Field notes from the founders
Marketing attribution: why no two tools ever agree, and what to do about it
November 17, 2026
Ask Meta, Google Ads, GA4, Klaviyo and Shopify how many new customers you won last month and you'll get five answers. Each is correct by its own rules, none of them match, and if you add up what every platform says it drove you'll usually find more customers than you have.We've written about specific versions of this: Meta against Shopify, Klaviyo against Shopify, GA4 against Shopify. This post is about the general problem, because the same two causes sit underneath every one of them.Every marketing attribution number is built in two layers.The collection layer is what a tool can see: which visits, which clicks, which orders, and what each one was worth. The credit layer sits on top: the model and the time window that decide which channel gets the credit for a sale.Most conversations about attribution go straight to the second layer, to arguments about last click against data-driven. In our experience the first layer is the bigger and less understood source of disagreement. Both layers differ between tools. Both also differ between two businesses running the same tool, because consent banners, tagging, checkout set-up and settings are configured differently everywhere. Two brands with identical customers can get different numbers from GA4 purely because of how each one set it up.Sessions and channelsGA4 publishes its channel rules. Google's default channel group documentation lists them, rule by rule: Direct is a session whose source is "(direct)" and whose medium is "(none)", and so on down the list. You can't edit them, and they're only as good as what they read. Source and medium come from UTM tags, Google's auto-tagging and the referring site, all of which go missing or get mangled more often than anyone would like.The rules run on sessions, and sessions are constructed. GA4 stitches events into sessions using a cookie, then stitches sessions into users according to a reporting identity you choose in the admin settings: Blended (your own user IDs, then device IDs, then modelling), Observed (user IDs, then device IDs) or Device-based. Switch between them and your past reports change, because the setting changes how the data is shown even though nothing was collected differently. GA4 also offers three channel views (by session, by first user and by key event), each of which can give a different answer for the same traffic.So "how many sessions came from paid social" depends on tags, cookies, identity settings and which of three channel views you opened. None of that is visible on the dashboard.GA4 against its own BigQuery exportIf you export GA4's raw events to BigQuery, the obvious move for a brand that wants to own its numbers, you'll find they never fully reconcile with the GA4 interface. Google's own guide says the export lacks the "value additions like modeling, Google Signals, and attribution". The interface estimates users and sessions with an approximate counting method, groups rare values under "(other)" and withholds some small numbers, and late-arriving data keeps updating the export tables for up to three days. If GA4 can't be reconciled exactly with its own raw data, matching it exactly to another tool is out of reach.Cookie consentUnder European rules, consent has to be specific to each purpose. The European Data Protection Board's guidance says the answer "lies in granularity", meaning separate consent for each purpose, and its later opinions treat audience measurement and advertising as different purposes. Google's consent mode follows suit, with separate signals for analytics storage, ad storage, sending user data for advertising, and ad personalisation. So consent for analytics and consent for advertising are usually similar on a given site, and almost never identical. GA4 and your ad platforms see slightly different slices of the same visitors.The slices aren't random either. When we've seen a consent banner go live, sessions fell far more than page views, and page views far more than transactions. The exact figures vary widely between sites, and opt-in rates have improved as banners got better designed and visitors got tired of clicking reject, but the pattern holds: the closer an event is to a purchase, the less of it consent takes away. Visitors who browse casually decline more often than visitors who come back and buy. One knock-on is that conversion rate can appear to jump overnight with no change in the business, because sessions disappear faster than orders.The published research is thinner than we'd like, but it points the same way. An NBER study of GDPR's effect found it cut the number of consumers firms could observe by 12.5%, and the ones who remained were trackable for longer. One server-side tagging provider's 2026 data shows full consent on 61% of page-view events and 74% of purchase events. You don't know what you don't know, and the visitors you can't see are a different crowd from the ones you can.How Google fills in what it can't seeGoogle's consent mode has two settings. In basic mode, Google's tags don't load until a visitor consents. In advanced mode, they send cookieless pings when consent is declined: the consent state and a measurement, with no cookie read or written. Google then models what it couldn't observe, and each product does it differently.GA4's behavioural modelling estimates users, sessions and new users, if your property qualifies (at least 1,000 declined events a day for seven days, among other thresholds) and uses the Blended identity. It isn't applied to event counts. Google's documentation says unconsented events are counted raw, which inflates first visits and session starts, because a visitor without a cookie looks new on every