What Is Attribution Modelling in Google Analytics 4?
When several things contributed to one enquiry, deciding which gets the credit is a choice rather than an observation. This is the subject that explains why two reports of the same period disagree, plus why search is consistently undercredited.
Somebody Chose How To Credit This
Every report you have been shown about which marketing produced your enquiries rests on a rule somebody selected. Change the rule and the same enquiries get credited to different things. That is not a flaw being exposed. It is how the subject works. Most people are never told.
Why a rule is needed at all. Journeys are not simple.
Somebody searches, reads a page, leaves, sees you again a fortnight later, returns and enquires. Several interactions, one outcome. The credit has to go somewhere.
What the tool cannot do. Split it by observation.
There is no fact of the matter about which interaction caused the enquiry. The customer could not tell you either. So a rule is applied instead.
What that makes the number. A model output.
The figure saying search produced eleven enquiries is the result of a calculation with assumptions in it, rather than a count of anything.
Why this matters commercially. Budgets follow it.
Money moves between channels on the strength of these figures. If nobody knows which rule produced them, the decision rests on something invisible.
The version most businesses meet. A pie chart.
Channels split neatly into slices, presented as though the split were measured. Every one of those slices is the output of a rule. The same enquiries would divide differently under another one.
What this page will not do. Tell you the true answer.
There is not one. What it will do is show you which rule you are using and what it systematically misses.
The Last Click Problem
The simplest rule gives all the credit to the final interaction before the enquiry. It is intuitive, it is easy to explain and it systematically undercredits everything that came earlier. That combination is why it has survived so long.
Why it appeals. It feels like proof.
They clicked that, then they enquired. The connection looks direct. The alternative requires explaining a model nobody wants explained.
What it discards. Everything before.
The article read three weeks earlier, the search that first found you, the page that convinced somebody you were competent. All present in the journey, all credited with nothing.
The pattern it rewards. Whatever comes last.
Which is usually something at the point of decision, such as somebody typing your name or clicking a branded advert. Neither of those created the demand.
Why it is not simply wrong. It is consistent.
Last click at least applies the same rule to every enquiry, so period against period comparisons hold. Its bias is predictable rather than random.
Where it actively misleads. Judging early stage work.
Anything whose job is to introduce you to people who did not know you existed will look poor under this rule by construction. That includes most content and a good deal of search work.
What follows from that. Read it knowing.
A last click report is not useless. It is a report with a known lean. Block three is the direction that lean runs.
Which Is Why Search Looks Worse Than It Is
Search work does its job early. Somebody finds you because a page answered a question, forms an opinion, then comes back later by another route to enquire. Last click reporting credits the other route, so the thing that started it appears in no report at all.
What search typically does. First contact.
It is how people who did not know you existed find out. That is the beginning of a journey rather than the end of one.
What gets credited instead. The return visit.
Which frequently arrives as direct, since somebody typed your name or opened a bookmark. Our traffic sources guide covers why that category absorbs so much.
How this compounds. Two effects, same direction.
Organic is already understated because returning visits lose their source, which our organic traffic guide covers. Then last click removes credit for the visits that were recorded properly.
What that produces. A believable underestimate.
Search looks like a modest contributor in the reporting while being the reason a large share of enquiries happened. Both effects run the same way, which is what makes it systematic.
What nobody can tell you. The size.
We publish no figure for it and neither should anybody else. The direction is knowable and the magnitude depends on your customers.
What it does to a decision. Cuts the wrong thing.
A business looking at a last click report and reducing its content or search spending is removing the introduction while keeping the closing handshake. The effect appears months later and looks like something else entirely.
Who this disadvantages. Whoever does the early work.
Which is a commercial problem as much as a measurement one. It is worth knowing before anybody judges a search engagement on a last click report.
What The Models Actually Are
Three approaches are currently available. That is fewer than most guidance describes. Four earlier options were withdrawn during 2023, so anybody following older instructions will look for settings that no longer exist.
Data-driven attribution. The default.
Credit distributed across interactions using patterns found in your own data rather than by a fixed rule. It generally treats earlier steps more fairly than last click does.
Last click across paid and organic. The straightforward one.
All credit to the final interaction, whatever kind it was. Simple to explain and carrying the bias in block two.
Last click across Google paid channels. The narrow one.
Credit confined to paid interactions with Google. Relevant only to businesses running advertising and rarely the right choice for judging overall performance.
What went away. Four fixed rules.
Approaches that gave everything to the first interaction. Approaches that spread credit evenly, weighted it by recency or split it between the ends of a journey. All withdrawn during 2023, which dates any guidance mentioning them.
The catch nobody expects. Not everywhere applies.
The model you select governs particular reporting rather than the whole tool. The standard acquisition reporting follows last click regardless, which is why a property set to the default can still show you last click figures.
What that means practically. Know which you are reading.
Two reports in the same property can disagree because they use different rules. That is expected behaviour rather than a fault. It catches experienced people.
The Lookback Window
The window decides how far back the tool looks when working out what contributed. Anything older than it receives no credit whatsoever, however important it was. Almost nobody changes this setting. For some businesses it is quietly deciding the answer.
