Marketing Attribution: Models, Limits and Alternatives

Marketing attribution is the practice of assigning credit for a conversion to the marketing touchpoints that preceded it, so budget can be allocated to what actually works. Attribution models differ in how they split that credit: some award it all to the first or last touch, others distribute it across every interaction in the path using fixed rules or a trained algorithm.

That definition is deliberately narrow, and the narrowness is the point. Attribution describes what happened before a conversion. It does not, on its own, tell you what would have happened if the campaign had never run. This guide covers the models searchers expect to find, corrects the parts of the standard explanation that are now out of date, and sets out where user-level attribution stops being the right tool.

What is marketing attribution?

Marketing attribution answers a bounded question: given a conversion that already happened, which recorded interactions should receive credit for it, and in what proportion?

It is an accounting exercise applied to observed paths. You collect the touchpoints a converting customer had with your marketing, you apply a rule or a model, and you distribute the value of the conversion across those touchpoints. The output is a per-channel revenue or conversion number that rolls up into a report.

Nielsen, one of the older practitioners in this space, defines multi-touch attribution as a technique that "takes all of the touchpoints on the consumer journey into consideration and assigns fractional credit to each so that a marketer can see how much influence each channel has on a sale." Nielsen also draws the important methodological line: "Rules-based methods are subjective, as they rely on marketers to define the rules of how credit is allocated," whereas "algorithmic methodologies use objective, statistical modeling and machine-learning techniques" (Nielsen).

Single-touch vs multi-touch attribution

The foundational split in every marketing attribution model is how many touchpoints are allowed to receive credit.

• Single-touch attribution gives 100% of the credit to one interaction, usually the first or the last. It is cheap, transparent and wrong in a direction you can predict: first-touch flatters awareness channels, last-touch flatters closing channels.
• Multi-touch marketing attribution splits the credit across every recorded interaction in the path, either by a fixed rule set by the analyst or by weights learned from the data.

Multi-touch is more sophisticated, but sophistication is not the same as accuracy. A model that distributes credit across five touches is still only describing the paths it can see, using an allocation rule that nobody can validate from the data alone.

What marketing attribution is not

Three distinctions save a lot of arguments later:

• Attribution is not incrementality. Incrementality testing asks what would have happened without the spend. Attribution never observes that world.
• Attribution is not marketing mix modelling. MMM works on aggregate time series, not individual paths, and does not need user identity.
• Attribution is not a measure of channel value in isolation. A channel's attributed revenue depends on the other channels in the path and on the rule you chose.

Why marketing attribution matters

Despite the caveats, attribution earns its place for specific jobs:

• Budget allocation between campaigns and audiences inside a single platform, where the platform sees most of the path.
• Diagnostics: spotting which creative, keyword or audience appears early versus late in converting journeys.
• Speed. Attribution reports refresh continuously, so they support in-flight decisions that a quarterly model cannot.
• A shared vocabulary. Even an imperfect attribution model gives teams one agreed way to describe a customer journey.

The value is bounded by the quality of the path data underneath. If a large share of conversions arrives with an incomplete path, the model is interpolating over a gap, and no amount of model choice fixes that.

Marketing attribution models, and which ones still exist

Here is the full model taxonomy, including current availability in Google's own products.

Model	How credit is allocated	Underlying assumption	Available in Google Ads or GA4 today
First-touch (first click)	100% to the first recorded interaction	Discovery is what creates the sale	No, removed
Last-touch (last click)	100% to the final interaction before conversion	Closing is what creates the sale	Yes, in both
Last non-direct click	100% to the last non-direct interaction	Direct traffic reflects demand you already earned	GA4 applies a direct-exclusion rule across its models
Linear	Equal share to every interaction	Every touch contributed equally	No, removed
Time decay	Exponentially more credit to recent touches	Recency predicts influence	No, removed
Position-based (U-shaped)	40% first, 40% last, 20% split across the middle	Discovery and close matter most	No, removed
W-shaped	30% first touch, 30% lead creation, 30% opportunity creation, 10% across the rest	B2B funnel milestones are the meaningful events	Never a Google model, found in CRM and B2B tools
Data-driven (algorithmic)	Weights learned from your own converting and non-converting paths	Observed path patterns reveal relative contribution	Yes, and the default in both
Marketing mix modelling	Not a touch-level model; estimates channel contribution from aggregate time series	Aggregate variation in spend and outcomes identifies effect	A separate discipline, not an attribution setting

Four of the models most guides still teach no longer exist in Google's products

This is the most out-of-date part of the standard marketing attribution explainer. Google announced the change on the Google Ads Developer Blog in April 2023 and completed the migration that September.

