Multi-touch attribution for ecommerce: models, setup, and limits
Learn how multi-touch attribution splits ecommerce conversion credit, how to choose and implement models, and how to measure AI-assisted discovery honestly.
Ecommerce shoppers rarely discover, compare, and buy in one session. A product can appear in search, get discussed by an AI assistant, return through an email, and convert after a paid reminder. A single-touch report compresses that path into one source. Multi-touch attribution gives the path more detail, while still leaving important questions about missing data and causality. This guide explains the models, an implementation plan, and the product-data foundations that make the output usable.
What multi-touch attribution measures
Multi-touch attribution (sometimes written as multi touch attribution, or MTA) assigns fractional or proportional conversion credit to multiple eligible touchpoints before a conversion. A touchpoint can be an ad click, organic visit, email click, referral, product-page visit, or another interaction your measurement system can observe.
The output is an allocation. It is not a record of which channel caused the purchase. Nielsen's guide to multi-touch attribution describes the method as assigning fractional credit to touchpoints that influenced a key performance event. The distinction matters when a team turns the report into a budget decision.
Every MTA result depends on five choices:
- Conversion event: The outcome being credited, such as a purchase, subscription, qualified lead, or product trial.
- Eligible touchpoints: The channels, campaigns, views, clicks, and owned interactions that can receive credit.
- Identity rule: How sessions, devices, logged-in users, and orders are connected to one journey.
- Lookback window: How far before the conversion a touchpoint remains eligible.
- Attribution model: The rule or algorithm that divides credit among the eligible touches.
Change any one of these inputs and the channel report can change. MTA does not discover a single timeless answer. It produces a consistent view of observed journeys under an explicit set of assumptions.
A worked ecommerce journey
Consider a $160 order with four observed touches:
| Day | Touchpoint | What happened | Linear credit | Position-based credit* |
|---|---|---|---|---|
| 0 | AI assistant referral | A shopper clicks a product link from ChatGPT | $40 | $64 |
| 2 | Organic search | The shopper returns through a non-branded product-guide query | $40 | $16 |
| 5 | Paid social | A retargeting ad sends the shopper back to the product page | $40 | $16 |
| 7 | A product email leads to checkout and purchase | $40 | $64 |
*This example uses a 40% first-touch, 20% middle-touch, 40% last-touch rule. It is an illustration, not a universal weighting.
Linear attribution gives each touch 25% of the order. Position-based attribution gives more weight to the first and last touches. Both approaches allocate the full $160. Neither proves that ChatGPT, organic search, paid social, or email created the order on its own.
If the assistant opens a webview that drops the referrer, the first touch may appear as (direct) / (none) or disappear from the tracked journey. The same linear model then divides the visible path among three touches. The model did not decide the assistant was unimportant. The source signal never arrived.
Single-touch, multi-touch, and marketing mix modeling
Single-touch models are easy to explain and cheap to operate. MTA is more detailed. Marketing mix modeling (MMM) answers a different question at a different level.
| Approach | How it assigns or estimates credit | Useful for | Main limit |
|---|---|---|---|
| First-touch | Gives all credit to the first recorded touch | Comparing discovery campaigns and top-of-funnel entry points | Ignores the work that turns interest into a sale |
| Last-touch | Gives all credit to the last eligible touch, often the final click | Fast channel reporting and short purchase paths | Favors channels that close demand, including branded search and retargeting |
| Multi-touch | Splits credit across multiple observed touches using rules or a data-driven model | Comparing the roles of discovery, consideration, and conversion channels | Depends on observable paths, identity matching, and model assumptions |
| Marketing mix modeling | Uses aggregate spend, exposure, and outcome data over time to estimate channel contribution | Budget planning across online and offline media when user-level paths are incomplete | Does not explain an individual shopper's journey and needs enough historical variation |
Use the approach that matches the decision. MTA can help decide which campaigns deserve more or less investment within an observable digital journey. MMM can provide a broader view when impressions, offline media, retail activity, or privacy limits make user-level tracking incomplete. Incrementality tests are the stronger tool when the question is whether removing or adding a channel changes outcomes.
