Attribution models: why last-click makes you cut the wrong SEO and GEO

By Miguel Maio · Published 5 October 2026

The journey attribution can't see

1. 1Botched cross-domain: each domain a new session
2. 2Subdomain counted as a new visit
3. 3Self-referrals and dark traffic fall into Direct
4. 4Last-click credits only the end of the journey

Broken data drives biased decisions.

An attribution model isn't an analytics technicality: it's the lens that decides where your next euro goes. Choose it badly, or don't choose at all, and the whole budget tilts toward whoever shows up at the end of the journey and against whoever did the work at the start. In practice, it's the difference between knowing what works and thinking you do.

What attribution models are (and why last-click lies)

An attribution model is the rule that splits the credit for a conversion across the touchpoints that led to it. The same sale can be assigned to the first click, the last, shared across all, or computed by an algorithm that weighs each channel's real contribution. It isn't neutral accounting: each model tells a different story with the same data.

Last-click, the default for a decade, gives 100% of the credit to the final touchpoint before the conversion. It's simple, it's auditable, and it's misleading: it treats the journey as if only the end existed. The customer who found you in an article, came back through a newsletter and only then clicked a brand ad shows up in the reports as a clean paid-search conversion. The article and the newsletter worked for free, and in the numbers they never existed.

Rule-based models tried to fix this: first click (credit to whoever started it), linear (equal for all), time-decay (more to what's near the conversion) and position-based (more to start and end). They all share the same flaw: the rule is arbitrary, decided by you and not by the data. Data-driven came to answer that, and that's where the serious conversation begins.

Last-click vs data-driven: the model that penalizes SEO and GEO

Data-driven attribution (DDA) uses your own data to estimate how much each touchpoint actually contributed, comparing paths that converted with paths that didn't. Instead of a fixed rule, a probabilistic model. It's the most advanced model GA4 offers, and for good reason: it spreads credit closer to reality than any manual rule.

The difference isn't academic. Under last-click, a channel that assists a lot but rarely closes, like informational SEO or an AI citation, shows up with near-zero credit. Under data-driven, that same channel gets the share it earns by appearing in the paths that convert. The read on each channel's ROI changes, and with it the budget decisions. What looked under last-click like a channel burning money becomes, under data-driven, the one feeding the funnel.

It isn't perfect. DDA is a relatively closed box, it needs data volume to stabilize, and it's still stuck with what the tracking can see, garbage in, garbage out. But between deciding on a rule you know is wrong and deciding on a model that at least looks at the data, the choice shouldn't be hard. It is, because last-click gives simple numbers and simple is comfortable.

Why non-paid traffic always loses

SEO, GEO, email and direct act early and in the middle of the journey, which is why any end-of-funnel model undervalues them by design. These channels create demand, form preference and bring the visitor back. But whoever closes is often a brand ad or a direct click, and that's who last-click hands the trophy to.

There's a technical aggravator that makes this worse than it looks. A huge share of non-paid traffic arrives with no reliable referrer and lands in "Direct": clicks from apps, from AI answers, from documents, from social. AI traffic is a case study, because ChatGPT, Perplexity and the rest don't always leave a clean trail, and what you can't see you can't credit. I worked this problem in depth elsewhere, and the conclusion holds: an inflated Direct is almost always mis-attributed traffic, and the bill for that is paid by organic.

The result is a vicious cycle. Non-paid looks weak in the reports, its budget gets cut, the demand it fed dries up, and paid ends up fighting over smaller demand at higher cost. Bad measurement isn't just a blurry photo: it's a self-fulfilling prophecy.

Broken journeys: silos, subdomains and botched cross-domain

Before debating which model to use, there's a more basic question most people skip: can your data even see the whole journey? In many brands, no. The journey is broken into pieces that analytics counts as different people, and no attribution model rescues data that is born broken.

The most expensive error is botched cross-domain. If your site lives on one domain and the checkout, booking or payment lives on another, and cross-domain isn't configured, GA4 treats the handoff as a new session and the origin domain shows up as a self-referral. The conversion gets attributed to your own site as "referral", and the real journey evaporates. It's the kind of error that makes a manager swear organic doesn't convert, when organic does convert and it's the tracking that loses it at the checkout door.

Subdomains bring their own trap: depending on setup, blog.brand.com and shop.brand.com can be counted as new visits instead of one journey, inflating sessions and Direct and cutting paths in two. Add a badly built referral exclusions list (payment gateways showing up as the source of the sale), inconsistent UTMs (each team tagging its own way, or not at all), the dark traffic falling into Direct, and the Consent Mode gap, where non-consented visits vanish or get modeled depending on the setup. Each one alone distorts; together, they turn the report into fiction.