What it does. Draws a line.
Interactions inside the window are candidates for credit. Interactions outside it are treated as though they never happened.
Who it hurts. Long buying cycles.
A business where people research for months before committing. Extensions, commercial installations, professional services, anything expensive. The first contact frequently falls outside the window entirely.
Who it does not. Emergency trades.
A boiler failing on a Tuesday produces a same day enquiry. Every interaction sits comfortably inside any window and the setting is irrelevant.
How to tell which you are. Ask your customers.
How long between first hearing of you and getting in touch. If the answer is regularly months, your reporting is losing the beginning of most journeys.
What changing it does. Alters the past reporting.
The credit gets recalculated rather than only applying forward, so figures you have already reported can move. Worth knowing before you adjust it mid year.
The related trap. Comparing two businesses.
Two companies using different windows are not producing comparable figures, so any benchmark shared between them describes two different calculations. That applies to anything you are shown about how you compare with others.
What it does not change. Enquiries you actually received.
The window governs how credit is assigned rather than how many enquiries arrived. Your total is unaffected by it, which is another reason to treat that total as the figure you trust.
What we advise. Look at it once.
Check what it is set to, decide whether it fits how your customers actually buy, then leave it alone and note the date if you changed it.
Change The Model, Change The Story
The same period can be reported very differently by switching models. Nothing about the business changes. Nobody has falsified anything. The figures simply describe a different question. This is where reporting gets flattered.
What switching does. Moves credit between channels.
A channel that looks weak under one rule can look considerably better under another, because it tends to appear earlier or later in journeys.
Why that is legitimate. Both are valid views.
Neither model is lying. They answer different questions about the same events. There is no experiment available to settle which is right.
Where it becomes a problem. Selection after the fact.
Choosing the model once you have seen which one makes the period look best. That is not falsification and it is not reporting either.
How to spot it happening. An unexplained improvement.
A channel that transforms between one report and the next with no corresponding change in activity. Ask which model each report used before accepting either.
The question worth asking any supplier. Which rule is this.
A supplier who can answer immediately is reporting carefully. One who cannot may not know that the choice exists.
What we have seen. A quiet switch.
An agency changing model between quarters without mentioning it, then presenting the resulting movement as an improvement. Nothing false was said at any point, which is precisely what makes it worth watching for.
Our own practice. Fixed and stated.
We pick a model, name it on the report and do not change it to improve a month. Block eight is that discipline in full.
What No Model Can Fix
Attribution rearranges credit among the interactions the tool observed. Anything it never saw stays invisible regardless of which model you select. For most local businesses that is where a substantial share of the truth lives.
The telephone call. The largest gap.
Somebody reading your page and ringing the number is not an interaction the tool can attribute, because the call happens somewhere it cannot see.
The conversation. Entirely absent.
A recommendation from a neighbour, a van seen on a driveway, a chat at a trade counter. All genuine causes and none of them are data.
Anything before installation. No history.
The tool knows nothing that happened before it was installed. It knows nothing at all from a device where measurement was declined.
Why this bounds the whole subject. Partial input.
A sophisticated model applied to an incomplete record produces a sophisticated answer about the part that was recorded. That is worth having and it is not the whole picture.
What actually closes the gap. Asking people.
How did you hear about us, recorded consistently. Unglamorous, cheap and more informative than any model, particularly for a trade.
Where the commercial version lives. Our measuring performance material.
That owns the attribution gap as a business problem rather than a technical one. This page owns the modelling inside the tool.
What To Do About It
Four habits. None requires understanding the mathematics. They exist to make your reporting comparable over time and resistant to being flattered, which is most of what good measurement is.
Pick a model. The default is fine.
For most businesses the property default is a reasonable choice and the selection matters less than the consistency. Deciding is the important part.
State it on every report. One line.
Name the model and the window somewhere on the document. That single line prevents most arguments about why two reports disagree.
Never change it to improve a period. The rule that matters.
If you switch models, switch permanently, note the date and expect a step in the figures. Switching to explain a bad month is where reporting stops being reporting.
Ask people how they found you. Alongside all of it.
Record the answers and compare them against what the tool says. Where they disagree persistently, believe your customers.
Watch the total, not the split. The stable number.
How many enquiries arrived is a count rather than a model output, so it survives every choice on this page. When the split between channels looks strange, check whether the total moved at all before investigating anything.
What to expect from doing this. Less certainty, better decisions.
You will report smaller and more defensible claims. That is an improvement even though it reads as a downgrade.
Where the rest sits. The hub.
Everything on this tool is on the Google Analytics guide. joining paid data is where these choices start to matter most.
The same month
can be reported
two different
ways, fairly.
Switch the attribution model and credit moves between channels without anything happening in the business. Neither version is false. Choosing which one to show after seeing the results is not falsification. It is not reporting either.
How we handle attribution:
If a channel transforms between one report and the next with no change in activity, ask which model each one used.
Every guide.
One tool.
Getting it installed and configured properly, reading the reports without being misled, measuring outcomes rather than activity, the advanced reporting and what to do when the numbers look wrong.
