Your Meta dashboard says 400 sales, Google Ads says 350, GA4 says 280 and your shop says 500. All four are “right”, because each answers a different question. Hunting for the one true model gets you nowhere. Match each measurement method to the decision it is good at, and stop asking it to do jobs it can’t.
What attribution actually answers
Attribution splits credit for a conversion across the marketing touches that came before it. That sounds like an answer to “which channel caused this sale?” It isn’t one. It answers “which tracked touches were present before this sale, and how should we share the credit between them?”
Two gaps hide in that sentence.
The first is the word tracked. Anything you can’t track gets no credit: word of mouth, a podcast mention, a WhatsApp forward, an ad seen on one device and acted on from another. Consent choices and browser restrictions widen this gap every year.
The second is “present before”. Being on the path does not mean causing the sale. A retargeting ad shown to someone who already had a product in their basket turns up on a lot of converting paths, and much of that credit would have happened anyway.
Everything below follows from those two gaps. Rule-based and data-driven models live inside the tracked world. Surveys, MMM and experiments are ways of looking outside it.
GA4’s models: last click and data-driven
Last click
Last click gives 100% of the credit to the final touch before conversion. It is simple, easy to explain and stable, which is why it survives.
As of October 2026, GA4 has only two rule-based options left. According to Google’s Analytics Help, its attribution reports offer three models in total:
- data-driven attribution, the default
- paid and organic last click, which ignores direct traffic and gives 100% of the value to the last channel the user clicked through
- Google paid channels last click, which gives 100% to the last Google Ads click and falls back to paid and organic last click if there wasn’t one.
The same page confirms that first click, linear, time decay and position-based models “are no longer available as of November 2023”. An older guide that tells you to compare first click against last click in GA4 is describing an option that no longer exists.
Two GA4 details catch people out. Direct visits get no credit in any model unless the whole path was direct, per the same Google article, so “direct” in an attribution report is a different thing from “direct” in a sessions report. And the lookback window moves the numbers. Google’s attribution settings page says acquisition key events (first visit, first open) default to 30 days, with 7 as an option, and all other key events default to 90 days, with 30 or 60 as options. Shorten the window and channels with long consideration cycles lose credit.
Use last click for day-to-day optimisation inside a channel, for comparing two landing pages or two ad sets at the same point in the funnel, and for spotting tracking breaks: a sudden drop to zero is a tagging problem, not a marketing one. Don’t use it to decide whether upper-funnel channels such as video, podcasts or influencer seeding deserve funding. It undervalues anything that starts journeys, every time.
Data-driven attribution
Data-driven attribution (DDA) uses machine learning to compare converting and non-converting paths, then shares credit according to how much each touch appears to change the chance of converting. That is how Google describes it in its attribution overview. Credit can be fractional, so a keyword might get 0.4 of a conversion.
For spreading credit across touches you can see, DDA beats last click. Both gaps still apply. It only sees tracked touches, and what it measures is association with conversion, not proof that a touch caused it. It is also a black box. You can’t inspect the weights, and the model shifts as your data shifts, so this month’s numbers aren’t strictly comparable with last year’s.
DDA gives you a fairer default view of the tracked funnel, and inside Google’s bidding it is the native model, so let it do that job. It is the wrong tool for comparing Google channels against channels Google can’t see, and it won’t prove incrementality to a finance team.
Platform-reported conversions
Every ad platform reports the conversions it believes it drove. The numbers are useful, and inflated by design. Each platform credits itself whenever its ad was on the path within its own attribution window, often counting people who only viewed the ad. Add Meta, Google, TikTok and LinkedIn together and the total will usually be higher than your actual sales, because the same customer gets claimed several times over.
Platforms still have the best view of what happened inside their own walls, and their optimisation algorithms run on that data. Keep the reporting consistent so trends mean something, and never add platform totals together.
So compare creatives, audiences and ad sets within a platform using its own numbers, and let those numbers feed its bidding. Splitting budget between platforms, or reporting total marketing return to the board, needs something else.
Self-reported attribution
Self-reported attribution means asking customers “How did you hear about us?”, usually as one question at checkout, at sign-up or on a demo form. It catches what tracking misses: podcasts, friends, communities, offline events, dark social links pasted into chats.
It has biases of its own. People remember the latest or most memorable touch, not the first. They pick the first plausible option in a list. Many answer “Google” when they actually searched your brand name after hearing about you somewhere else.
A free-text box or a randomised list stops option order from deciding the answer. Ask at the moment of conversion, not in a follow-up email a week later. Then read the results as a share of answers over time, never as the truth about any one customer.
Handled like that, survey answers are the best way to find channels tracking can’t see, and to check whether “direct” and “branded search” are other channels in disguise. They are too loose for precise return on ad spend calculations.
Marketing mix modelling
Marketing mix modelling (MMM) is a statistical model that relates spend in each channel over time to sales over time, controlling for things like seasonality, price and promotions. It needs no cookies or user-level tracking, so consent and browser changes don’t break it.