Google Ads Help now states plainly: "The first click, linear, time decay, and position-based attribution models are no longer supported by Google." What remains is last click and data-driven attribution, the latter being the default for most conversion actions because it "uses your account's data to calculate the actual contribution of each interaction across the conversion path" (Google Ads Help).

Google Analytics 4 followed. Its Attribution reports now offer three models only: data-driven attribution, paid and organic last click, and Google paid channels last click. The documentation confirms that "the first click, linear, time decay, and position-based attribution models are no longer available as of November 2023," and notes that all GA4 models exclude direct visits from receiving credit unless the entire path is direct (Google Analytics Help).

The practical consequences:

• If a guide tells you to select a time-decay model in GA4, it is describing an interface that has not existed since November 2023.
• Historical reporting built on those models is not comparable to anything you can generate now.
• The direct-exclusion rule in GA4 quietly reallocates credit away from direct traffic, which changes channel numbers independently of the model you pick.

You can still build linear, time-decay and position-based logic yourself in a warehouse or a BI tool. Nothing stops you. But you are then maintaining a rule that the largest advertising platform in the world removed because it did not consider it defensible.

One conversion path, five models, five different budgets

Every guide draws diagrams of attribution models. Almost none show the same path scored under each one, which is where the disagreement becomes obvious.

Take a single customer who places a $400 order after five recorded touches over 28 days:

1. Paid social ad, day 0
2. Organic search to a blog post, day 14
3. YouTube ad, day 21
4. Email click, day 26
5. Brand paid search click, day 28, which converts

Now score that identical path under five models. Time decay uses a seven-day half-life. Position-based uses the standard 40/20/40 split.

Touchpoint	Day	First-touch	Last-touch	Linear	Time decay	Position-based
1. Paid social	0	$400	$0	$80	$10	$160
2. Organic search	14	$0	$0	$80	$38	$27
3. YouTube ad	21	$0	$0	$80	$76	$27
4. Email	26	$0	$0	$80	$124	$26
5. Brand paid search	28	$0	$400	$80	$152	$160

Read the paid social row across. Under first-touch it earned the entire $400. Under last-touch it earned nothing. Nothing about the customer, the media or the order changed between those two numbers. Only the accounting rule did.

Scale that to a full account and the model choice becomes a budget decision by proxy. A team that switches from last-click to a position-based rule will see upper-funnel channels look dramatically more efficient overnight, and will move money accordingly, on the strength of an assumption nobody tested.

Attribution windows move the number more than the model does

For most accounts, the lookback window is a bigger lever than the model, and it gets a fraction of the attention.

The window defines how far back from a conversion the system will look for qualifying interactions, and whether views count alongside clicks. Three things follow:

• A longer click window pulls more early touches into the path, which mechanically shifts credit towards discovery channels.
• View-through windows are where platform numbers diverge most. A one-day view window and a seven-day view window on the same campaign produce very different conversion counts.
• Comparing two platforms with different default windows is not a comparison at all. It is two different questions rendered in the same units.

Document the window alongside every attribution number you publish, and change it deliberately rather than accepting each platform's default. When a channel's performance appears to change overnight, check the window before you check the creative.

Where multi-touch marketing attribution breaks down

To be fair to multi-touch attribution: inside a single walled garden, for a logged-in user, on a short purchase cycle, it is a genuinely useful optimisation signal. The platform sees the impressions it served, the clicks it received and the conversion it recorded. That is a reasonably complete picture.

The problems begin when you try to use it as a cross-channel source of truth.

Identity and coverage gaps

Multi-touch marketing attribution requires stitching interactions to a person across sessions, devices and browsers. Every year that job gets harder: Safari and Firefox restrict third-party cookies and cap script-set first-party cookie lifetimes, mobile apps operate under platform frameworks rather than web tracking, consent choices remove a share of users from measurement entirely, and people simply switch devices mid-journey.

