Attribution Modeling in 2024: What to Trust, What to Ignore, and What to Build Around

Attribution modeling sounds like a measurement problem. It isn't. It's a philosophy problem — a question about what you believe caused a conversion, dressed up in algorithmic clothing. Every attribution model makes a set of assumptions about human purchasing behavior, the causal relationship between ad exposure and conversion intent, and the relative importance of different touchpoints in a customer's path. Those assumptions are never fully correct. The models disagree with each other by design. And the data they're working from has been getting progressively less complete since 2020, as privacy legislation, browser changes, and Google's own data policies have removed a growing share of the signal that attribution models depend on to function.

Most accounts don't treat attribution this way. They pick a model — or accept whatever Google defaults them to — and then read the reported numbers as though they accurately describe what happened. Campaigns with strong attributed ROAS get more budget. Campaigns with weak attributed ROAS get cut. Keywords that attribution credits with conversions get higher bids. Keywords that attribution doesn't credit get paused. The model's assumptions become the account's strategy, without anyone explicitly deciding that's what should happen.

The consequence is predictable: accounts optimized hard against a single attribution model tend to over-invest in whatever touchpoints that model favors, and under-invest in whatever it discounts. For last-click, that meant over-investing in bottom-funnel brand and exact match terms and systematically under-investing in upper-funnel discovery. For data-driven attribution, the failure modes are different but equally real — and less visible, because the model presents itself as objective rather than opinionated.

Google has progressively consolidated its attribution options. What was once a menu of seven models — last click, first click, linear, time decay, position-based, data-driven, and last Google Ads click — is now effectively three: last click, data-driven, and a legacy access to the remaining rule-based models for accounts that explicitly opted to keep them. Understanding what each model actually does — not what Google says it does, but what mathematical operation it performs on your conversion data — is the prerequisite for reading attribution reports honestly.

The critical insight from this table is that every model has a systematic bias — a consistent direction in which it misattributes credit. This isn't a bug that will be fixed in the next update. It's a structural property of how the models work. Last click will always over-credit bottom-funnel terms. Time decay will always favor remarketing. Data-driven will always struggle with thin-data segments. Knowing the direction of the bias doesn't eliminate the problem, but it tells you which parts of your attribution report to read with the most scepticism.

Data-driven attribution (DDA) is Google's current default and its most heavily promoted model. The pitch is compelling: instead of applying a fixed rule to credit allocation, DDA uses machine learning to observe which combinations of touchpoints actually preceded conversions in your account data and distributes credit accordingly. It sounds like it replaces opinion with observation. It doesn't. It replaces one set of assumptions with a different, less visible set.

The first problem is the correlation-causation gap. DDA observes that certain touchpoint sequences correlate with higher conversion rates and allocates more credit to those sequences. But correlation in conversion path data is not causation. A branded search term that appears late in long conversion paths correlates strongly with conversion — not because it caused the conversion, but because users who were already highly intent on converting are the ones who search branded terms. DDA credits the branded term heavily. The credit reflects the user's pre-existing intent, not the ad's causal contribution. The campaign looks like it's performing brilliantly. Budgets shift toward it. Actual incremental performance doesn't change, because the conversions would have happened anyway.

The second problem is model opacity. Last-click attribution is wrong in a predictable, understandable way — you know exactly which touchpoints it ignores. DDA is wrong in ways that are much harder to detect, because the model's internal weightings are not exposed to the user. You cannot inspect the model to understand why it credited a particular touchpoint at a particular rate. You can observe the outputs, but you cannot audit the reasoning. This makes DDA's errors harder to identify and harder to correct for than last-click's errors — even though DDA is, on average, more accurate at a portfolio level.

Attribution models depend on observing complete conversion paths — the full sequence of touchpoints a user encountered before converting. Since 2020, that signal has been systematically degraded by a combination of privacy legislation, browser-level tracking restrictions, iOS consent changes, and Google's own data policies. The result is that a growing share of conversion paths are now partially or completely invisible to the models — and the models have no reliable mechanism for communicating how incomplete their data is.

1. 1

   Safari Intelligent Tracking Prevention (ITP) — Apple's ITP has progressively shortened the window in which cross-site tracking is possible in Safari, which holds approximately 25% of browser market share. Conversion paths that cross sessions or involve non-Google touchpoints are increasingly unobservable in Safari traffic, which means a significant share of multi-touch paths are truncated or invisible to Google's attribution models.

