A savings report needs to distinguish suspicious activity observed, charges actually incurred, credits received and future spend plausibly prevented. Multiplying flagged visits by average CPC does not establish money saved.
A report can be useful while some of those figures remain unknown. Its job is to make the evidence and assumptions visible enough that you can decide what to do next.
Start with the unit being counted
Requests, page loads, sessions, click identifiers and billed ad clicks are different units. One visitor can load several pages after one ad click. A scraper can generate many requests without arriving through an ad at all.
Ask for the eligible paid-click denominator, deduplication method, attribution coverage, time zone and observation window. Keep unattributed activity visible rather than silently assigning it to a campaign.
Then inspect the reasons for the suspicious label. An unexplained score or a failure to convert is insufficient evidence of billed fraud.
Reconcile what the platform already removed
Google explains that invalid traffic detected before billing is removed from charges, with credits possible for later detection. It also warns that third-party flags can overlap traffic already filtered. A visit remaining in website analytics therefore does not establish that the advertiser paid for it.
Later adjustments can change the financial picture. Google's Invalid Activity Credit Report includes credited amounts and adjusted campaign costs.
Campaign totals can support reconciliation, but they do not necessarily identify the billing status of each click your detector flagged. Preserve that uncertainty instead of assigning every flagged click a charge.
A worked example: exposure is not savings
Suppose a fictional report flags 100 identified clicks and uses a campaign average CPC of $10. The resulting $1,000 is an illustrative exposure estimate under the assumption that those clicks cost the average amount.
It is not proof of $1,000 saved. The clicks already occurred; their individual costs and platform treatment remain unresolved. Nor does the multiplication say how many later clicks an intervention prevented.
| Report label | Evidence required |
|---|---|
| Observed suspicious activity | Defined cohort, deduplication and inspectable detection evidence |
| Estimated cost exposure | Explicit costing assumptions and unresolved billing overlap |
| Confirmed credit received | Platform billing record for the relevant account and period |
| Estimated prevented spend | A defensible comparison estimating what would otherwise have happened |
| Improved acquisition performance | Comparable mature outcomes, spend and intervention history |
Even a confirmed credit does not by itself prove that a particular tool caused it. The platform may have issued it independently.
How to test incremental benefit
A useful evaluation declares its comparison before examining the result: eligible campaigns, assignment method, action, outcome horizon and stopping conditions.
Track the intervention version and actual application alongside spend and independently confirmed outcomes. Review competing explanations such as changed bids, targeting, landing pages, seasonal demand and conversion tracking. Shared account learning can also complicate comparisons between campaigns.
The test should be capable of finding no benefit or harm. More blocked addresses or fewer recorded conversions are not sufficient success criteria. For a lead business, compare qualified outcomes after a suitable follow-up period; for a shop, distinguish completed orders from success-page observations.
What to request from your provider
Ask for one worked example from observation through action to financial interpretation, with missing evidence labelled. Request the calculation assumptions, later-credit handling and method for checking legitimate-customer impact. Apply the same standard to Button and to any service bundled with your agency.
For related checks, read what IP exclusions can prevent, how conversions differ from lead quality and which campaigns need investigation.
AI disclosure
ChatGPT (AI) drafted this article using the linked sources. Prepared for editorial review by Nathan Pankowsky. The dollar example is hypothetical; no Button customer savings or causal performance result is claimed.