Click fraud or a bad setup: a 10-step diagnosis
Contents
Share
A sharp rise in clicks in Direct is possible even when there is no invalid traffic. Real users with no interest in the offer are not invalid traffic, and Direct’s filter does not remove them. A strategy needs at least 10 achievements of a goal per week before it can learn on that goal. Until those three boundaries are laid out, an “ad diagnosis” mixes click fraud, the wrong demand and a hungry strategy.
Signs of click fraud stay a separate review. This is an order of ten steps that says where to stop and which action the help text actually offers. Figures below that are not from the help text are marked as an example.
What the diagnosis is allowed to decide
The statistics help explicitly allows a rise in clicks with no invalid traffic. The first chart, “clicks up”, is not a conclusion. It is only a reason to walk the steps.
Direct’s filter does not take live people who are not interested in the offer. Their clicks stay in spend. Treating them as fraud edits the wrong layer: negative keywords, the offer and the landing page belong to interest, not to recognising traffic as invalid.
Strategy learning is the third layer. Fewer than 10 goal achievements in a week is a reason to take a goal higher in the funnel, change the budget or the conversion price, not a reason to call the week fraud. Ten achievements are a learning threshold, not a sign of fraud. Separately, check the goal by its own conversion and see whether Metrica recorded it.
Ten steps
Write down the period and the campaign. A single day with a spike and the strategy’s calendar week are different windows. The strategy is judged on a week of achievements. The help text allows a click spike with no invalid traffic.
Separate “there are more clicks” from “the traffic was recognised as invalid”. Clicks can rise in both cases. Until there is a recognition, a spike remains a spike.
Note who among these clicks looks like a live person with no interest in the offer. The filter does not remove them. If the hypothesis “they do not need what we sell” holds, the next move is demand setup, not a fraud complaint.
Open the signs you already know how to read: how to find bots in Yandex Metrica and how to look for them in Webvisor. Their internal methods are not copied into this diagnosis. You bring back only a label: the visit looks like a robot or like a person with no interest. The filter does not remove the second case.
If the campaign runs on networks, read placements in their own report: suspicious YAN placements. A placement with spend is not the same as invalid traffic. Clicks from a placement can also rise without it.
Count achievements of the chosen goal from Monday through Sunday. Learning needs at least 10 in the week. Fewer, and the strategy is not in the mode the help text calls learning.
If achievements are under ten, the help text suggests one of three moves: a goal higher in the funnel, a different budget, or a different conversion price. Pick one. Do not change all of them at once, or the next week will not show which lever moved the counter.
Check the goal by its own conversion and whether Metrica recorded the achievement. A strategy will not learn on a goal the counter did not write down. That is a setup check. It does not turn missing achievements into fraud.
Match leads to clicks separately: clicks without leads. An empty lead when the users are live and uninterested is a gap the filter is expected to leave: their clicks stay in spend. An empty lead when Metrica did not record the goal is a setup gap.
Write the conclusion in one line: “no invalid traffic is visible in this spike”, “live people with no interest are visible”, or “fewer than 10 achievements this week, the strategy is not trained”. The three conclusions can sit together. One cannot replace another.
Stop at the first step that already explains the spend. Finishing the rest is useful, but the decision belongs to the layer that matched.
How not to mix the three answers
| What you saw | Which layer | What the help text supports |
|---|---|---|
| Clicks jumped, and the traffic was not recognised as invalid | An allowed spike | Do not treat the spike as proof of fraud |
| Live people with no interest in the offer | Not Direct’s filter | Do not expect them to be removed as invalid |
| Fewer than 10 goal achievements in the week | Strategy learning | A higher goal, the budget, or the conversion price |
| The goal is missing in Metrica, or its own conversion is empty | Goal setup | Accounting first, then learning |
| Leads are missing while clicks are there | A separate funnel review | Do not glue this to the spike until steps 2–3 are done |
The table adds no new thresholds. The ten in it is only the learning threshold from the recommendations help.
A week where bots are not the cause
A teaching example.
The campaign got noticeably more clicks this week than last week. There is no report that recognised the increase as invalid. Step 2 closes like this: the help text allows a rise in clicks with no invalid traffic.
Visit recordings show people reading the offer and leaving: the price does not fit. These are real users with no interest. The filter does not take them. Spend on them remains, and that is not a hole in antifraud.
Achievements of the “order” goal this week: 4. The strategy does not reach ten. The help text suggests a goal higher in the funnel. In the example, “product card view” replaces “order”, and the budget is not touched the same week. The new goal’s own conversion is checked in Metrica: achievements are there, so the strategy at least sees a signal.
Leads stay scarce. That agrees with step 3. It does not cancel it. Clicks without leads here follow from missing interest, plus an “order” goal that did not reach ten. The example does not change both layers on the same day: otherwise it is unclear what brought the ten back.