2. 2

   iOS 14.5 App Tracking Transparency — Apple's ATT framework requires explicit user consent for cross-app tracking on iOS devices. Opt-in rates average around 25–35%, meaning the majority of iOS app traffic is now attribution-dark. For accounts with significant mobile traffic or app conversion events, this represents a structural hole in attribution data that no model can fill without consent.

3. 3

   Google's search term privacy threshold — Since 2020, Google has withheld search term data for queries that don't meet a minimum privacy threshold. This doesn't directly affect conversion path data, but it reduces the ability to audit attribution outputs at the query level — making it harder to identify cases where DDA is misattributing credit to high-volume terms at the expense of lower-volume intent signals.

4. 4

   Third-party cookie deprecation — Chrome's phased deprecation of third-party cookies removes the mechanism that many attribution models use to track cross-site conversion paths. Conversions that involve touchpoints outside the Google ecosystem — display, video, social, direct — become harder to connect to their originating ad exposures. The share of conversion paths that attribution models can fully observe shrinks with each phase of the deprecation.

5. 5

   Consent Mode and modeled conversions — Google's Consent Mode fills attribution gaps with modeled conversion data — statistical estimates of what conversions probably occurred based on consenting users' behavior. This is a reasonable response to an unavoidable data gap, but it means that a meaningful portion of what your attribution reports show is not observed data but modeled inference. The two are not distinguished in the standard reporting interface.

The cumulative effect of these changes is that attribution models in 2026 are working from a significantly less complete dataset than they were in 2019. Estimates of the share of conversion paths that are now partially or fully unobservable vary widely, but figures around 35–45% are consistent with what practitioners report seeing in enhanced conversion audits. The models adapt — they use modeled conversions, they infer from consenting user behavior, they adjust their weightings — but adaptation to incomplete data produces outputs that are more uncertain, not less. The confidence intervals around attribution outputs have widened substantially, even as the presentation of those outputs in Google's reporting interface has remained equally precise.

Attribution data isn't useless. It's useful for specific questions and misleading for others — and most accounts use it for both without distinguishing between them. Understanding which use cases attribution data genuinely supports, and which ones it systematically distorts, is more valuable than chasing a more accurate attribution model.

Most practitioners know, in the abstract, that attribution data is imperfect. The problem is translating that abstract knowledge into concrete scepticism about specific numbers in specific reports. Here are the attribution report outputs that are most systematically misleading — not because the data is wrong, but because the numbers present themselves with a precision and completeness that the underlying measurement doesn't support.

1. 1

   Branded campaign ROAS under DDA — Branded search campaigns almost always show exceptional ROAS under data-driven attribution because users who search brand terms have typically already decided to convert. DDA observes the high conversion rate and credits the branded touchpoint heavily. The number is real; what it doesn't tell you is how many of those conversions would have happened through organic search if the branded campaign hadn't existed. Branded campaign ROAS is a measurement of intent capture, not of incremental value. Treating it as the latter leads to systematic over-investment in branded spend.

2. 2

   Remarketing campaign conversion volume under any model — Remarketing campaigns show strong conversion rates by definition: they target users who have already demonstrated intent by visiting your site or engaging with your content. Attribution models credit remarketing touchpoints that appear in conversion paths, but cannot determine whether those touchpoints caused the conversion or merely accompanied it. A user who was going to convert anyway and happened to see a remarketing ad in the process looks identical in attribution data to a user who was persuaded to convert by the remarketing ad. The numbers cannot be separated.

3. 3

   Year-over-year comparisons across a model change — If your account changed attribution models at any point in the comparison period — which most did when last-click was deprecated in 2021 — any year-over-year performance comparison that spans that change is comparing numbers produced by different measurement systems. The performance shift you observe may be real, or it may be an artefact of the model change, or both. Without a controlled holdout period under both models, the comparison is unreliable.

4. 4

   Conversion volume on campaigns below DDA's data threshold — As noted above, DDA requires 300 conversions and 3,000 ad interactions per month to function reliably. Below that threshold, the model produces outputs that may be partially rule-based without disclosure. For any campaign that doesn't clearly exceed these thresholds, DDA's credit allocations should be treated as approximate at best and potentially misleading at worst.

