Attribution Model in Plain Terms: What It Shows and What It Doesn't
An attribution model shows which touchpoints were recorded before a conversion. Here is what it measures, where the numbers disagree, and what stays invisible.
ADS Beast editorial teamPublished 10 min read
An attribution model assigns credit for a conversion to the touchpoints a person passed through before it. It works from observable data: clicks, impressions, sessions, form fills, purchases. What it produces is a version of the path, built from what was recorded, not the path itself. That gap explains most disagreements about numbers.
In short:
- An attribution model distributes credit across recorded touchpoints. It does not observe the decision.
- Google Ads and GA4 can report different conversion counts for the same campaign because they count different things over different windows.
- No model sees a conversation, a friend's recommendation, or an ad someone saw and forgot.
- A campaign still in the learning phase produces numbers that are settling, not numbers you can judge a channel by.
- Broken tracking ruins the model before the model runs: fix measurement first, then read the report.
What does an attribution model actually measure?
It measures recorded contact between a person and your marketing, then splits conversion credit across those contacts according to a rule. The rule is the model. Last click gives everything to the final touchpoint. A multi touch attribution model spreads credit across several. Data-driven attribution in GA4 uses your own conversion data to weight touchpoints by how much they appear to contribute.
The important word is "recorded." A marketing attribution model can only work with events that fired and were stored: a click with a tracking parameter, an impression counted by a platform, a session in analytics, a purchase event on your site. Anything that happened without leaving a trace is not in the dataset, and the model will not miss it, because it never knew it was there.
This is why a lead attribution model is really a statement about your tracking setup as much as about your channels. If the form fires twice, the model distributes credit across two conversions where one happened. If a click loses its parameter in a redirect, the touchpoint disappears from the path. The math stays correct; the input was wrong.
Why do Google Ads and GA4 show different conversion numbers?
They count different things. Google Ads credits conversions by click through a window that is 30 days by default and can be set from 1 to 90 days. GA4 keeps event data for 2 or 14 months and offers last-click or data-driven attribution. A click that lands outside one window or inside another appears in one report and not the other.
Neither number is wrong. They answer different questions. Google Ads is telling you what it believes it delivered against its own counting rules, which is what its bidding system optimizes toward. GA4 is telling you what your site recorded and how a model distributed credit across sessions. When you compare them side by side without accounting for the window, the retention period, and the attribution rule, you are comparing three variables at once.
The practical fix is to write down the settings before you compare anything: conversion window in each ad platform, data retention in analytics, attribution model in each report, and which events feed each channel. Then decide which number you manage by. Most teams pick one source of truth for channel decisions and use the platform numbers only for in-platform optimization.
| Question you are asking | Where to look | What the number means |
|---|---|---|
| How many conversions did this channel deliver against its own rules? | The ad platform's conversion column | Conversions credited by click, inside that platform's window |
| How did recorded sessions distribute credit across the path? | GA4 attribution report | Credit split by the selected model over stored events |
| Did revenue cover spend across all channels? | Blended revenue and total spend | A ratio, not a path; no touchpoint credit assigned |
That third row matters because channel-level attribution and blended measurement answer different questions. If you want the ratio view, it is worth understanding how MER vs ROAS differ before you mix them into one dashboard.
What a multi touch attribution model can and cannot tell you
A multi touch marketing attribution model can tell you which recorded touchpoints appear most often on paths that end in conversion, and how credit shifts when you change the rule. It cannot tell you which ad caused a sale.
Correlation is what you get. If display touchpoints appear on most converting paths, the model will give them credit. That may mean display helps, or it may mean display is cheap enough to reach many people, including people who were already going to buy. The model cannot separate those two stories because both produce the same data. This is the core limit of marketing attribution, and no modeling choice removes it.
What you can do with correlation is still useful:
- Compare the same period under two different models and look at which channels change rank. A channel that only wins under last click is doing a different job than one that wins under a spread model.
- Check whether channels that get credit also get credit in a holdout test, where you turn one channel off in a region or audience and watch what happens to total conversions.
- Look for paths where a channel appears early and often but rarely last. Those are the ones a last-click report will undervalue.
- Watch for channels that appear almost exclusively on paths that also contain branded search. Branded search tends to sit at the end of paths it did not start.
