Multi-touch attribution is a lie (sort of)

I’ve done over 50 attribution projects at this point. Built custom models, implemented off-the-shelf solutions, presented results to CMOs and CFOs and board members. And here’s the thing I’ve learned that nobody in analytics wants to say out loud: every attribution model is wrong.

Not “slightly imprecise.” Wrong. They all produce a number that looks precise and scientific and is fundamentally a made-up allocation of credit. The models disagree with each other wildly, and there is no ground truth to test them against.

Let me show you what I mean.

The attribution disagreement problem

I had a client last year, a B2B SaaS company spending about $80k/month across Google Ads, LinkedIn Ads, organic content, and email. We ran their Q3 data through four different attribution models. Same data, same conversions, same time period. Here’s what each model said about the value of their LinkedIn Ads spend:

• Last-click: LinkedIn drove 8% of conversions. LinkedIn looks like a waste of money.
• First-click: LinkedIn drove 31% of conversions. LinkedIn is the hero.
• Linear: LinkedIn drove 19% of conversions. LinkedIn is doing okay.
• GA4 Data-Driven: LinkedIn drove 23% of conversions. LinkedIn is doing pretty well.

Same data. Same reality. Four different stories. The difference between the lowest and highest number is nearly 4x. If you’re the marketing director deciding whether to increase or cut the LinkedIn budget, which number do you use?

This isn’t a bug. It’s the fundamental nature of attribution. Every model makes assumptions about how credit should be distributed. Those assumptions are arbitrary. There’s no experiment you can run to determine whether the last click or the first click “actually” deserves more credit. The question itself is meaningless because the customer’s decision was influenced by all of the touchpoints, and the contribution of each one can’t be isolated from observational data alone.

Last-click is garbage, but at least it’s consistent garbage

I have a soft spot for last-click attribution. Not because it’s accurate. It’s clearly not. The last ad someone clicked before converting gets 100% of the credit, which is absurd for any purchase cycle longer than five minutes.

But last-click has one thing going for it: everybody agrees on how it works. When I say “last-click conversions,” every marketer, every analyst, and every ad platform knows exactly what I mean. There’s no debate about the methodology. No black box.

Last-click is also stable over time. If you’re comparing this quarter’s performance to last quarter’s, and you used last-click both times, the comparison is valid even though both numbers are “wrong” in absolute terms. The bias is consistent. Consistent bias is something you can work with. Random bias is not.

This is why I don’t get too upset when companies use last-click as their default reporting model. It’s bad in theory and useful in practice. Just don’t use it to make major budget reallocation decisions. It systematically overvalues bottom-of-funnel channels and undervalues awareness channels. If you’re using last-click to justify cutting brand spend, you’re going to regret it in twelve months when your branded search volume drops.

What GA4’s data-driven attribution actually does

Google’s Data-Driven Attribution (DDA) in GA4 sounds impressive. It uses machine learning. It analyzes your actual conversion paths. It “assigns credit based on how each touchpoint changes the probability of conversion.” Fancy.

Here’s what it actually does under the hood, as much as we can tell from Google’s published documentation and the API outputs.

DDA builds a model of conversion probability using Shapley values, a concept from cooperative game theory. (If your GA4 data itself is unreliable, none of this modeling matters — make sure your GA4 migration isn’t broken first.) Imagine every marketing channel is a “player” in a game, and the “prize” is the conversion. Shapley values calculate each player’s marginal contribution by looking at all possible combinations of channels and measuring how the conversion probability changes when you add or remove a specific channel.

In practice, Google builds this model from your GA4 data. It looks at converting paths and non-converting paths, then estimates the marginal contribution of each touchpoint.

The problems with this approach are real and worth understanding.

Sample size dependency. DDA needs a lot of data to produce stable results. Google requires at least 400 conversions per 30 days with at least 2 different touchpoints in the path. Below that threshold, it falls back to a rules-based model. Many businesses, especially B2B companies with long sales cycles and lower conversion volumes, don’t have enough data for DDA to work properly. I’ve seen DDA results fluctuate dramatically month-to-month for clients with 500-800 monthly conversions. The model just doesn’t have enough signal.

Cross-device blindness. DDA only sees what GA4 sees. If a user researches on their phone during lunch, sees a retargeting ad on their laptop at work, and converts on their home desktop, DDA only connects these if the user is logged in across all three. For most sites, it can’t, so it treats this as three separate users. The phone session and the work session get zero credit because, as far as the model knows, those users never converted.

Walled garden problem. DDA can’t see impressions. It can’t see YouTube views that didn’t result in a click. It can’t see the LinkedIn ad someone saw but didn’t click, which primed them to Google your brand name the next day. View-through conversions, especially on social platforms, are invisible to GA4’s model. This systematically undervalues upper-funnel channels.

Black box opacity. You can see the output (credit percentages per channel) but you can’t inspect the model. You can’t verify the Shapley value calculations. You can’t understand why the model shifted 5% of credit from organic to paid social this month. It just did. Trust Google, I guess.

I use DDA as one input. Never as the only input. It’s better than last-click for most companies, but treating it as ground truth is a mistake.