Common multi-touch attribution models
Rules-based models are transparent. They are also opinions encoded as formulas. Data-driven models estimate weights from observed converting and non-converting paths, which can be more responsive to the business and harder to inspect.
| Model | Logic | When it helps | Trade-off |
|---|---|---|---|
| Linear | Splits credit equally across every eligible touch | Establishing a neutral baseline when no stage has a defensible weight | Treats a brief reminder and a high-intent product visit as equally important |
| Time-decay | Gives more credit to touches closer to the conversion | Longer consideration cycles where recent interactions tend to matter more | Can understate early discovery and brand-building work |
| Position-based (U-shaped) | Gives extra weight to the first and last touch and shares the rest across the middle | Journeys where discovery and conversion are both important | The chosen percentages are arbitrary unless validated |
| W-shaped or custom | Reserves weights for milestones such as first touch, lead, opportunity, and close | B2B or commerce journeys with clearly defined stages | Adds complexity and can create false precision when stages are inconsistent |
| Algorithmic or data-driven | Estimates contribution from patterns across converting and non-converting paths | Teams with enough clean journey data and a specific optimization question | Sensitive to data quality, selection bias, and the model's training rules |
Google Analytics currently supports data-driven attribution and last-click variants in its reporting settings for paid and organic channels. Google says first-click, linear, time-decay, and position-based models were removed from its attribution reports in November 2023. Those models remain useful industry concepts, but do not describe them as current GA4 reporting choices. See Google's attribution settings and its current model guidance when documenting a GA4 setup.
Google's data-driven model uses property-specific data to estimate the contribution of interactions. That is more grounded than assigning a fixed share to every touch, but it is still an estimate of allocated credit. It is not causal proof or a universal ranking of channels.
How to implement multi-touch attribution
The software comes after the measurement design. Start with the questions the business needs to answer and make the path auditable.
1. Define the conversion and the decision
Choose one primary outcome for the first implementation. For most ecommerce teams, that is a completed purchase with order ID, revenue, currency, and item-level details. You may also track subscription starts, lead submissions, or checkout starts, but keep those events separate from purchase reporting.
Write down the decision the report should support. Examples include:
- Should discovery content receive more budget?
- Does a paid campaign create incremental demand or capture existing demand?
- Which product categories need more qualified traffic?
- Do assistant referrals reach products that convert or only pages that attract browsing?
The question determines the required touchpoints and the right level of detail.
2. Map identities, events, and product relationships
Create a data map before selecting a model. At minimum, define how these fields connect:
- user or consented customer identifier, where permitted;
- session and event timestamps;
- source, medium, campaign, content, and landing page;
- referrer or referring application when available;
- order ID, revenue, currency, refunds, and discounts;
- product ID, variant ID, SKU, quantity, and item revenue.
Do not treat an order ID as a user identity. An order can join the conversion to the transaction, while a consented user or session key joins eligible interactions to that transaction. Cross-device and anonymous paths will remain incomplete when the required signal is unavailable.
3. Instrument the touchpoints you control
Use consistent campaign parameters for email, SMS, affiliate, influencer, QR, PDF, and partner links. Preserve those parameters through redirects. Keep naming conventions stable, and prevent internal links from overwriting the original campaign source.
For events, capture the page type and the product or variant involved. A product-page visit tells a different story from a blog visit, even when both came from the same source. Exclude or clean payment gateways, authentication providers, checkout return URLs, and self-referrals before assigning revenue to a channel.
Known AI assistant referrals deserve their own channel rule when the source is visible. ChatGPT, Perplexity, Gemini, Claude, and Copilot can appear as referrals or as a recognized AI Assistant channel depending on how the click arrives. Their visible sessions are useful evidence, not a complete count of AI-assisted demand.
4. Set eligible channels and the lookback window
A 30-day window can make sense for a fast-moving product. A considered purchase may need 60 or 90 days. Use the actual decision cycle, then keep the window stable while comparing periods. If you change it, annotate the report because the eligible path has changed.
In GA4, the key event lookback window determines how far back a touch can receive credit. Current settings allow different windows for acquisition key events and other key events. Google's settings reference documents the available options and the fact that the window applies across attribution models.
5. Start with a simple baseline model
Use linear attribution as a transparent baseline when the team has no evidence for a weighting scheme. Compare it with last-touch and, where available, data-driven output. A large change between models is a finding. It shows that the budget decision depends on an assumption about where value sits in the journey.
