State of Ecommerce Attribution 2026: What the Evidence Shows

What ads add differs by channel: in 225 DTC geo tests run on one vendor's platform, the median incremental ROAS was 2.31x overall and 0.70x for branded Google search, from a self-selected US sample (Stella, vendor-published, 2026). And measuring it is hard: across 25 field experiments, the median confidence interval on return on investment was over 100 percentage points wide (Lewis and Rao, Quarterly Journal of Economics, 2015).

What do controlled experiments say about attribution?

Three peer-reviewed papers and one preprint set randomized experiments against estimates built from observed data, the raw material of attribution reports, or show how hard such experiments are to run. Their abstracts describe tests of methods, not audits of a platform's own report.

Gordon, Zettelmeyer, Bhargava and Chapsky (Marketing Science, 2019) compared 15 US advertising experiments at Facebook, covering 500 million user-experiment observations, with estimates from multiple observational models. The observational methods often failed to produce the same effects as the randomized experiments, even after conditioning on extensive demographic and behavioral variables.

A larger preprint by Gordon, Moakler and Zettelmeyer (arXiv, revised October 2022) analyzed 663 Facebook experiments with over 5,000 user-level features, richer data than most advertisers can access. The authors write that, despite large-scale experiments and rich user-level data, they are unable to reliably estimate an ad campaign's causal effect.

Blake, Nosko and Tadelis (Econometrica, 2015) ran paid search field experiments at eBay. Returns from paid search were a fraction of non-experimental estimates, and brand keyword ads had no measurable short-term benefits.

Lewis and Rao (Quarterly Journal of Economics, 2015) report that across 25 field experiments with major US retailers and brokerages, the median confidence interval on return on investment was over 100 percentage points wide, and that informative advertising experiments can easily require more than 10 million person-weeks.

What does GA4 count by default?

Google's GA4 documentation lists 3 attribution models in its Attribution reports: data-driven, paid and organic last click, and Google paid channels last click (Google Analytics Help, 2026). All attribution models exclude direct visits from receiving credit, unless the whole path is direct. Google says its data-driven model takes a counterfactual approach and trains on data from randomized controlled trials. Google trains and reports the model itself, and the page prints no sample size, accuracy figure or outside audit for it.

How much signal has privacy removed?

The sources below describe limits, not a loss rate.

Safari. WebKit's documentation says Intelligent Tracking Prevention blocks all third-party cookies by default. It caps the expiry of cookies created in JavaScript on a landing page at 24 hours when it detects click IDs added to links as URL parameters, and deletes script-created cookies and other script-writeable storage after 7 days of no user interaction with the website.

Chrome. Google's post of 17 October 2025 refers to its decision to maintain Chrome's current approach of offering users third-party cookie choice, and retires the Attribution Reporting API.

GA4 consent mode. Among its conditions, behavioral modeling needs at least 1,000 events per day with analytics_storage='denied' for at least 7 days, and at least 1,000 daily users with analytics_storage='granted' on at least 7 of the previous 28 days. Without enough consented traffic, events from users who decline consent aren't reported, and data export such as BigQuery is listed among the features that do not use modeled data.

iOS apps. Vendor-published figures disagree on how many users accept Apple's tracking prompt: Kochava (26 June 2026) says 69.7% of iOS users shown it accept, Branch (13 August 2026) says around 35% of app users consent when offered it.

None of these sources measures what share of your own purchases an ad platform or GA4 fails to see. Your Shopify orders can.

What do published incrementality tests in DTC find?

Stella, an incrementality testing platform, posted results dated 26 May 2026 for 225 geo-based tests run between August 2024 and December 2025 on its own platform.

The page states its limits: DTC ecommerce brands, 90% on Shopify, in US markets, and users Stella calls self-selected, measurement-sophisticated advertisers. Grade: vendor-published, not independently audited. On brand search it points the same way as the eBay experiment above, though one is peer-reviewed and one is not.

Where are shoppers coming from now?

Adobe (company-published, January 2026) said that traffic to retail sites from generative AI tools, counted as shoppers clicking a link, increased by 693.4% year on year over the 2025 holiday season, in data covering over 1 trillion visits to US retail sites, and that the base of users remains modest. Shopify said, as reported by TechCrunch (company statement, second-hand, 5 August 2026), that AI-driven traffic and orders to its stores had tripled year on year in the second quarter.

Both are growth rates, not shares of traffic, and neither source gives a Dutch or EU series. GA4's default channel group has an AI Assistant channel for sources like ChatGPT, Gemini, Deepseek, Copilot or Grok, and counts Google's AI Overviews and AI Mode under Organic Search.

How strong is each piece of evidence, and what should a brand do with it?

A control group, as in a holdout or a geo test, is how the experiments above isolated what an ad added; estimates from observed data alone need stated assumptions. None of these sources says how big the gap is for your store. Each platform reports on its own ads; a holdout does not depend on one.

Three checks, in this order:

1. Reconcile claims to orders. For one closed week, add up the purchases each ad platform claims, note its attribution window, and compare the total with Shopify orders for the same days and time zone. For illustration: if Meta claims 400 purchases, Google Ads 350 and TikTok 100 in a week with 600 Shopify orders, the platforms claim 850 purchases for 600 orders, a ratio of 1.42 (850 / 600). Pass: the claims add up to no more than your orders. Fail: they add up to more, so some orders are claimed twice or counted on a different definition. A fail is a reason to test, not a verdict on one platform.
2. Watch MER. Divide total revenue by total ad spend for the same weeks; no platform's attribution enters it. For illustration: 120,000 EUR of revenue on 30,000 EUR of ad spend is a MER of 4.0 (120,000 / 30,000). If you raise spend and MER falls, revenue grew by less than spend did, whatever a platform reports.
3. Hold out regions. Split non-overlapping regions at random into two groups, as Google Research describes for geo experiments (2011), and check that their past revenue moved together. Pick a channel and regions you can afford to pause, pause it in one group for a planned period, and compare revenue with the group that kept spending. Stella's published tests ran a median of 33 days. For illustration: if the group that kept spending 2,000 EUR sold 52,000 EUR and a similar group with the channel paused sold 49,000 EUR, the gap is 3,000 EUR for 2,000 EUR of spend, a return of 1.5 (3,000 / 2,000). Pass: the groups moved together before the test and the gap is larger than their week-to-week swings. Fail: they did not, or the gap sits inside the swings. Call that inconclusive, not negative.

Later, once a holdout has told you what one channel added, a causal attribution read like Causality Engine's can give a second opinion on what each channel caused, from your GA4 export.

Sources, 30 September 2026: Gordon, Zettelmeyer, Bhargava and Chapsky (Marketing Science, 2019); Close Enough? (arXiv, 2022); Blake, Nosko and Tadelis (Econometrica, 2015); Lewis and Rao (Quarterly Journal of Economics, 2015); Stella incrementality benchmarks (Stella, 2026); attribution, consent mode and channel group (Google Analytics Help, 2026); WebKit tracking prevention (WebKit, 2026); Privacy Sandbox update (Google, 2025); Adobe holiday results (Adobe, 2026); Shopify on AI search (TechCrunch, 2026); Eurostat, online purchases by individuals (Eurostat, 2026); Kochava and Branch (vendor blogs, 2026); geo experiments (Google Research, 2011).

Related answers

• Causal attribution
• Incrementality testing
• Geo testing for marketing incrementality
• Blended MER vs platform ROAS: which should drive spend?
• AI referrals broke your attribution quietly