What I actually use: triangulation

After 50+ attribution projects, I’ve landed on an approach I call triangulation. It’s not original. Other smart people in measurement have described similar frameworks. But it works, and I’ve refined it over enough engagements to feel confident recommending it.

The idea is simple: use three independent measurement approaches and look for convergence. Where they agree, you can be more confident. Where they disagree, you know you have uncertainty that no single model will resolve.

Leg 1: Data-driven attribution (DDA)

This is your click-level, user-journey view. GA4’s DDA or a similar model. It tells you what’s happening at the individual touchpoint level. Its strengths are granularity and recency. Its weaknesses are everything I described above.

Leg 2: Marketing Mix Modeling (MMM)

MMM takes a completely different approach. It’s a statistical model (typically regression-based) that looks at aggregate data: how does total spend in each channel correlate with total conversions over time, controlling for seasonality, promotions, and external factors?

MMM doesn’t care about cookies. It doesn’t care about ad blockers. It works on aggregate data, so the individual tracking gaps don’t matter as much. It can also measure offline channels, TV, radio, out-of-home, which DDA can’t touch.

The downsides: MMM needs 2-3 years of historical data to produce reliable results. It’s a blunt instrument. It can tell you that “LinkedIn as a channel drives approximately X conversions per $1,000 spent” but it can’t tell you which LinkedIn campaign or which audience segment. And it’s slow to react. If you launched a new channel last month, MMM won’t have enough data to measure it for months.

I use Google’s open-source Meridian for most MMM implementations now. It’s Bayesian, handles small datasets better than traditional regression, and it’s free. The setup takes real statistical knowledge though. You need someone who understands priors, posterior distributions, and model diagnostics. This is not a marketing analyst’s tool. It’s a data scientist’s tool — and if you’re ready to work with raw data yourself, BigQuery is a good place to start.

Leg 3: Incrementality testing

This is the gold standard, and the hardest to execute. Incrementality tests are experiments. You turn off a channel (or hold out a geographic region, or pause spend for a time period) and measure what happens to conversions.

If you turn off Meta ads for two weeks and conversions drop by 15%, you now have causal evidence that Meta was driving roughly 15% of your conversions. No models, no assumptions. Actual cause and effect.

The problems: you have to be willing to lose money during the test. Turning off a channel means forgoing the conversions it was driving. Most companies can’t or won’t do this for their largest channels. Geo holdout tests are easier to stomach but require enough geographic diversity in your customer base to be statistically valid.

I try to run 2-3 incrementality tests per year for each client, targeting the channels where DDA and MMM disagree the most. If DDA says Meta drives 25% of conversions and MMM says it’s 10%, an incrementality test on Meta is the way to resolve that.

How triangulation works in practice

You end up with three numbers for each channel’s contribution. They won’t match perfectly. They shouldn’t. But patterns emerge.

For that B2B SaaS client I mentioned earlier, here’s what triangulation revealed about LinkedIn:

• DDA said: 23% of conversions
• MMM said: 14% of conversions
• Incrementality test (2-week pause in one region): estimated 17% contribution

The triangulated view: LinkedIn drives somewhere around 15-20% of conversions. That’s a range, not a precise number. And that’s honest. Claiming LinkedIn drives exactly 18.7% of conversions would be false precision.

The range was enough to make a decision. LinkedIn was clearly driving meaningful value. Not as much as DDA suggested, but significantly more than last-click’s 8%. The client increased LinkedIn spend by 20% and saw results consistent with the triangulated estimate.

Why your CFO doesn’t care about attribution models

I’ve sat in a lot of rooms with finance people. Here’s what I’ve learned: CFOs don’t care about attribution models. They care about two questions.

“If I give you $100k more, how much incremental revenue will it produce?” This is the investment question. Attribution models can’t answer it directly because they describe correlations in historical data, not causal predictions about future spend. MMM gets closer. Incrementality tests get closest.

“If I cut $100k from channel X, what happens?” This is the risk question. Again, attribution models describe the past. They don’t predict what happens when you change inputs. The only reliable answer comes from incrementality testing.

What CFOs actually want is a reliable estimate with explicit uncertainty ranges. “We believe cutting LinkedIn spend by $20k/month would reduce pipeline by $40k-$70k/month, based on our incrementality test results and regression analysis. There’s real uncertainty in this range.”

That’s a conversation a CFO can work with. Presenting it in a dashboard that actually gets used helps too. “Our DDA model shows LinkedIn drives 23% of conversions” is a conversation that goes nowhere because the CFO will ask how confident you are in that number and you’ll have to say “it’s a model output” which translates to “I don’t know” in finance-speak.

The honest conclusion

Attribution is a measurement problem without a clean solution. Every model is wrong. Some are useful.

My advice after doing this for years: stop looking for the “right” model. Start triangulating. Use multiple approaches. Accept ranges instead of point estimates. Run experiments when you can.

And when someone asks you “which channel is driving our conversions,” resist the urge to give a clean answer. The honest answer is always “here’s what we know, here’s what we don’t, and here’s what we’d need to test to narrow the uncertainty.”

That’s not a satisfying answer. But it’s the right one.