Do not switch models every week to produce a favorable channel ranking. Keep a primary model for operating decisions and use alternatives as sensitivity checks.
6. Reconcile before interpreting
Compare the attribution dataset with the source of truth for orders and revenue. Check:
- purchase counts by day and order ID;
- total revenue, currency, tax, shipping, discounts, and refunds;
- duplicate events and missing transactions;
- item IDs and variants in ecommerce events;
- consent rates and known identity gaps;
- checkout, payment, and cross-domain referrals;
- campaign tags, referrer coverage, and Direct growth;
- changes to the site, consent banner, feed, or analytics implementation.
If the report cannot reconcile to orders, fix the data before debating the model. A sophisticated allocation over broken inputs only makes the error harder to see.
7. Report allocation with coverage and confidence
Report allocated conversions or revenue with the model name, lookback window, eligible channels, and observable-path coverage. Call the result “allocated revenue” or “influenced revenue.” Reserve “caused” for an experiment or another design that can estimate incremental impact.
At product level, include item ID, variant, landing page, channel, and outcome. This makes it possible to see whether a channel is sending shoppers to the right products, rather than only whether it sent sessions to the domain.
How to measure AI-assisted discovery without false precision
AI-assisted shopping adds a source problem to an existing attribution problem. A shopper can ask an assistant to compare products, click a recommendation, and arrive on a product detail page. An app, webview, copied link, privacy setting, or redirect may remove the referrer. GA4 then has no way to reconstruct the missing source after the fact.
Catalog's guide to referral traffic in Google Analytics explains the practical split:
- Visible AI referrals: A detectable source or recognized AI Assistant channel. These are the measurable floor.
- Unknown-source or Direct sessions: A mixed bucket that includes typed URLs, bookmarks, untagged campaigns, private apps, copied links, and some assistant clicks.
- Likely assistant-influenced Direct: A test segment based on landing page, new-user behavior, engagement, timing, and supporting visibility evidence. It is an estimate, not a source label.
Do not label every Direct product-page session as AI traffic. Start with a baseline and investigate a narrow segment, such as new users whose first session lands on a product, category, comparison, or buying-guide page, excludes account and support pages, and includes meaningful engagement or a purchase. Then cross-check against:
- known AI referrals in the same period;
- product and prompt visibility in answer engines;
- branded and product-name searches;
- paid, email, affiliate, influencer, and SMS calendars;
- product launches, promotions, PR, and marketplace activity;
- device and browser patterns;
- server logs when they preserve referrer information.
Catalog's guide to dark traffic in agentic commerce recommends reporting a range rather than a single invented number. Use three layers:
| Layer | What to report | What it means |
|---|---|---|
| Known AI floor | Visible AI sessions, purchases, and revenue | Confirmed by the source data that reached analytics |
| Likely dark estimate | Direct product-page activity above a tested baseline | A defensible hypothesis after competing explanations are removed |
| Influence signals | Visibility, citations, branded search, and product-level movement | Evidence that may align with AI discovery without identifying a click |
In an MTA report, include only observed AI touchpoints in the path. Keep likely hidden influence as a separate estimate or diagnostic layer. Folding the estimate into allocated revenue without labeling it turns uncertainty into fake precision.
Why product-data quality changes attribution quality
Attribution follows identity. If the same product appears as three URLs, two variant IDs, and a stale feed record, the journey and the outcome can split across records. If an assistant links to a non-canonical page or an unavailable variant, a visible referral may still produce a poor commerce outcome.
Use a product-data layer that keeps the following facts aligned across the product page, feeds, analytics events, marketplaces, and AI shopping surfaces:
| Product-data requirement | Discovery role | Measurement role |
|---|---|---|
| Stable product and variant IDs | Lets systems recognize the item and its sellable versions | Joins page events, item events, orders, and revenue |
| Canonical product URL | Gives assistants and shoppers one reliable destination | Groups landing pages and avoids splitting one product across URLs |
| Clear parent-child variant relationships | Helps match size, color, pack, and configuration requests | Prevents product and variant revenue from being mixed |
| Current price, currency, and availability | Keeps recommendations commercially valid | Adds context when traffic converts, abandons, or returns |
| Consistent taxonomy and attributes | Makes products comparable by use case and constraints | Enables channel and category cuts that mean the same thing over time |
| Provenance and update state | Gives systems confidence that facts are current | Helps explain changes in recommendations and conversion rates |
Catalog is the product data layer for AI commerce. It helps brands structure product data, keep facts current, and publish machine-readable product objects to AI shopping surfaces without replacing the existing storefront. Catalog is not an attribution platform and does not assign fractional conversion credit. Your analytics stack or warehouse still owns touchpoint collection, model logic, and revenue reporting.
