This page discloses how AdWhiz measures “how much it saved you,” in full, so it can be reproduced and checked. It describes what the product actually computes, not marketing language.
1. What is measured
We do not take credit for “you spent less this year than last” — that could be seasonality, your own changes, or the market. We compare two randomly formed groups of your campaigns, across your connected ad accounts in one currency, over the same period: campaigns eligible for AdWhiz recommendations (treatment) and campaigns randomly held out that get no AdWhiz recommendations (control). The difference between them is the measured savings.
2. The control group
- Campaigns are the unit: about 15% of your campaigns are randomly assigned to the control group (not budgets, ad groups or audiences).
- Random and stable: the assignment comes from a hash of the campaign ID and your AdWhiz account; it is recorded, with its date, when the campaign is first seen and never changes.
- No AdWhiz recommendations: AI recommendations, budget-optimizer suggestions and recommendations returned by the API / MCP all leave control campaigns out. You can still change them yourself; that is part of normal operation in both groups.
- The one exception: if every campaign across all your connected accounts happens to be in the control group (no comparison is possible then), a one-off paid audit report covers them instead and says so.
3. Which campaigns are compared
- A campaign counts once a full reporting window has passed since its assignment (all of its data in the window is post-assignment).
- Only campaigns in the same currency and reporting window are compared; money is never converted or mixed. TikTok accounts count once the platform has confirmed their currency.
- Every campaign that spent in the window counts, including campaigns paused later (so a campaign paused mid-window is not dropped together with its spend). Campaigns deleted, removed or archived on the platform drop out of the dashboard comparison; the period reports used for the guarantee figure include them.
4. The measurement
Measured savings = (control cost per result − treatment cost per result) × treatment results, where each group’s cost per result = its total spend ÷ its total results. Equivalently, with k = control cost per result ÷ treatment cost per result, measured savings = treatment spend × (k − 1). The result can be negative, and a negative result is shown as such. A result is the conversion count the platform reports for the campaign — Google Ads “Conversions”; Meta purchases (the first of omni_purchase / purchase / website pixel purchase / on-site purchase the platform reports); TikTok “Conversions” — the same definition for both groups, on the dashboard and for the guarantee figure.
5. Uncertainty and the 95% range
- Test: the control group is re-drawn many times, the same way it was drawn (same number of control campaigns), to see how unusual the actual result is (a permutation test). With no real effect it calls a difference in at most about 5% of cases.
- 95% range: for each candidate effect k, treatment results are scaled back by k and the same test is run; the range is every k the test does not reject (test inversion).
- Best estimate: the k at which the test is exactly balanced; this also corrects the upward lean a ratio has when there are few control campaigns.
- On the dashboard: the headline shows a figure only when the test is significant and the range excludes zero — the end of the range nearest zero (“at least …”, or “a loss of at least …” when negative); otherwise the headline says “no measurable difference yet”. Once the data thresholds are met, the best estimate and the range are always shown beneath it.
6. Enough data
A number is shown only when all of these hold: at least 8 control campaigns with spend in the window; at least 30 results in each group; and results not concentrated in a few campaigns (at most 1% of re-draws leave the control group with no results, and the range can be closed). Until then the dashboard shows “building your control group” with progress toward each threshold, and no number.
7. The guarantee figure
- Computed once over the whole Measurement Period: after the Measurement Period ends, each campaign’s spend and results from the latest of its assignment date, 2026-10-16 (the day the control group began covering every recommendation surface) and the period start, until the period end, are measured as in Sections 4–6. The best estimate is the “Measured Savings” under the Guarantee Terms and is reported to you in writing together with its 95% range.
- Why not the daily dashboard figure: the dashboard updates every day, and looking many times gives chance differences more opportunities to appear; the guarantee uses only this single end-of-period computation.
- Data thresholds: if the thresholds in Section 6 are not met over the Measurement Period, the Guarantee’s eligibility conditions are not met (Terms, Section 3). The dashboard shows the same thresholds for its latest reporting window (usually 7 days) as a guide; the period totals are larger, so an account that meets them there will normally meet them for the period, and one that does not yet may still meet them over the whole period. The number of campaigns matters most: with about 55 campaigns the 15% control group reaches 8 campaigns only about 60% of the time; with about 70 or more it usually does.
- Where the period data comes from: each campaign’s spend and results for those dates come from the advertising platform’s own campaign-level reporting; you can export the same dates from your own ads manager to check them.
- More than one currency: money is never converted; the figure uses the currency group with the most campaigns that meets Section 6 (if none does, the largest), and campaigns in other currencies are left out, as on the dashboard.
- Non-USD accounts: Measured Savings are converted to USD at the U.S. Federal Reserve H.10 rate for the last business day on or before the last day of the Measurement Period, then compared with the subscription fees; for a currency H.10 does not publish, the European Central Bank reference rate for that day is used, or failing that the rate of the currency’s central bank.
- Excluded days: days excluded from the Measurement Period under Section 4 of the Terms (force majeure, material platform changes) are left out of the period data.
8. Two different savings figures
- Per-change savings (the dashboard’s “Estimated savings this month”): for each budget change actually applied (suggestions that were only recorded are not counted), the counterfactual result of the old budget at the ROAS predicted when the change was decided, with a 95% interval. This is a model estimate, not a control-group measurement, and it is not used for the guarantee.
- Measured savings (the control-group comparison): as described in Sections 4–7. Only this figure is used for the guarantee.
9. Data source & record
- All performance data comes from the advertising platforms’ own campaign-level reporting (Google / Meta / TikTok): the dashboard reads it through their official APIs, and the guarantee figure uses the same platforms’ campaign reports for the period’s dates (Section 7). Nothing is estimated.
- DEXUN computes Measured Savings in good faith from that objective data.
- You have a 14-day reconciliation right: if you in good faith believe there is a manifest error or technical malfunction, you may dispute in writing within 14 days of the figure being reported, and we will review and correct it if confirmed (see ROI Savings Guarantee Terms, Section 5).
10. What it is not
- This methodology measures cost saved — it is not a guarantee of profit, revenue, or ROAS.
- Platform algorithm or policy changes, account suspensions and seasonal swings affect both groups and cancel out in the randomized comparison rather than being deducted separately.
11. Where the method comes from
Holding out a random control group to measure the increment is the same family of method as Meta’s, Google’s and TikTok’s Conversion Lift studies; the difference is that we randomize campaigns, not people.
12. Versioning policy
This methodology is versioned. The version applicable to you is the version published on this page at the start of your Measurement Period; we will not change it to your detriment during your Measurement Period, and any change is prospective. For a Measurement Period that started before 2026-10-16, you may choose in your claim to have it determined under v1.0 or under this version; without a choice, v1.0 applies. Under this version, Measured Savings for such a period are computed from 2026-10-16, and Section 2 of the Guarantee Terms (including the refund amount and its cap) applies with the subscription fees pro-rated by days to that part of the period.