The Signal Your Attribution Model Calls a Conversion Belongs to a Different Customer Than the One Your Campaign Reached

Attribution feels like math. An impression fires, a conversion happens, a graph connects them, and a report tells you the campaign worked. What that sequence quietly skips is a question worth sitting with: the identity resolution step in the middle is doing more interpretive work than most campaign reports acknowledge.

When your attribution model credits an impression for a downstream conversion, it is not observing a direct link. It is inferring one. A device ID, an email hash, a cookie, or a household cluster on the impression side gets matched to a customer record or a conversion event on the outcome side. That match runs through an identity graph, and identity graphs are probabilistic tools with coverage that varies by population segment, device type, geography, and data vintage.

The practical consequence is that the person your ad reached and the person your attribution model says converted may not be the same person at all. They may share a household. They may share a device. They may share nothing except a graph edge that exists because a probabilistic model gave the connection a confidence score above its threshold.

Why this is harder to see than it sounds

Attribution vendors present match rates and coverage statistics in aggregate. A platform might tell you it resolved 78 percent of your impression log to persistent identifiers. That number sounds reassuring, and in aggregate it may be accurate. But aggregate resolution rates do not tell you how resolution quality distributes across the specific segments your campaign targeted.

If your campaign targeted a behavioral or demographic segment that skews toward environments where identity signals are sparse, such as Safari browsers with Intelligent Tracking Prevention active, connected TV devices that rotate identifiers, or mobile app inventory without a logged-in user state, your resolution rate inside that segment may be meaningfully lower than the aggregate. The graph fills those gaps by extending to the nearest connected node, which may be a household member, a shared device, or an inferred associate rather than the person who saw the impression.

The conversion side has its own version of this problem. Conversion events are frequently collected through a pixel, a CRM update, or a first-party log. Each of those sources carries its own identifier type. When the attribution model tries to join an impression-side identifier to a conversion-side identifier, it may traverse two or three graph edges to make that connection. Each edge adds uncertainty. The final attribution credit can end up assigned to an impression that reached a real person in your target audience who did nothing, while the actual conversion came from someone who never saw your ad.

What a hypothetical example makes visible

Consider a simplified scenario: a brand runs a connected TV campaign targeting in-market auto buyers. The campaign delivers impressions to a household. The identity graph on the impression side resolves those impressions to a household cluster containing two adults. One adult sees the ad. The other adult, who shares a household email domain or a WiFi router signal, later visits a dealership and converts. The conversion event, logged through a CRM system, resolves to the same household cluster. The attribution model joins the impression to the conversion and reports a view-through conversion.

The credit is not fabricated. The graph connection is technically valid. But the causal inference, that the ad influenced the conversion, may be entirely wrong. The converting person may have been searching for vehicles independently for weeks before the campaign launched. The impression reached the wrong household member. The reported lift is real only in the accounting sense.

The structural gap most post-campaign reviews miss

Post-campaign analysis rarely audits the identity resolution step explicitly. Teams look at impression volume, reach, frequency, and conversion counts. They compare those numbers to a control period or a holdout group. They do not typically ask: what was the resolution confidence distribution of the impressions we are crediting, and how does that distribution compare to the resolution confidence of the conversions we are claiming?

That question is worth adding to your standard post-campaign review. Most attribution vendors and identity partners can provide resolution confidence bands if you ask for them, even if they do not surface them by default. Asking specifically for the share of attributed conversions that relied on probabilistic household extension rather than deterministic individual matching will tell you something useful about how much interpretive work your attribution report is doing on your behalf.

Practical steps for more defensible attribution

First, ask your attribution vendor to segment resolved impressions by identifier type and resolution method before you read the conversion report. Impressions resolved to a deterministic, logged-in identifier carry more confidence than those resolved through probabilistic household extension. Treating them as equivalent in your analysis will overstate certainty.

Second, where your campaign ran in environments with strong deterministic signals, such as publisher log-in environments or authenticated email newsletters, consider running a separate attribution pass for that inventory specifically. That subset will give you a cleaner read on causal relationship than the blended campaign total.

Third, compare your attribution output to your incrementality results rather than treating them as parallel validations of the same finding. Attribution tells you where credit was assigned. Incrementality tests, when designed carefully, tell you whether purchases would have happened without the campaign. When the two numbers diverge substantially, the most common explanation is identity resolution error in the attribution path, not a flaw in the incrementality design.

Fourth, review the vintage of the identity graph your attribution vendor used at the time of the campaign, not the current graph. Some platforms run attribution retroactively against a refreshed graph, which means the identities they connect today were not the identities they would have connected during delivery. A retroactive resolution pass can introduce connections that did not exist when your impressions actually ran.

What this means for how you read performance reports

None of this means attribution is useless. It means attribution is an estimate produced by a specific identity resolution method at a specific moment, and understanding the method tells you how much confidence the estimate deserves. A campaign that shows strong attributed conversion volume through predominantly probabilistic, multi-hop identity resolution deserves more skepticism than one where most attributed conversions connected through deterministic individual identifiers.

Building that audit into your regular reporting cadence is not a large lift. It is mostly a matter of knowing which questions to ask your vendors before a campaign ends, while the resolution logs are still accessible. The teams that do this consistently find that their optimization decisions improve, not because they distrust their tools, but because they understand what those tools are actually measuring.