5. 5

   Cross-channel attribution that includes Google-assisted paths — Google's attribution models observe Google touchpoints. They do not observe touchpoints on other platforms — Meta, LinkedIn, programmatic display, email — unless those platforms' conversion events are explicitly integrated. When Google's model shows that a conversion path involved only Google touchpoints, it may mean the user only encountered Google ads, or it may mean the user also encountered non-Google touchpoints that are simply invisible to the model. The absence of non-Google touchpoints in a path is not evidence that they didn't occur.

The goal isn't to find a perfect attribution model — no such model exists, and the signal environment in 2026 makes perfect measurement structurally impossible. The goal is to build a decision framework that uses attribution data appropriately, supplements it where it's weakest, and avoids the specific failure modes that attribution over-reliance consistently produces. Here's what that framework looks like in practice.

1. 1

   Use DDA for Smart Bidding inputs and within-campaign optimization — Data-driven attribution is the right signal for Smart Bidding target-setting and for evaluating performance trends within campaigns over consistent time periods. Use it for what it's designed for: feeding Google's bid models and tracking directional performance shifts. Don't use it to compare campaigns with fundamentally different conversion path profiles or to make budget allocation decisions between branded and non-branded spend.

2. 2

   Use geo-based incrementality tests for your highest-stakes budget decisions — The most reliable way to measure incremental performance is to run a controlled geographic experiment: hold spend constant in a test region, reduce it in a matched control region, and measure the difference in conversion volume. This approach bypasses attribution entirely and measures actual causal impact. It's time-intensive and requires careful market matching, but for major budget reallocation decisions — particularly those involving branded spend, prospecting versus remarketing mix, or cross-channel investment — it's the only measurement method that actually answers the incrementality question.

3. 3

   Apply consistent model weighting to your portfolio-level reporting — Choose one attribution model and use it consistently across all campaigns and all time periods in your performance reports. This doesn't make the model right, but it makes directional performance shifts reliable: if attributed conversions increase under a consistent model, something real has changed. Model-switching mid-analysis introduces a confound that makes it impossible to separate genuine performance changes from measurement artefacts.

4. 4

   Flag campaigns below DDA's data threshold separately — Build a segment in your reporting for campaigns that consistently operate below 300 monthly conversions. Report their attribution data separately, with explicit caveats about reliability, and weight it accordingly in budget decisions. These campaigns are the ones where DDA is most likely to be producing rule-based outputs without disclosure — and where the attributed numbers are least trustworthy.

5. 5

   Use search impression share and auction insights as attribution-independent performance signals — Impression share, top-of-page rate, and auction insights data are not attribution-dependent. They measure your actual presence in the auctions your keywords are eligible for, independent of what happens downstream in the conversion path. A campaign with deteriorating impression share and stable attributed ROAS is losing competitive ground that attribution can't see. Including these signals in your performance framework gives you coverage over the parts of the funnel where attribution is weakest.

6. 6

   Treat attribution data as evidence, not verdict — The most important shift is epistemological: attribution data is one input into a decision, not the decision itself. When attribution data and other signals — auction performance, landing page engagement, incremental test results, CRM revenue data — point in different directions, the right response is to investigate the discrepancy rather than default to the attributed number. The attributed number has a clear provenance and a precise format, which makes it feel more authoritative than signals that are harder to quantify. That feeling is not the same as accuracy.

Attribution modeling in 2026 is not a solved problem, and anyone telling you it is — including Google — is oversimplifying. The models have improved. DDA is genuinely more accurate than last-click at a portfolio level. Enhanced conversions and server-side tagging recover meaningful signal that cookie-based tracking loses. Consent Mode modelling fills gaps that would otherwise be invisible. But the gap between what attribution models claim to measure and what they actually measure remains wide — and it has widened, not narrowed, as the signal environment has degraded over the last four years.

The practitioners who make the best decisions in this environment are not the ones who have found a better attribution model. They're the ones who understand their current model's specific failure modes well enough to work around them — who know which numbers to trust, which ones to treat as approximate, and which ones to supplement with measurement methods that attribution can't replace. That's a harder discipline than picking the right model. It's also the only one that actually works.