That asymmetry follows from what the reviewer is being asked to decide. They cannot verify your intentions and will not disclose their evidence. The only question they can act on is whether the condition that produced the classification still exists. An appeal is therefore a change report, not a defence — and a change report is impossible to write if you never established what changed.

Where the money goes when this happens

Publishers are often surprised that revenue they have already earned — sometimes already been paid — can be taken back. It can, and the mechanics are published.

Invalid traffic is removed from earnings at three different points. Some is filtered in real time and never appears in your reporting at all. Some is removed during finalisation, which is why a month's estimated earnings can drop when it is finalised. And some is deducted up to sixty days after payment, appearing as a negative adjustment on a later invoice.

This is why a month can look healthy in your dashboard and then quietly shrink. The reporting you watch daily is an estimate, and the estimate is generous by construction — it counts everything that has not yet been disqualified. Finalisation is where the disqualification happens.

On a disabled account, a payment hold of around thirty days applies while the final review runs. Deductions for invalid traffic are refunded to the advertisers whose budgets bought the invalid impressions — which is the part that makes the whole mechanism coherent rather than punitive. The money goes back to the party that paid for something they did not receive.

What Google is actually looking for

The single most common misconception is that enforcement is driven by traffic volume — that a big spike gets you flagged. Volume is a weak signal on its own. Sites go viral. Stories get picked up. Google has seen a great deal of legitimate traffic growth.

What draws attention is a mismatch between metrics that should move together.

The logic behind mismatch detection is worth understanding, because it tells you what to watch. A real human who clicks an ad usually did some other things first: they scrolled, they spent time, they arrived from somewhere plausible, they may have visited a second page. Those behaviours are expensive for a bot to fake convincingly and cheap for a human to produce naturally. So the useful question is not "did clicks rise" but "did the things that normally accompany clicks rise with them". When clicks move and nothing else does, the most economical explanation is that something is producing clicks without producing readers.

That fourth item is where a surprising number of well-intentioned publishers get caught, so it is worth being specific about layout.

The placement problems that generate real clicks nobody meant

• Ads sitting immediately beside or beneath navigation, where a mis-tap is likely.
• Sticky units that overlap content or move under a thumb mid-scroll.
• Ad labels styled to resemble menu items or article links.
• Units placed so close to buttons that the two targets overlap on a small screen.

The commonly cited working rule is to keep ads at least around 150 pixels away from buttons and interactive elements. It is a guideline rather than a published hard threshold, but it encodes the right instinct: if a reasonable person could click your ad while trying to do something else, the resulting clicks are invalid regardless of what you intended.

Two kinds of banned publisher

In practice the population splits cleanly, and the split matters because only one half is a policy problem.

The first group sends invalid traffic unintentionally. They tested their own live ads to check a layout. They put a unit somewhere that generates accidental taps. They bought traffic from a vendor whose quality they had no way to inspect. They ran a campaign that brought in a burst of low-engagement visitors from a single source. None of this involves an intention to defraud anybody, and most small-publisher enforcement lands here.

The second group does it deliberately to inflate revenue. Click exchanges, bought bot traffic, incentivised clicking.

The distinction matters commercially as well as morally. The first group's problem is fixable and usually cheap to fix — move two ad units, stop testing on live inventory, drop one traffic vendor. The second group's problem is their business model. If you are in the first group, the single most valuable thing you can do is find out before an algorithm does, because everything that group does wrong is visible in their own analytics if anyone is looking.

The enforcement system does not, and largely cannot, distinguish between the two. It observes traffic patterns, not motives. This is worth internalising because it reframes the whole problem: your good intentions are not a control. The only thing that protects an account is what the traffic actually looks like.

The part nobody warns you about: it can be done to you

Because enforcement keys off traffic patterns rather than intent, invalid traffic can be weaponised against a publisher who did nothing at all.

The attack is straightforward: point bot traffic at a competitor's ad units until the pattern trips enforcement. There are documented cases of extortion built on exactly this — an email demanding payment in cryptocurrency, with the threat of flooding the target's ads until the account is suspended. The technique needs no access to the victim's site and costs very little.

It is worth being calm about how common this is. Most suspensions are not attacks — the boring explanations, a bad traffic vendor or an accidental-click layout, are far more likely, and reaching for sabotage as the first hypothesis is usually how a publisher avoids finding their real problem. But it happens often enough that it belongs in your threat model, particularly if you are in a competitive niche or have recently annoyed somebody.

There is no clean defence, which is an unsatisfying thing to write. What there is:

• Detection. Knowing it is happening while it happens, rather than learning it from an enforcement email three weeks later.
• An evidence trail. Contemporaneous logs showing the anomaly's onset, its source composition and how it differs from your baseline. An appeal that says "this was not us" is worth little. One that shows exactly when foreign traffic arrived, from where, and how it diverged from twelve months of normal is at least an argument.
• Speed. The window in which anything can be done is the ad-serving-limit stage, and it is short.