Data is the catch. Google’s documentation for its open-source MMM, Meridian, includes a worked example. A national model with 12 media channels, 6 control variables and 8 seasonality knots has 26 effects to estimate. Two years of weekly data gives 104 data points, about four per effect, which the documentation calls too low to estimate reliably. Its suggested fix combines channels and uses three years of data to reach about 15 points per effect, which “may allow some directional insight”.
Meta’s equivalent, Robyn, is also open source, and Meta describes it as experimental. Both tools support calibrating the model with experiment results, which tells you how much these models need ground truth.
Once you have enough history, MMM is the right tool for annual and quarterly budget splits across channels, offline and upper-funnel spend included. It is no help for weekly creative decisions, or at small spend levels where channel spend barely varies.
Incrementality tests
An incrementality test splits people or places into a group that can see your ads and a group that can’t, then compares outcomes. The difference is the lift your marketing caused. Of the six methods here, it is the only one that measures causation directly.
Options run from platform tools to geo tests you set up yourself. Google’s Conversion Lift is one example: at the time of writing, its help page lists a minimum of 1,000 observed conversions and a minimum campaign budget of $5,000, and recommends running for at least 14 days. Academic work such as the “ghost ads” method, published in the Journal of Marketing Research in 2017, improved how platforms build clean control groups without paying for placebo ads.
Tests are slow. They cost you the sales you hold back from the control group, and each one answers a single question. That is why you use them to calibrate the other methods, not to run the business week to week.
A test can tell you whether a channel causes sales at all, and whether your retargeting or branded search spend is mostly paying for sales you’d get anyway. Its answer covers the campaign, period and market you tested and nothing beyond. Results go stale as creative, prices and competition change.
Plumbing that every model depends on
None of the above works on broken data. Agree a UTM naming convention and stick to it. Google’s own guidance is that parameter values are case sensitive, so Facebook and facebook become two sources. A free extension such as our UTM builder can enforce lowercase and approved values.
Then run a tracking audit before scaling. Duplicate tags, missing purchase events and broken cross-domain journeys corrupt every model above, the expensive ones included.
Which number to use for which decision
When someone asks “which number is right?”, the answer depends on the decision in front of them.
| Decision | Best method | Sense-check with | Avoid |
|---|---|---|---|
| Which ad, audience or creative to scale inside one platform | Platform-reported | GA4 last click trend | Adding platform totals together |
| Which landing page or offer converts better | GA4 last click or a proper A/B test | Shop or CRM data | DDA (too slow, too opaque) |
| How to split next quarter’s budget across channels | MMM (if you have the history) | Incrementality test on the biggest channel | Platform-reported |
| Whether a channel is worth funding at all | Incrementality test | Self-reported share | Last click |
| Whether brand search or retargeting is mostly capturing existing demand | Holdout or geo test | Self-reported data | Platform-reported |
| Finding channels tracking cannot see | Self-reported | Branded search volume trends | Any click-based model |
| Reporting payback to finance | Blended CAC from actual revenue and total spend | Incrementality results | Sum of platform ROAS |
Get the last row wrong and the rest hardly counts. Platform ROAS is for steering. Finance needs blended customer acquisition cost built from real revenue, real spend and real margins, with the LTV to CAC payback view on top.
No single model has to be perfect if each layer checks the others. The cheapest layer to add is one free-text survey question at checkout or sign-up, asked at the moment of conversion. The most useful in the next quarter is an incrementality test on one big channel, starting with the one you most suspect of claiming sales it didn’t cause. Leave MMM until you have two to three years of weekly data, several channels with real variation in spend, and someone who can maintain the model.
Questions people ask
Which attribution models are still available in GA4?
At the time of writing (October 2026), GA4's attribution reports offer three models: data-driven, paid and organic last click, and Google paid channels last click. First click, linear, time decay and position-based were removed in November 2023, according to Google's Analytics Help.
Why do Meta, Google Ads and GA4 report different numbers for the same campaign?
Each platform counts conversions with its own rules, windows and data. Ad platforms credit themselves whenever their ad touched the path, often including views, while GA4 uses its chosen model and ignores direct visits unless the whole path was direct. None of them is wrong, they are answering different questions.
Is data-driven attribution the same as measuring incrementality?
No. Data-driven attribution shares out credit for conversions that were tracked, based on patterns in converting and non-converting paths. It does not tell you how many of those conversions would have happened without the ad. Only a controlled experiment, such as a holdout or geo test, measures that directly.
Should a small business bother with marketing mix modelling?
Usually not as a first step. MMM needs years of spend and sales history with enough variation to separate channel effects. Most small teams get more from clean UTMs, a post-purchase survey and one well-designed holdout test.
Sources
- Google Analytics Help: Get started with attribution
- Google Analytics Help: Select attribution settings
- Google Analytics Help: Collect campaign data with custom URLs
- Google Ads Help: Set up Conversion Lift based on users
- Google for Developers: Meridian amount of data needed
- Meta Open Source: Robyn
- Kellogg School of Management: Ghost Ads, Improving the Economics of Measuring Online Ad Effectiveness