page. Several reports and explorations don't use the modelled data at all, and the BigQuery export never has it.GA4's key-event modelling fills in conversions lost to consent, short cookie lifetimes and cross-device journeys. Where there isn't enough data, Google says modelled key events are credited to Direct.Google Ads puts its modelled conversions straight into the Conversions column, mixed in with the observed ones.So the same declined visitor can be modelled in Google Ads, partly modelled in GA4, counted raw as a new user in some GA4 reports, and absent from the export.Browsers and ad blockersSafari's Intelligent Tracking Prevention deletes cookies written by a site's JavaScript after seven days without a visit, and caps them at 24 hours if the visitor arrived through a tracked link from a known tracker. Since 2023, Safari's Link Tracking Protection strips click identifiers such as Google's gclid and Meta's fbclid in Private Browsing, Mail and Messages, while plain UTM tags pass through. Firefox isolates cookies site by site by default, and strips some tracking parameters in its strict mode. Chrome, after years of warnings, decided in April 2025 to leave third-party cookies alone. Apple's App Tracking Transparency, the iOS 14.5 change everyone remembers, governs tracking across apps, Meta's own app included. It's why Meta lost sight of so many journeys and leans on modelling, but on the web the browsers' own rules do the damage.Then there are blockers. Around one in three US internet users runs an ad blocker, according to GWI's 2025 survey data, and the most common privacy blocklist blocks both Google's tag and Meta's pixel by name.New and returning customersAll of this hits sessions first, and everything built on sessions inherits the damage: session counts, pages per session, session duration, conversion rate. It does the most harm to new against returning. Analytics tools decide that a visitor is returning by finding their cookie. A loyal customer who last visited eight days ago on an iPhone has no cookie left, so GA4 calls them new. For a high-retention business that's a serious trap: the tools overstate new customers and understate returning ones, which flatters acquisition and hides retention. Count new and returning customers from your store's own customer records, where a customer is an email address.Server-side trackingMost tracking still runs client-side: a script in the shopper's browser sends events to Google, Meta and the rest. Server-side tagging changes the route. With server-side Google Tag Manager, the browser sends events to a server on your own domain, which forwards them on. Google tag gateway, launched in 2025, serves Google's tag from your own domain through your content delivery network. Meta's Conversions API sends purchases from your server, and Meta deduplicates them against the browser pixel by event ID.Server-side tracking helps. Requests to your own domain are harder for blockers to catch, you control what's passed to whom, pages carry less script, and cookies set by your own server can last longer than Safari's seven-day cap, provided they come from the same site's own infrastructure. Cloudflare reported an 11% uplift in data signals for early users of Google tag gateway.The consent rules still apply. European regulators' guidance says they apply whenever code in the browser tells it to send tracking requests, wherever those requests end up, so a server-side setup needs exactly the same consent as a client-side one. Safari still applies too: it caps cookies from setups that merely disguise a third party as your domain at seven days, and Link Tracking Protection still strips click IDs. Server-side tracking helps, but it's no magic bullet.RevenueEven when the order is seen, its value differs. Tools disagree on tax, shipping and duties; on refunds, which Shopify books on the day they're processed while most platforms never hear about them; on store credit and gift cards; on currency conversion and time zones; and on subscription renewals, which are created on a server with no browser for a pixel to fire in. We covered how this plays out for Klaviyo and GA4.Other sources of differencePayment gateways such as PayPal and Shop Pay can take the credit for a purchase session unless they're excluded as referrers. UTM tags are case-sensitive and inconsistently applied. Redirects drop click IDs. Each platform filters bots its own way, and GA4 won't tell you how many it removed. A browser pixel and a server event that aren't deduplicated count the same sale twice. Modelled conversions keep arriving for days after the sale.Why attribution can't be solvedMarketing attribution tries to answer a counterfactual question: would this customer have bought if they hadn't seen that ad? You can never observe the answer for a single customer, so no dataset contains the right answer to check a model against. Only a randomised experiment, where a group is held back, reveals the causal effect. When researchers compared observational attribution methods with 15 large randomised experiments run on Facebook, they found the observational methods "often fail to produce the same effects as the randomized experiments" (Gordon and colleagues, Marketing Science, 2019).That's why attribution is unsolvable, and why every number in the credit layer is an estimate.Data-driven attributionData-driven attribution, Google's default in both GA4 and Google Ads, compares the paths of people who bought with the paths of people who didn't. For each touchpoint it asks how much that touch changed the predicted chance of a sale, and shares out the credit in proportion. Google describes it as "a counterfactual approach", and it's descended from Shapley values, a way of dividing a team's result fairly between its players.It's more sophisticated