The MCM angle

If your inventory is managed through Multiple Customer Management — a network representing you in Google Ad Manager — the enforcement picture changes shape. Traffic quality becomes a shared concern: problems on one represented site can draw scrutiny that affects the relationship, and a network with any competence screens for that before it becomes everyone's problem.

For a publisher this cuts both ways, and it is worth being honest about both. You gain a party with more data, stronger monitoring and a direct commercial interest in your traffic staying clean. You also gain a party whose other partners' behaviour is not under your control. Ask a prospective network what they screen for, what they will tell you, and how quickly — the answer is a reasonable proxy for how seriously they take the rest of the job.

Flying blind, by design

Google's invalid-traffic filtration is deliberately opaque, and the reasoning is defensible: the system never signals to the user — or to a bot — that traffic has been flagged, because telling an adversary what you detected is how you stop being able to detect it. A transparent filter is a filter with a published test suite.

The consequence for honest publishers is nonetheless real. You cannot see what was filtered, or why, or which pages it came from. You get an outcome without an explanation.

That opacity has also increased over time. Google deprecated the network-domain and service provider dimensions in Google Analytics — fields publishers had used to spot suspicious ISP concentrations in their own traffic — and did not offer a replacement. A diagnostic tool that existed was removed, and nothing took its place.

The practical position, then: you are accountable for traffic quality, you have limited native visibility into it, and the feedback arrives as a verdict rather than as a warning.

What actually reduces the risk

None of this is exotic, and all of it is cheaper than an appeal.

Think of it as the ad-ops equivalent of bookkeeping: unglamorous, mostly preventive, and conspicuously valuable exactly once, at the worst possible moment.

Know your own baseline

You cannot detect an anomaly without knowing what normal looks like. Write down your ordinary ranges — CTR by placement, sessions per country, average session depth, referral composition — and check them weekly. Most publishers discover their baseline for the first time while trying to explain a deviation from it.

Watch CTR against behaviour, not alone

A CTR rise on its own is not a signal. A CTR rise with flat scroll depth, flat time-on-page and flat pages-per-session is the specific pattern that draws attention. Track them as a set.

Treat sharp growth as a thing to verify, not celebrate

A traffic jump is the most pleasant possible warning sign, which is exactly why it gets waved through. When volume rises sharply, check that the accompanying behaviour rose with it — session depth, returning visitors, pages per session, geographic spread. Growth that arrives with its behaviour intact is growth. Growth that arrives as raw sessions with nothing behind them is a problem wearing a nice hat.

Vet traffic sources like a buyer, not a hopeful

If you buy traffic, ask where it originates, insist on source-level reporting, and start small enough that a bad vendor is an expense rather than an extinction event. "Guaranteed real users" is a sentence, not a control.

Never test on live ads

Do not click your own ads to check a layout. Not once. Use Google's own preview tooling, or a test property. This causes a meaningful share of accidental self-inflicted enforcement, and it is entirely avoidable.

Keep the trail before you need it

Whatever you use — your analytics, your ad server's reporting, a monitoring layer — make sure it retains enough history to describe a normal year, and make sure you can export it. An appeal written from memory is an assertion. An appeal accompanied by twelve months of source composition, with the anomaly visible as a departure from it, is evidence. The time to arrange that is while nothing is wrong.

Audit placements for accidental clicks

Open your own site on a mid-range Android phone — the device most of your audience actually uses — and try to navigate it normally. Anywhere your thumb nearly hits an ad while reaching for something else is a placement to move.

Where a network layer helps

Everything above is work a publisher can do alone, and should. The limit is structural: a single site does not have enough data to distinguish "unusual for us" from "unusual anywhere", and the signal that matters most — behavioural mismatch — is exactly the kind that needs a baseline broader than one property to read confidently.

That gap is the reason traffic monitoring exists at network level in BeeOS. BeeGuard screens traffic quality across properties and watches for the behavioural-mismatch pattern this article describes, so the signal surfaces while it is still at the recoverable stage rather than after a platform has acted. BeeScope keeps the reporting trail — what your traffic looked like, day by day, before and during an anomaly — which is the difference between an appeal that asserts and an appeal that demonstrates.

Neither is a guarantee, and anyone who offers you one on this subject is selling something. The enforcement decision belongs to the platform, the criteria are deliberately undisclosed, and no third party controls the outcome. What is available is earlier warning and better evidence — and given that the recoverable window is measured in weeks, earlier warning is most of what there is to win.