Search the guides

Filtering & analytics

Bots in paid traffic: find the loss before adding a filter

A visit without a purchase is not proof of a bot. Separate automated requests, weak traffic and broken measurement with a practical diagnostic sequence.

Different counters describe different events

The ad account shows clicks, the landing page shows fewer visits and enquiries are scarce. Do not call the gap bot traffic yet. Find the missing step: did the page arrive, did a visitor begin the form and did the business receive a valid enquiry?

A bot is software making requests automatically. Some requests are useful, such as monitoring and link previews; others can produce spam or consume resources. A request in a server log is therefore not automatically a paid human visit or a fraudulent advertising click.

Google describes invalid clicks as interactions without genuine business interest and explains its own filtering and billing treatment. That is separate from a rule on your landing page. A local block does not itself establish that a billed click was invalid or that its cost has been recovered.

Follow a discrepancy through the funnel

Take an illustrative funnel: 1,000 reported clicks, 780 page loads, 90 form starts and 20 accepted enquiries. These are teaching numbers, not a benchmark. Investigate the first gap as a delivery question; investigate later losses as possible form, offer or audience problems.

Event flowLocate the first meaningful loss
  1. Ad clickPlatform interaction
  2. Page deliveryA usable landing page appears
  3. IntentVisitor starts the next step
  4. OutcomeA valid enquiry or purchase is recorded

Investigate the first unexplained gap, then move to the next stage.

Before comparing totals, match the period and time zone. Check whether an order is grouped by its click date or payment date and whether repeated requests count separately. Write each column’s definition beside the number. Otherwise a calculation can turn incompatible measurements into a false loss rate.

Separate “what reached our site?” from “what did the advertising platform bill?” A local log can help with the first question. It cannot answer the second by itself. For example, a useful link preview can reach the page without being a billed advertising click.

Look for combinations, not a single giveaway

A short visit is weak evidence. Repeated submissions with identical impossible field values give you a more specific pattern to investigate. Check the sequence and endpoint—the address receiving an action such as a form submission—then look for an ordinary explanation.

Compare a limited suspicious sample with visits that produced confirmed results. Keep only relevant fields and restrict access. Do not copy customer form contents into a general team sheet just to make the analysis convenient. The goal is to understand the pattern, not collect everything.

  • Separate page requests from static resources and health checks.
  • Record the exact rule or signal behind a suspicious label.
  • Compare time, campaign, endpoint and completion behaviour.
  • Check whether verified useful automation should remain accessible.
  • Look for the same pattern in accepted orders before broad blocking.
Signal layersDo these clues point to the same cause?
  • Request pattern

    Repeated actions with a regular sequence.

  • Form content

    The same invalid values recur.

  • Journey

    Actions do not match ordinary page use.

  • Countercheck

    Could monitoring or a test explain it?

Combine the pattern with a check for useful automation or an internal test.

Run a controlled change

Protect the action where the harm occurs. For repeated junk enquiries, start at the form. For excessive requests to one address, assess a rate limit: a rule restricting how frequently an action can occur. Do not ban a broad class of devices merely because that control is available.

Track accepted enquiries and failed legitimate attempts alongside blocked requests. A rising block percentage can mean stronger protection or worse selectivity. Define a rollback condition for new failures on a previously working browser or route, and keep a record of the change.

Do not let a traffic mix change impersonate an attack

A new placement or creative can bring a different audience. Lower conversion then does not necessarily mean more automation. Compare relevant groups before interpreting the aggregate change, and do not treat a few visits as a dependable rate.

Imagine moving from a specific search query to a broad mobile placement. The new audience may be curious but not ready to enquire. This is a hypothetical example. Clearer qualification in the ad or a better page explanation may help; a network filter cannot create missing intent.

Look at your own release history too. A new required field, changed consent behaviour or slow asset can reduce recorded activity during the same period. Include these possibilities in the investigation instead of assigning every unexplained change to incoming traffic.

If you contact the advertising platform, provide the period, campaign and concise pattern evidence. Do not label every unconverted click fraudulent. A precise question helps distinguish a billing concern from a broken landing-page journey.

Match the remedy to the finding

Choose the next task from the finding. Fix delivery when the page fails to open, reconsider the message when people misunderstand the offer, and apply a narrow protective rule when harmful automation repeats. If the evidence cannot distinguish those cases, keep investigating before buying another tool.

Decision pathChoose the remedy after locating the loss
Where does the journey fail?

Use delivery and outcome evidence.

  • Page does not load

    Fix delivery, redirects or access.

  • People leave

    Review the audience and offer.

  • Abuse repeats

    Apply a narrow protective rule.

The same conversion drop can require three different responses.

Save a short incident note: symptom, evidence, alternative explanations, change and result. The next operator can then recognise a recurrence or avoid repeating an old measurement mistake. The same note improves a support conversation.

Use traffic labels to explain a demonstrated problem. Start with delivery and accepted outcomes, then examine automation where it remains a plausible cause. A useful filter reduces harmful activity without quietly removing the customers the campaign was meant to reach.

Questions answered

Does zero engagement prove a visit was automated?

No. Fast exits can result from slow loading, a wrong offer or measurement limits. Combine evidence from the request and the actual visitor journey.

Will blocking a bot refund the original click?

A local block and an advertising platform’s billing decision are separate. Use the platform’s invalid-traffic process for billing questions.

ADSTRAFFIC / ENAll guides →