The pattern I see in the field is always the same: brands with the house split into silos, SEO, paid and CRM teams looking at different dashboards, each defending its own number, and no one owning the full journey. Attribution, in these cases, isn't wrong: it's measuring a reality that doesn't exist.

Attribution models in GA4: what you have and what you no longer do

GA4 simplified, and it's worth knowing what's left. GA4 now keeps three models, data-driven and two last-click variants (paid and organic, and Google paid channels); the rule-based models from Universal Analytics, first click, linear, time-decay and position-based, were removed in November 2023. In practice, the real choice is now between data-driven and last-click, with configurable conversion windows.

There's a distinction that trips a lot of people: GA4 separates attribution across paid and organic channels from the attribution that feeds Google Ads. They're different lenses on the same event, and comparing numbers across reports without knowing each one's model and window is a recipe for wrong conclusions. Before debating strategy, always confirm which model is active and which window it's counting, because the same conversion changes owner with the definition.

This guide stays at the concept and the read. The step-by-step, setting up cross-domain, fixing referral exclusions, designing a UTM taxonomy, is left for dedicated articles, because each is a serious topic and cramming them here would only give you a shallow guide. What matters, for now, is understanding what the number in front of you is actually saying.

How this turns into bad paid-media decisions

The real cost of bad attribution isn't in the report, it's in the decision that comes out of it. A biased read doesn't sit still: it becomes budget, becomes a brief, becomes next quarter. And because the bias always points the same way, the error doesn't correct itself, it compounds.

The script repeats. Last-click shows SEO and GEO with low ROAS, so their budget is cut. It shows brand paid with sky-high ROAS, so it's doubled down, when brand paid is often just harvesting demand that organic and awareness created, paying for clicks that would have come for free. You over-invest in what harvests, cut what sows, and next quarter's numbers confirm the wrong thesis, because demand, that one, dried up where no one was looking.

And then there's the opposite reaction, just as bad: the brand realizes the numbers don't add up, loses faith in all of it, and throws in the towel. It stops looking at attribution, goes back to deciding on gut and on whoever is loudest in the meeting. Swapping an imperfect model for no model isn't healthy skepticism, it's a return to the dark ages with an expensive dashboard on top. The problem was never that attribution exists; it was that no one fixed the data before trusting it.

Beyond deterministic attribution: MMM and incrementality

User-level attribution is going blind, and not from incompetence: by design. The reliable end of third-party cookies, Consent Mode, privacy restrictions and referrer-less traffic keep cutting the deterministic data attribution depends on. Deciding only on last-touch attribution, in a world that no longer feeds it, is flying with half the instruments broken.

The senior answer isn't a new tool, it's an old one that came back strong: marketing mix modeling (MMM), which estimates each channel's contribution at the business level through statistical methods, without needing to track the individual. Alongside it, incrementality tests and geo experiments answer the one question attribution never answers well: what would have happened if I hadn't invested in this channel? It's the difference between correlation and cause.

It isn't attribution or MMM, it's attribution and MMM, in layers. Attribution serves the day-to-day and tactical optimization; MMM and incrementality give the causal read at the overall-budget level, especially for the channels privacy made opaque. For most brands this still sounds like a luxury, but the direction is clear, and whoever ignores it will decide more and more in the dark as deterministic data shrinks.

What to do before trusting a number

Order matters, and almost everyone inverts it. First you fix the tracking, then you choose the model, only then you draw conclusions. Trusting an attribution model with a broken journey is decorating a house with no foundations: it looks fine until the first shake.

In practice, in order: make sure the journey is seen end to end (cross-domain and subdomains resolved, self-referrals and payment gateways on the exclusions list, UTMs on a single taxonomy, Consent Mode conscious of what it models). Only then compare models, and compare them properly, watching how per-channel credit shifts from last-click to data-driven. And for the big decisions, don't rely on the deterministic alone: triangulate with incrementality tests, however simple, to validate what attribution suggests. The detail of each of these steps is left for dedicated articles; here is the sequence, which is what most people get wrong.

The rule I use is short: without reliable measurement, channel strategy is faith; with reliable measurement, it's management. And reliable measurement starts long before the model, it starts with seeing the whole journey.

The cost of deciding on broken data

As long as you decide on broken data and on last-click, you'll systematically cut what compounds your growth and reinforce what merely harvests it, with the calm certainty of someone holding a report that agrees with them. The worst isn't having no data; it's having wrong data and trusting it, because that gives you the courage to make bad decisions at scale. The question to ask this week isn't which attribution model you should use. It's this: can my data see the whole journey, or am I always rewarding whoever shows up last to a party others paid for?

Sources

• Google - About attribution and attribution models (GA4)
• Google - Cross-domain measurement in GA4
• Analytics Mania - Cross-domain tracking in GA4

Frequently asked questions