The connection is operational. Better product identity makes the downstream path easier to read. Better price, availability, variant, and attribute data helps assistants recommend the right offer. AI visibility for ecommerce and Catalog's guide to trusted data sources cover the discovery side. MTA shows what can be learned after a shopper reaches the site.
What multi-touch attribution cannot tell you
MTA is useful when its limits are visible. Plan for these failure modes:
Missing touchpoints
Ad impressions, private-app clicks, offline conversations, untagged links, and assistant interactions may never enter the path. A model cannot allocate credit to a signal it did not receive.
Identity and consent gaps
Cookie restrictions, consent choices, browser privacy controls, and device changes break or limit journey stitching. Do not use fingerprinting or other methods that violate a user's privacy to make the path look complete.
Correlation mistaken for causation
People who see an ad or visit a product page may already be more likely to buy. A model can assign that interaction credit because it appears on converting paths. It cannot isolate the incremental effect without a suitable control or experiment.
Modeled events and delayed updates
Google Analytics can model key events that are not directly observable because of privacy or technical limits. Google's documentation on modeled key events explains that modeled data is estimated from patterns between observed and unobserved events, and that reports can be updated after a conversion. Treat modeled and observed numbers as different evidence layers when the distinction is available.
Platform-specific views
Ad platforms often use different conversion windows, eligibility rules, and identity graphs. Their totals will not necessarily match your analytics property or order system. Reconcile definitions before comparing numbers.
Product and revenue changes after the click
Price changes, stockouts, shipping promises, returns, promotions, and substitutions can change an order after the original discovery touch. A channel may send qualified demand to a product that could not be purchased. Include product state and item-level outcomes in the analysis.
When causal evidence matters, pair MTA with incrementality tests, holdouts, or marketing mix modeling. Use MTA to organize the observed path and experiments to test what changes when exposure changes.
Multi-touch attribution FAQ
What is multi-touch attribution in simple terms?
It is a method for splitting conversion credit across the eligible interactions that preceded a conversion. The model, lookback window, identity rules, and available data determine the split.
What is the difference between last-touch and multi-touch attribution?
Last-touch gives all eligible credit to the final touch before a conversion. Multi-touch allocates credit across several touches. Last-touch is easier to operate. MTA gives more visibility into the path, but its answer depends on what the system can observe and the weighting or algorithm it uses.
Which multi-touch model should an ecommerce team use?
Start with a transparent baseline, usually linear, then compare it with last-touch and a data-driven model when you have sufficient clean journey data. A short purchase cycle may not need a complex model. A long cycle may benefit from time-decay or a custom rule, provided the weights match a documented business question and are tested against outcomes.
Does GA4 have a multi-touch attribution model?
GA4 currently offers data-driven attribution and last-click variants for its paid and organic channel reporting. Google removed first-click, linear, time-decay, and position-based models from its attribution reports in November 2023. You can still implement those approaches in a warehouse or another analytics system when their assumptions suit your question.
Can multi-touch attribution prove that ChatGPT caused a sale?
No. A visible ChatGPT referral can document an observed touch. An MTA model can allocate part of an order to that touch. It cannot prove that the referral caused the purchase, and it cannot see assistant influence when the referrer is missing. Use experiments or incrementality methods for causal claims.
Does better product data improve multi-touch attribution?
It improves the inputs that make commerce journeys interpretable. Stable IDs, canonical URLs, clear variants, current offers, and consistent taxonomy make it easier to connect AI discovery, product-page behavior, item events, and orders. Product data does not replace an attribution model.
Multi-touch attribution is most useful when it is treated as a measurement system rather than a verdict. Define the conversion, document the path you can observe, make product identity consistent, report ranges for missing AI influence, and use experiments when the business needs causal evidence. If the weak link is the product-data layer behind AI discovery, talk to Catalog about making those records structured, current, and machine-readable.