Next week, count achievements again. If they are still under 10, return to step 7 and change the budget or the conversion price — one lever, not both.
Checklist before the sentence “we are being click-fraudded”
The spike period and the strategy week are written down separately.
The rise in clicks is not called invalid traffic only because it rose.
Live uninterested users are separated: the filter does not remove them.
The Metrica and Webvisor conclusion is brought back as “robot or person”, without retelling another article.
Network placements, if any, were read apart from the campaign spike.
Goal achievements for the calendar week are counted.
If the count is under 10, one lever is chosen: a higher goal, the budget, or the conversion price.
The goal’s own conversion and Metrica’s accounting are checked.
Empty leads are matched to interest and to goal accounting, not to the single word “bots”.
The outcome is a line you can check next week.
If step 7 is empty while there are four achievements, the click-fraud conversation is early: the help text still offers a fix for learning. If step 3 is filled with people who have no interest, a conversation about refunding those clicks is early too: the filter does not treat them as invalid.
Where the diagnosis ends
After ten steps a spike may remain that is explained neither by live indifference nor by a hungry strategy. That is the entrance to a review of signs and placements, not a reason to invent a fourth cause inside this article. The causes here are only these: a rise with no invalid traffic, live uninterested users, or fewer than ten achievements and an unchecked goal.
Repeat the diagnosis once a calendar week, because the ten achievements are weekly too. A Monday click spike does not close learning and does not prove fraud. Wait for Sunday, recount achievements, and only then decide which lever from step 7 is still unused.
What to write down at each step
The diagnosis falls apart when the note only says “looks like bots”. Each step needs one line you can check a week later.
Step 1 looks like this: “spike — 12 March, strategy week — 9–15 March”. A day and a week do not replace each other. Ten achievements are counted for the week, and the help text discusses a click spike as a statistics event, with no required link to invalid traffic.
Step 2 names whether the traffic has already been recognised as invalid. If it has not, write that. A rise in clicks with an empty line remains a case the help text considers possible.
Step 3 describes interest, not “traffic quality” in general. “They read the price and left” is live people with no interest, and the filter does not take them. “They did not read, and the visit broke off in an identical batch” is not yet a conclusion of this article: that pattern goes to the Metrica and Webvisor review and comes back only with the label “robot” or “person”.
Step 4 does not copy another article into the note. “The visit looks like a person in Metrica” or “it looks like a robot” is enough. The second label does not cancel step 2: this article does not automatically glue a robot and Direct’s invalid traffic together. The first label, if there is no interest, stays in spend.
Step 5 is the placement name and whether you read it separately. Placement spend without its own report does not become invalid traffic just because clicks rose.
Step 6 is the number of goal achievements in the calendar week. 4, 9 and 10 read differently only against the learning threshold. 9 is still under ten. 10 is the minimum the help text names for learning, not a profitability score.
Step 7 is one chosen lever. “Moved the goal from order to card” is different from “changed everything”. Budget and conversion price wait for the next week if they were not the lever you chose.
Step 8 is whether Metrica recorded the achievement and what the goal’s own conversion is. Empty accounting closes the learning conversation: the strategy has nothing on which to hold a ten, even if clicks are there.
Step 9 ties empty leads to step 3 or step 8. “No leads because people do not need the offer” is a complete sentence. “No leads because the goal is not being written” is a different complete sentence. “No leads, therefore bots” is a sentence these steps do not confirm on their own.
Step 10 can record the three allowed outcomes together. An example line: “a spike with no recognition of invalid traffic; live uninterested users stay in spend; 4 achievements, next week we take a goal higher in the funnel”. Seven days later that line is checked again: are achievements at least 10, and has the click spike been replaced by a fraud story.
The second week of the same example
A continuation of the teaching example, not a norm.
In the week of 16–22 March clicks are again higher than a neighbouring campaign, and there is still no recognition of invalid traffic. Step 2 repeats with the same answer. Live bounces because of price are still there; the filter did not remove them.
The “card” goal produced 12 achievements. The learning threshold is passed. Budget and conversion price were not touched this week, so the ten can be tied to the goal change, not to three levers at once. Orders stay scarce: interest did not appear because the strategy saw a card. That is an honest limit of step 3. The next conversation is about the offer, not about changing the goal again the same week.
If the card had produced 6 achievements, the example would return to step 7 and choose the budget or the conversion price. It would not change the goal again the same week: otherwise the line from step 10 can no longer be checked.
The ten steps become a cycle, not a form for one evening. The click spike lives in statistics. The ten achievements live in the week. Live people with no interest live in spend until the offer fits them. Click fraud in this cycle occupies only the line where traffic has already been recognised as invalid, or the line you reached from a separate review of signs — and you brought a conclusion, not a guess from the chart.