The commonly repeated framing that the industry is preparing for a cookieless future is now wrong in both directions. Third-party cookies remain in Chrome. At the same time, the standards-based replacement for privacy-preserving measurement was withdrawn: on 17 October 2025 Google announced it was retiring ten Privacy Sandbox technologies, including the Attribution Reporting API, Topics and Protected Audience, explaining that "after evaluating ecosystem feedback about their expected value and in light of their low levels of adoption, we've decided to retire the following Privacy Sandbox technologies" (Privacy Sandbox blog).

So the honest position is not that attribution is about to break. It is that the coverage gaps are structural, they vary by browser and platform, and no cross-industry standard is arriving to close them.

Walled gardens and double counting

Ask three platforms who drove a conversion and three platforms will tell you they did. Each applies its own attribution window, its own view-through rules and its own model, to the subset of the journey it can observe. None of them can see the others.

The result is familiar to anyone who has built a board deck: the sum of platform-claimed conversions exceeds the number of orders in the finance system, sometimes considerably. There is no shared identifier that would let the platforms deduplicate against each other, and no commercial incentive for them to try.

The only workable response is procedural:

• Treat your own order or CRM system as the single source of truth for how many conversions happened.
• Treat platform-reported conversions as directional optimisation signals, not as revenue accounting.
• Report a reconciliation ratio, platform-claimed conversions against actual conversions, and track how it moves. A drifting ratio is a measurement warning long before it is a performance story.

Correlation, not causation

This is the structural limit, and no amount of model sophistication removes it.

Attribution scores the observed paths of people who converted. It cannot distinguish a touchpoint that caused a conversion from a touchpoint that simply appeared on the way to one. Brand paid search is the clearest case: a customer who has already decided to buy searches your brand name, clicks the ad, and converts. Last-click attribution records that as an ad-driven sale. In many cases the sale would have happened anyway through the organic result immediately below.

Retargeting has the same shape. Retargeting audiences are, by construction, made of people who already showed intent. Any model that scores their subsequent conversions will make retargeting look excellent.

Data-driven attribution helps with the credit split, because the weights come from your own converting and non-converting paths rather than from an analyst's opinion. It does not create a counterfactual. The academic work on algorithmic attribution has pushed in sensible directions, modelling interaction effects between channels, the decay of advertising effects over time and heterogeneity in individual response, and reporting confidence intervals rather than point estimates. That is a real improvement in rigour, and it is still inference from observed paths.

Marketing attribution vs marketing mix modeling

Marketing mix modelling approaches the same business question from the opposite end. Instead of tracking individuals, it models aggregate outcomes against aggregate inputs over time: spend by channel, price, promotions, distribution, seasonality, competitor activity and macro conditions. It estimates adstock, or carryover, and saturation curves that show diminishing returns as spend increases in a channel.

Because MMM never touches user-level data, it is unaffected by cookie loss, consent rates and cross-device journeys. It also covers channels that user-level attribution cannot reach at all: television, out of home, print, radio, sponsorship, retail media and word of mouth.

The approach is not niche. Google made Meridian, its open-source marketing mix model, generally available on 29 January 2025, describing it as using "Bayesian causal inference that allows you to blend your prior knowledge with real-world data" and noting that it "accounts for reach and frequency, not just impressions" (Google). Meta maintains Robyn, an open-source MMM package from its Marketing Science team. The two largest sellers of user-level advertising both now publish aggregate modelling tools, which tells you something about where they expect measurement to land.

Dimension	Multi-touch attribution	Marketing mix modelling	Incrementality experiment
Question answered	Which touches preceded the conversion?	How much did each channel contribute to the outcome?	What would have happened without this spend?
Unit of analysis	Individual user path	Channel by time period	Treatment cell versus control cell
Data required	Tagged user-level events and a stable identity	Aggregate spend, outcomes and control variables over a long history	A deliberately withheld group
Covers offline, TV, OOH	No	Yes	Yes
Requires user identity	Yes	No	No
Privacy and consent exposure	High	Low	Low
Produces a counterfactual	No	Modelled, not observed	Observed
Typical refresh	Continuous	Monthly or quarterly	Per test
Best used for	In-platform optimisation, creative and audience decisions	Budget setting, saturation curves, scenario planning	Settling a contested channel question

At Aryma we build the interpretation layer that sits across those three, connecting attribution, incrementality, saturation and prediction into a single view rather than three competing reports. That is the job MMM Singularity is designed for.