For the economics behind those paths, customer value over time belongs in the same conversation as attribution. The formula and its inputs are covered in LTV calculation, and it changes which touchpoints look worth paying for.
Where do offline and cross-device journeys go?
They stay partly invisible, and no model recovers them. A person who saw your ad, asked a colleague, then typed your brand name into a search engine three days later produces one recorded touchpoint: the branded search. The conversation and the earlier ad are gone from the data.
Cross-device journeys have the same problem in a different shape. A click on a phone, a purchase on a laptop, and a login that does not connect the two leave you with a session that never converted and a purchase that appears to have arrived from nowhere. Identity resolution helps where a person is signed in to the same account across devices, and it does nothing where they are not.
Offline influence is the harder case. A radio mention, a conference booth, a recommendation in a group chat: these move people and leave no event. A market attribution report built on digital events will not show them, and the absence is not evidence that they did nothing. When a channel's platform numbers look weak but total sales hold steady, the honest read is that your model is missing part of the picture, not that the channel failed.
How does the learning phase distort what you see?
On Meta, a campaign needs roughly 50 results in a week after the last significant edit to leave the learning phase. Until then, delivery and costs move around, so the numbers you attribute are still settling. Judging a model on data from an unstable campaign usually says more about the learning phase than about the channel.
The same caution applies beyond Meta. Any platform that adjusts delivery as it gathers data will produce a period where performance is not yet representative. If you compare two channels and one of them was edited last week, you are comparing a settled system to an unsettled one. Wait for stability, or at least mark the period and avoid drawing conclusions from it.
The habit that prevents most of these mistakes is simple: before you open an attribution report, confirm that every campaign you are comparing has been running without significant edits for long enough to be past its learning period. If it has not, the report is a snapshot of a system in motion.
What should you check before trusting an attribution report?
Confirm that conversion tracking fires once per real conversion, that the lookback window matches how your customers actually decide, and that the same events feed every channel you compare. If deduplication is off, or one platform counts a lead twice, the model is distributing credit across numbers that never existed.
Work through it in this order:
- Fire the conversion once. Submit a test form or make a test purchase and check how many conversion events land in each system. Two events where one happened means every downstream number is inflated.
- Match the window to the decision. If your sales cycle is long, a 30-day click window will cut off the touchpoints that started the journey. If it is short, a long window pulls in conversions that had nothing to do with the click.
- Feed every channel the same events. If one platform sees a lead event and another sees a qualified lead event, their conversion counts will never line up, and no attribution model fixes that.
- Check deduplication where two sources report the same conversion. Matching event names and identifiers is what prevents double counting; there is no switch that does it for you.
- Write the settings down. Model, window, retention, and event definitions, dated. The next person to read the report needs them.
If you want the numbers to be comparable across channels without rebuilding the setup each time, a tool that keeps spend, CPA, CTR, CPC and CPM in one place removes a layer of manual reconciliation. That is what the forecast and reporting views are for.
How should attribution fit into your measurement stack?
Treat it as one input among several, not as the answer. An attribution model is good at showing where recorded touchpoints cluster on converting paths, and at comparing how credit shifts when the rule changes. It is bad at proving cause, and it cannot see what never left a trace.
Three things make it more useful. First, keep the setup clean, because a model on broken data is worse than no model: it produces confident numbers that are wrong. Second, pair it with tests that change reality, such as turning a channel off in one region and comparing totals. Third, connect it to the rest of your planning, so that channel credit informs budget and creative decisions rather than sitting in a report nobody acts on. The broader frame for that is the marketing 4P mix, where promotion decisions depend on product, price, and place rather than on a single channel's attributed conversions.
For paid social specifically, attribution reports are most useful when read alongside how the content itself was planned, which is the subject of social media marketing strategy. And as ad inventory expands into new surfaces, the measurement questions change shape again; what to measure in OpenAI ads covers that shift.
Next step
Open your attribution report and write down four settings before you read a single number: the model in use, the conversion window, the analytics retention period, and which events feed each channel you compare. If any of the four is inconsistent between channels, fix that first. Then re-read the report. You can see how a setup like this looks when spend, conversions and cost metrics sit in one place on the forecasts and reports page.