Beta feature — Contact your Customer Success Manager for access.
TL;DR
Impact Analysis tells you whether the changes you make to campaigns, adsets, or adgroups actually drove the performance you're seeing — or whether it would have happened anyway. It works on your historical data and returns a concrete absolute number ("this change added 1,466 extra conversions").
What problem does it solve?
The hardest question in performance marketing: "Metrics moved after my change — was it me, or would this have happened anyway?"
You raise your Brand Search daily budget from 7,000₺ to 21,000₺. A week later, conversions are up 148%. Two questions sit underneath:
What happened? → The before/after numbers tell you that.
Did your change actually cause it? → Much harder. Seasonality, demand shifts, organic momentum could all be in play.
Impact Analysis answers question 2 with statistical rigor — and reports the result as a concrete number ("+1,466 conversions over 7 days") rather than a percentage you have to mentally translate.
What kinds of changes can you analyze?
Available now
Type | Entity | Platform |
Campaign Budget Change | Campaign | Google Ads, Meta Ads |
Adset Budget Change | Adset | Meta Ads |
Campaign Bidding Strategy Change | Campaign | Google Ads |
Adset Bidding Strategy Change | Adset | Meta Ads |
Change Impact (Fixed Date) | Workspace Level | Any — you supply the date |
Coming soon: New Campaign Launch · Campaign Pause · Creative Change · Custom
About Change Impact (Fixed Date): Use this when the change isn't auto-detectable from platform metadata — landing page swaps, audience exclusions, tracking changes, or external events (competitor launches, market shifts). You provide the intervention date and the model measures from there.
How the methodology works
Impact Analysis builds a counterfactual model: an estimate of what your entity's performance would have looked like if you had not made the change. The gap between that projection and reality is the causal impact.
In the detail panel:
Without The Change = the counterfactual
What Happened = the actual
Your Impact = the difference
Covariates
A change isn't the only thing happening. To separate signal from noise, the model uses covariates — additional metrics that help it understand what was happening around the change. Defaults are Clicks and Cost; you can add up to 5.
If unsure, leave the defaults. Add Impressions if you suspect auction-side fluctuations. Don't pick redundant metrics (e.g. both Clicks and CTR). 2–3 well-chosen covariates beat 5 noisy ones.
Verified? status
Status | Meaning |
🟢 Real Impact | The change drove a measurable effect. |
⚪ Not Clear Yet | The model can't isolate your change from noise yet. |
Setup
Basic settings
Which campaigns / adsets / adgroups should be included? You don't pick entities by name — you define a filter, and the system pulls every change inside that filter automatically. The entity level depends on the change type.
When did it happen? Set a Start and End Date for the period you want to scan for changes. This is "when did the change happen?" — not the model's training window (those are configured separately under Advanced).
What do you want to measure? Pick one target metric. Default is Conversions. Other typical choices: Conv. Value, ROAS, CPA, Clicks. Fixed at run time — re-run with a different target if needed.
What else might affect results? (Covariates) Up to 5 metrics. Defaults: Clicks and Cost.
Advanced settings
These have sensible defaults — only touch them if you know why.
Group | Setting | Default |
Analysis Window | Days Before Change (baseline window) | 60 |
| Days After Change (observation window) | 7 |
Minimum Data | Min. total clicks (entity excluded if below) | 200 |
Eligibility | Min. Days Since Change ✅ Editable | 4 |
| Min. Active Days Before 🔒 Locked | 0 / 60 |
| Min. Active Days After 🔒 Locked | 0 / 7 |
The two locked thresholds are managed by Orphex to keep the model statistically valid. Entities that fail any check are dropped with the reason shown.
Reading the results
Results page
The header summarizes the run: "We scanned Apr 1 – Apr 30 and found 10 campaigns. For each, we compared 60 days before the change to 7 days after."
Three summary cards: Analyzed, Real Impact, Not Clear Yet.
The table has one row per entity. Key columns:
Entity name,
Changed (date),
Real Impact (absolute lift, e.g. +1,466.1),
the change itself (e.g. 7,000₺ → 21,000₺),
Covariates Δ,
Verified? badge.
Detail panel
Three sections per entity:
What Changed (Layer 1 — descriptive) The intervention itself + a metric table with daily averages (Conversions, Clicks, Cost: Before / After / %).
Did [Target] Change Drive Results? (Layer 2 — causal) The headline (e.g. +1,466.1 Real Impact: Your change added approximately 1,466 Conversions over 7 days) plus three cards:
Card | Example |
Without The Change (counterfactual daily avg) | 143.8 |
What Happened (actual daily avg) | 353.2 |
Your Impact (causal lift, daily avg) | +209.4 |
When the result is Not Clear Yet, this section lists the reasons instead and recommends a retry.
Trend Line chart with Your Campaign (solid), Without the change (dashed), and the change date marker. Visually, the gap that opens up after the marker = the causal lift.
When does "Not Clear Yet" happen?
Common reasons, in order of frequency:
Post-change window is too short
Pre-change window is too sparse
Multiple confounding changes too close together
Low conversion volume — noise overwhelms signal
Indirect-attribution campaigns (Shopping, PMax, Discovery) — delayed effects
The change was too small to be measurable
Covariates didn't correlate with the target metric
Each result lists the reasons that apply, so you know whether to wait, re-run with different settings, or accept the change wasn't measurable at this scale.
FAQ
Why an absolute number instead of a percentage? "+1,466 conversions over 7 days" tells you exactly what the change bought you. "+18%" doesn't until you do mental math against base volume.
Why no p-value? The model is Bayesian. Significance is read from the credible interval rather than a frequentist p-value.
I made multiple changes to the same entity — which one will it measure? The system flags concurrent changes; the model still runs but its attribution will be muddied. Clean setup means spacing changes apart in time.
Cost effect came back near zero even though I doubled the budget — why? Likely a tracking gap, an unlifted delivery cap, or a learning-phase reset. Check platform-side delivery before trusting the conversion-side numbers.
When should I use Change Impact (Fixed Date)? When the change isn't visible in platform metadata — landing page swaps, audience exclusions, tracking changes, external events. You supply the date manually.
Which platforms are supported? Google Ads and Meta Ads in beta. TikTok and Criteo are coming.
Beta limitations
Google Ads and Meta Ads only
Manually triggered (no scheduled runs)
Requires conversion data (impression-only entities not supported)
Target metric fixed at run time
Each entity analyzed independently (no cross-entity aggregation)
Some change types not yet live (see "Coming soon" above)
Feedback
Reach out via Intercom or your account manager. Most useful feedback areas:
Does the Without The Change / What Happened / Your Impact framing help you make decisions?
Are the "Real Impact / Not Clear Yet" labels clear?
Which of the upcoming change types do you need most urgently?
For Change Impact (Fixed Date) — what use cases are you using it for?