Resources
Broken attribution: how it distorts your CAC and your decisions
Short answer
Attribution is broken when there's a material, systematic gap between the credit your measurement stack assigns and the real causal impact of each channel. When that happens, it isn't just the report that breaks — the whole decision system does, because that same signal feeds automated bidding. An over-credited channel shows artificially low CAC and earns more budget; an under-credited one looks expensive and gets cut even while it keeps producing incremental customers.
Why this isn't a reporting problem
It's tempting to treat broken attribution as a dashboard issue. It isn't. Attribution now feeds three things simultaneously:
- The report your team reads.
- The bidding algorithm at Meta, Google and TikTok, which optimizes toward whatever conversions you declare.
- Next quarter's budget decision.
Which means bad measurement doesn't just distort the read — it degrades actual acquisition. In a large-scale Meta study (Wernerfelt, Tuchman, Shapiro and Moakler, 2024), advertisers optimizing to purchases reached a median cost per incremental customer of $38.16, while optimizing to clicks pushed it to $49.93 — a 31% increase. The same work estimated that losing offsite tracking data raised the median cost per incremental customer from $112.69 to $154.77 at six months — roughly a 37% increase.
The signal you hand the algorithm becomes the cost you pay.
The two families of causes
Technical causes
These are the fixable ones. They cluster in five places:
| Cause | Typical symptom | Where you catch it |
|---|---|---|
| Click ID lost in redirects | Paid traffic surfacing as organic or direct | Platform clicks vs. sessions carrying the parameter |
| Misconfigured cross-domain | Self-referrals, split sessions, attribution to your own domain | Referral report; domain hops inside the journey |
| SDK + server-side duplication | Inflated conversions, artificially low CAC | Ratio of conversions to real backend orders |
| Mixed scopes | Numbers that don't reconcile between reports from the same system | Metric definition audit |
| Incomplete offline conversions | The channel driving real pipeline looks like the worst performer | Offline import match rate to CRM |
Structural and privacy causes
These get managed, not fixed: third-party cookie restrictions, ITP limits on storage and link decoration, consent and modeling in GA4 and Google Ads, ATT/IDFA constraints on iOS, and the aggregated, thresholded nature of Apple's frameworks.
The distinction matters. Spending three weeks trying to "fix" a structural signal loss is wasted time; the right response there is to model and validate with experiments, not to chase granularity that no longer exists.
How CAC gets distorted, with numbers
Take a simple portfolio: $100,000 monthly spend, 1,000 real new customers. True portfolio CAC: $100.
| Channel | Spend | Real customers | Attributed customers | True CAC | Reported CAC | Error |
|---|---|---|---|---|---|---|
| Branded search | $15,000 | 90 | 190 | $167 | $79 | −53% |
| Prospecting social | $45,000 | 520 | 380 | $87 | $118 | +36% |
| Retargeting | $20,000 | 140 | 290 | $143 | $69 | −52% |
| Paid-assisted organic | $20,000 | 250 | 140 | $80 | $143 | +79% |
The important part is how to read this. Portfolio CAC is correct — $100 either way, because total customers don't change. What's wrong is the distribution. And since budget is allocated by channel rather than by portfolio, the team will shift money out of prospecting and into retargeting and branded, buying more and more demand that already existed.
That's the pattern the eBay experiment documented most sharply: brand keyword ads showed no measurable short-run benefit, and in non-brand the positive impact concentrated among new and infrequent users, while frequent users absorbed spend without changing behavior.
How it contaminates payback and cohorts
The damage doesn't stop at this month's CAC. If attribution mislabels the source channel, every cohort built on that field inherits the error:
- LTV by channel is computed over mislabeled populations.
- CAC payback — how many months to recover the investment — stretches or compresses artificially per channel.
- Retention by source stops being comparable, because one channel's users show up under another's banner.
A channel bringing high-retention users but receiving little credit shows high CAC and low LTV at the same time. That's the combination that guarantees you cut it.
Diagnostic protocol
This is the order we run a measurement audit. It's deliberately cheap at the start and expensive at the end.