Choosing a method: a decision framework

The right question to ask is not which marketing attribution model is best. It is which method answers the decision you are about to make.

1. Tactical, in-platform, short cycle. Optimising bids, audiences or creative inside one platform, with a purchase cycle measured in days. Use the platform's data-driven attribution, accept its blind spots, and do not export the number as truth.
2. Strategic budget setting across channels. Deciding next quarter's split across paid, brand, retail media and offline. Use marketing mix modelling. Attribution cannot see most of what you are allocating to.
3. One contested channel. Somebody insists brand search or retargeting is carrying the account. Run an incrementality test. This is the only method that produces an observed counterfactual, and it settles the argument permanently.
4. Everything above the smallest accounts. Triangulate. Run MMM for the strategic split, experiments to calibrate it, attribution for day-to-day steering, and treat disagreement between them as information rather than as a data quality problem.

Business type changes the emphasis:

• B2B with long sales cycles. Individual paths span months and multiple stakeholders in one account. Milestone-based models such as W-shaped are common in CRM tooling, and account-level rather than person-level analysis usually matters more than the model choice.
• Ecommerce with retail and offline sales. A meaningful share of revenue never appears in a web analytics path at all. MMM is not optional here.
• Mobile apps. Measurement runs through platform frameworks with their own aggregation and delay rules, not through web attribution. Design for those constraints rather than fighting them.

How to validate marketing attribution

This is the section almost no guide includes. Every page tells you how to build attribution. Very few tell you how to find out whether it is right.

Attribution outputs are model outputs. They deserve the same scepticism as any other model output.

• Run geo holdouts. Withhold a channel in a matched set of regions, keep it running elsewhere, and compare outcomes. This is the cleanest practical counterfactual most advertisers can build.
• Use platform lift tests where they exist. Ghost ad and PSA-based designs give a randomised control group inside the platform.
• Try switchback tests for always-on channels where geographic splits are impractical, alternating on and off periods over time.
• Compare the experiment result against what attribution claimed for the same channel and period. If last-click says brand search drives a large share of revenue and a geo holdout shows a small drop when it is paused, the attribution number is measuring intent, not influence.
• Calibrate MMM against experiments. Experiment results make excellent priors in a Bayesian MMM, which is exactly the blend of prior knowledge and observed data that Meridian's documentation describes.
• Backtest model forecasts on a held-out period before you trust them for planning.
• Establish a cadence. Attribution reviewed weekly, MMM refreshed quarterly, at least one incrementality test per major channel per year. Attribution is one input into marketing effectiveness overall, not a substitute for it.

When the methods disagree, the experiment wins. That rule, agreed in advance and written down, prevents most measurement disputes from becoming political ones.

How to implement marketing attribution

A practical sequence, in the order that actually works:

1. Define the conversion and name the system of record. One definition of a conversion, one system that holds the truth about how many happened, agreed before any tooling decision.
2. Map the customer journey, including the parts you cannot track. Knowing where the blind spots are is more useful than pretending they do not exist.
3. Fix collection before modelling. Consistent UTM conventions, server-side event collection, first-party identifiers where consent allows, deduplication rules for repeat events.
4. Set the attribution window deliberately and document it. Click window, view window, and the rule for direct traffic.
5. Choose one model and freeze it for a full reporting period. Comparability over time is worth more than incremental model sophistication.
6. Reconcile platform numbers against the system of record every period, and publish the ratio.
7. Add marketing mix modelling once you have enough aggregate history, particularly if offline or brand spend is material.
8. Calibrate with experiments, and re-calibrate on a schedule.

Most of the value lands in steps 1 to 4. Teams that skip them and start at step 5 end up with an elegant model built on inconsistent inputs.

When marketing attribution is not worth building

Every ranking guide on this topic is published by a vendor whose article resolves to its own tool, so nobody says this: sometimes the answer is not to build attribution.