than a fixed rule, but it's still a model of the patterns in observed paths, trained per advertiser, and you can't inspect its weights. It can only credit the touchpoints it sees, and it inherits every gap in the collection layer.Last click and other rule-based modelsLast click is a model too: a rule that says the final touch did all the work. Google retired first-click, linear, time-decay and position-based models from GA4 and Google Ads in 2023, leaving data-driven and variants of last click. Shopify offers last non-direct click by default, plus first click, last click, linear and "any click", which credits every channel in full and adds up to more than your orders. Each gives a different answer to the same question, and none can be checked against the truth.Lookback windowsEvery tool also decides how far back to look. GA4 looks back 90 days for most key events by default. Google Ads counts conversions up to 30 days after a click, three days after an engaged video view and one day after an impression. Meta's default is seven days after a click or one day after a view, and since March 2026 a "click" means a link click, with likes, shares and saves moved into a separate engage-through window.So a customer who clicks a Meta ad on the 1st and buys on the 20th is invisible to Meta's default report, while a Google ad clicked on the same day still counts in Google Ads and GA4. A Meta ad seen but not clicked the day before a purchase earns Meta the sale, and earns GA4 nothing.This is also why GA4 and Google Ads give you different CPAs while both use data-driven attribution. Google Ads credits only Google's own ads, while GA4 shares credit across every channel, email and organic included. Google Ads reports a conversion against the date of the click, GA4 against the date of the purchase. The windows differ, and Google Ads includes its modelled conversions.Improve data quality firstNo attribution model will save you if the data underneath is poor. Most brands can switch on Meta's Conversions API today, free, through Shopify's Facebook & Instagram app. Wire consent properly into Google's consent mode, tag every link consistently, exclude payment gateways as referrers, and count customers from your store's records.At or near $10m in revenue, we'd strongly recommend full server-side tracking. You can self-host server-side Google Tag Manager on Google Cloud (Google suggests a minimum of three small servers, a few hundred dollars a month at most), or use a hosted service such as Stape, which we've found fairly cheap and fairly easy to set up; its plans run from free to $200 a month depending on volume. The real cost is setting it up and keeping it running, and at that scale the improvement in what your ad platforms can see is worth it.Agree a default everyone works toIf you can't agree on the numbers, you won't act on them. Finance runs on one figure, the paid team on another, the retention team on a third, and every budget conversation turns into a debate about whose dashboard is right. A sensible default that everyone agrees to work from beats a mix of sophisticated approaches nobody trusts. For most brands our size that means Shopify as the ledger, blended CAC as the headline, and one agreed model and window for sharing credit between channels (GA4's data-driven attribution reconciled to Shopify's new customers is a reasonable start), with sources cited every time.Calibrating with experimentsThe only way to check a model is an experiment. A geo test switches a channel off in some regions and compares them with the rest. Platform lift studies hold back a random group of users. Marketing mix modelling uses years of spend and sales history to estimate each channel's contribution.If a geo test shows that Meta's true CAC is about $150, you can choose the attribution model and window that put Meta closest to $150, and do the same for Google. The choice of model is then anchored to a measured result. Repeat the test when spend or season shifts, because the gap moves, and remember it calibrates a channel's total and says nothing about the split between campaigns inside it.The bar is high. Meta's open-source geo testing tool recommends 20 or more geographic units and plenty of history. Meta's lift studies want at least $5,000 of campaign spend and 500 conversions, and Google's want 1,000 conversions. Google's open-source MMM, Meridian, struggles on two years of weekly national data. Beyond the numbers, a geo test or an MMM takes significant planning, careful experiment design and time that most lean teams don't have. In practice the bar is normally well above $10m in revenue, even where conversion volumes say it's possible. If you can do it, great. If it feels like a step too far, you're not alone, and a clean collection layer with an agreed default will do far more for you than a model you can't validate.Marketing attribution is two hard problems stacked on top of each other, set up differently in every business, and nobody can solve it. What a small or mid-sized brand can do is collect the data as cleanly as possible, agree one view of the world with sensible defaults, write down why each tool disagrees, and check the defaults against an experiment if the money ever justifies it.That's what a company brain for D2C is for. We rebuild one common view of the numbers from structured data across every system you run, with sources always cited, read the business LTV:CAC first, and our agents act on that view in the open.If you're below roughly $1m in revenue, the samples are usually too small to separate signal from noise, and the data-quality steps above are the right place to start.If you'd like a second pair of eyes on why your tools disagree, book a call at signalovernoise.app/book-a-call or drop us an email at hello@signalovernoise.app.