Tier 1 — Reconciliation (1 day)
- Sum of conversions reported by all platforms vs. real backend orders/leads. If the sum exceeds reality by more than 15–20%, you have over-counting.
- Platform-vs-analytics delta by channel, and its stability across 90 days.
Tier 2 — Technical integrity (2–3 days)
- Click ID survival through every redirect and consent wall.
- Deduplication rules between SDK/tag and server-side.
- Cross-domain configuration and self-referrals.
- Conversion window alignment across systems.
Tier 3 — Coverage (3–5 days)
- Share of traffic in denied consent and the modeling method applied.
- Offline conversion match rate to CRM.
- Business event coverage, not just surface conversions.
Tier 4 — Causal validation (2–6 weeks)
- A holdout or geo-experiment on the most suspicious channel. Usually branded search or retargeting.
Tiers 1 and 2 resolve most cases. Tier 4 is the only one that answers whether assigned credit corresponds to real impact.
How ADEX approaches it
Our measurement audit follows exactly that order and produces three concrete deliverables: a discrepancy map with defined acceptable thresholds, a prioritized list of technical fixes with estimated effort, and an explicit recommendation on which channel deserves an incrementality test first.
We don't hand over a 60-page report. We hand over the order in which things should be fixed.
What to expect, honestly
Fixing attribution does not lower your CAC. It makes your reported CAC resemble your real one — and that usually means some channels look worse after the audit, not better.
That's the correct outcome. A number that gets worse and is true is worth more than one that improves and is false. But it's worth preparing leadership for that conversation before you start, not after.
FAQ
How do I know if my attribution is broken? Fastest signal: add up the conversions all your platforms report and compare against real backend orders or leads. If the sum meaningfully and persistently exceeds reality, you have over-counting. If it falls well short, you have signal loss.
Why do Google Ads and GA4 never match? They use different windows, models and attribution criteria by design. Some discrepancy is expected and normal. The problem starts when the delta moves without a known cause, or exceeds the threshold you defined as acceptable.
Does broken attribution affect automated bidding? Yes, and it's the most expensive effect. Automated strategies optimize toward whatever conversions you declare. If those conversions are biased or duplicated, the algorithm systematically buys the wrong kind of user.
Is consent the main cause today? It's a significant cause, but rarely the only one. In most audits, technical causes — lost click IDs, duplication, cross-domain — explain more discrepancy than consent does.
Do I need incrementality, or is fixing tracking enough? Fixing tracking corrects measurement errors. Incrementality corrects interpretation errors. Different problems: you can have flawless tracking and still overvalue retargeting.
How long does an attribution audit take? Reconciliation and technical integrity tiers typically resolve in one to two weeks. Causal validation takes additional weeks, because it depends on running an experiment with sufficient statistical power.
Next step
If you suspect your numbers don't describe your real business, start with reconciliation. Audit my measurement →
Related reading
- Attribution models and how to evaluate an MMP
- ROAS vs incrementality: which metric for which decision
- Service: MMP attribution
- Service: marketing data dashboards
External sources
- Gordon, Moakler and Zettelmeyer (2023). Close Enough? A Large-Scale Exploration of Non-Experimental Approaches to Advertising Measurement. Marketing Science. https://doi.org/10.1287/mksc.2022.1413
- Wernerfelt, Tuchman, Shapiro and Moakler (2024). Estimating the Value of Offsite Tracking Data to Advertisers: Evidence from Meta. NBER Working Paper 32765. https://www.nber.org/papers/w32765
- Blake, Nosko and Tadelis (2015). Consumer Heterogeneity and Paid Search Effectiveness. Econometrica. https://doi.org/10.3982/ECTA12423
- Shao and Li (2011). Data-driven Multi-touch Attribution Models. ACM SIGKDD. https://doi.org/10.1145/2020408.2020453
- Google Ads and GA4 documentation on attribution models, conversion modeling and enhanced conversions; Apple documentation on ATT and AdAttributionKit; WebKit ITP.
The numeric scenarios in this article are illustrative, built to explain the distortion mechanism — not figures from a specific client.