Consider skipping or radically simplifying it when:

• Conversion volume is low. Data-driven attribution learns weights from path patterns. With few conversions and little path variety, those weights are noise dressed as insight.
• You run essentially one channel. If almost all spend sits in one place, use that platform's reporting and spend the engineering effort elsewhere.
• Most sales happen offline. Attribution can only score what it can observe. Go straight to MMM and experiments.
• The engineering and licence cost exceeds the money at stake. Identity resolution, server-side collection and warehouse modelling are real budget lines. If the decision they inform is worth less than the build, do not build it.
• Nobody owns the number. Attribution that no one is accountable for becomes a reporting artefact that three teams quote differently.

In those cases, a disciplined last-click report plus one well-designed incrementality test per year will beat a half-finished multi-touch build, at a fraction of the cost. If you are unsure which side of that line you sit on, our custom solution process starts by sizing the decision before it sizes the build.

Who owns the number

The organisational side of marketing attribution is rarely discussed and frequently the actual problem.

• One owner. A single named person or team is accountable for the definition, the model and the published figure.
• One definition. Written down, versioned, and referenced by every dashboard.
• A fixed cadence. Weekly tactical review, monthly channel review, quarterly strategic review tied to the MMM refresh.
• A change log. When the window, the model or the tagging changes, the number changes. Record it, or you will spend the next quarter explaining a step change nobody can locate.

If three teams currently report three different revenue figures, that is not a modelling failure. It is a governance one, and no attribution tool fixes it.

How Aryma approaches this

Our position is straightforward. The statistical core of measurement should be rigorous, causal and human-led. The work around it, generating insight, running scenarios, translating model output into a budget recommendation, is where AI agents genuinely help. We call that Peripheral Agentic MMM.

Applied to this topic, that means we do not treat attribution as a standalone answer. MMM Singularity acts as the interpretation layer, connecting attribution, incrementality, saturation and prediction so that a channel number carries its context: what the model estimated, what an experiment confirmed, and where the saturation curve says the next dollar should go. Aryma Nebula turns that into an actual allocation, and MMMGPT lets a team interrogate the reasoning in plain language rather than waiting for an analyst's availability.

The measurement question worth answering is not which marketing attribution model to select. It is how much of your outcome each channel actually caused, and how confident you are in that answer.

Frequently asked questions

What is the difference between MMM and MTA?

Multi-touch attribution (MTA) works at the level of individual user paths, assigning fractional credit to the touchpoints it can observe before a conversion. Marketing mix modelling (MMM) works on aggregate time series, estimating each channel's contribution to outcomes without needing user identity. MTA is tactical, continuous and privacy-exposed. MMM is strategic, periodic, and covers offline and brand media that MTA cannot see at all.

What attribution models are still available in GA4?

Three. Google Analytics 4 offers data-driven attribution, paid and organic last click, and Google paid channels last click. Google's documentation states that the first click, linear, time decay and position-based models have not been available since November 2023. All GA4 models also exclude direct visits from receiving credit unless the entire conversion path is direct, which shifts channel numbers independently of your model choice.

Why did Google remove the first-click, linear, time-decay and position-based models?

Google announced the deprecation in April 2023 via its Ads Developer Blog and completed the migration that September, extending it to GA4 in November 2023. The rules those models applied were analyst-defined rather than data-derived, and Google directed advertisers towards data-driven attribution, which learns weights from an account's own converting and non-converting paths instead of applying a fixed template.

Why do Google Ads and Meta report different conversion numbers?

Because each platform applies its own attribution window, its own view-through rules and its own model, to the portion of the journey it can observe. Neither can see the other, and there is no shared identifier to deduplicate against. The sum of platform-claimed conversions routinely exceeds real orders. Use your order or CRM system as the source of truth and treat platform figures as optimisation signals.

What is an attribution window?

The attribution window, or lookback window, is how far back from a conversion the system searches for qualifying interactions, and whether ad views count alongside clicks. It often changes reported performance more than the model does. A seven-day view window and a one-day view window on identical campaigns will produce materially different conversion counts, so always publish the window alongside the number.

What is incrementality testing and how is it different from attribution?

Incrementality testing withholds a channel from a randomised or matched control group and measures the difference in outcomes, producing an observed counterfactual. Attribution never does this: it scores the paths of people who converted, so it cannot separate influence from coincidence. Where a clean incrementality test and an attribution report disagree, the experiment is the more reliable answer.