Single-touch models are one of the fundamental attribution modeling methods, still popular due in part to their relative simplicity. How do they work, and are they really that simple?
Previous article (part 6): Comparing attribution models
In single-touch models, 100% of the conversion credit is assigned to a single interaction, e.g. the click that took place before the conversion, while the remaining interactions are ignored — unlike multi-touch models, which potentially take into account all the interactions that occurred on the path within the lookback window.
In 2024, Google Analytics introduced a terminology change and what were previously called conversions in Analytics were renamed key events. Some illustrations in this article, created earlier, may point to key events in Analytics as conversions. In this article, the use of the word “conversion” (in the general sense) will not always match Analytics terminology.
The two basic types of single-touch attribution are the models:
- first interaction
- last interaction
They assign all of the conversion credit to the interaction that was, respectively, the first or the last one on the conversion path — within the limits of the lookback window (conversion window).
It’s worth noting right away that, unlike the last interaction model, the first interaction model is sensitive to changes in the lookback window. The last interaction is in principle always the same, whereas what counts as the first interaction depends precisely on how far back in time we reach. By shortening the conversion window from the example above, we can make Google Ads, rather than Direct, the first interaction.
The First user source dimension in Google Analytics assigns key events and revenue to the source from which the user visited the site for the very first time in their entire history. It is therefore the absolute first source of the visit, and the lookback window does not apply in this case (in a sense, the lookback window is “infinite”). See also the article on attribution in Google Analytics.
Assisted conversions
The last interaction model is linked to the concept of assisted conversions and assist interactions (clicks).
An assist interaction is any interaction that is not the last one on the path leading to a conversion, and which therefore was not credited with the conversion by the last interaction model. So on the path we have the last interaction before the conversion and the remaining interactions — and it is these that we call assists. A special case of an assist interaction is the first interaction.
For each traffic source, assisted conversion reports show the conversions after the last click alongside the assisted conversions, in which that source appeared at other positions on the path. This gives you an idea of what the last click report doesn’t tell you about the role of a given source, and how significant that may be.
Keep in mind that last click conversions and assisted conversions must not be added together. This is because a given source can be both the last click source and an assisting source on the path of the same conversion:
On the path above, Organic is the last click, but it is also an assist click. For this conversion, Organic will therefore appear both in last click conversions and in assisted conversions, even though both concern one and the same conversion.
The sum of last click conversions and assisted conversions will therefore contain duplicated conversions.
For the same reasons, assisted conversions should not be added to each other either. In the example above, this conversion will appear as assisted for Direct, Organic and Google Ads.
The term assist interaction itself is not entirely accurate, because it suggests that the interaction assisted the conversion process, when all we actually know is that it occurred on the conversion path — we don’t know whether it had any influence on it at all. A more accurate term would be accompanying interaction.
Modifications of the last interaction model
Single-touch models can be modified by defining additional rules that determine which interaction the conversion will be attributed to — meaning it will sometimes not be the absolute last interaction, but the last interaction that meets a specific condition.
The last click model
In the last click model used in Google Analytics reports, 100% of the credit for key events and revenue is assigned to the last non-direct visit (so it is not the absolute last visit). It is sometimes called the last non-direct click model, but formally the adjective “non-direct” is not needed, because in Google Analytics terminology a direct visit is not a click.
The last click model is also used in Google Analytics 4 to determine the session source in traffic acquisition reports — see the article on attribution in Google Analytics.
When we see a visit from organic search results in a traffic acquisition report, it may just as well be a direct visit made by a user whose earlier visit to the site came from organic search results:
This means that traffic reported in Google Analytics as (direct) / (none) concerns users who, within the lookback window, visited the site exclusively through direct visits. In reality there are usually far more direct visits, but they are assigned to earlier traffic sources.
The situation is analogous for the attribution of key events: if a key event in Google Analytics is attributed to direct traffic, it means there were no interactions on its path other than direct visits. Changing the attribution model won’t change anything — there are simply no other interactions on the path.
Last click (Google paid channels)
The last click model can be modified so that it assigns key events and revenue to the last interaction from a selected channel, e.g. Google Ads. In Google Analytics this model is called Last click with the annotation Google paid channels, and it first tries to assign the key event to the last Google Ads click (unlike the other models, which consider all non-direct interactions and carry the annotation paid and organic channels).
If there is no Google Ads click on the path, this model behaves like the (non-direct) last click model and assigns the key event to the last interaction that is not a direct visit. Only when the path contains nothing but direct visits will the key event be attributed to (direct) / (none).
This model comes in handy if we want to compare conversions reported in Google Ads with Analytics reports.
Google Ads, like any other advertising system, reports conversions by attributing them exclusively to interactions with its own ads. All other interactions are invisible to Google Ads, so as long as a conversion happens after an interaction with Google Ads, the Google Ads system will take full credit for driving it. And it makes no difference whether the click on a Google ad was just one of many different interactions.
When there are more Google Ads interactions on the path, the internal Google Ads attribution model (e.g. the data-driven model) will distribute that conversion across the interactions on the path — but only (naturally) across Google Ads interactions.
The total number of conversions reported in Google Ads should be close to what Google Analytics assigns in a model that prefers Google paid channels (e.g. the last click Google paid channels model).
It’s worth making sure that both reports use the same reporting time (the default reporting time in Google Ads is the interaction time; in Google Analytics — the time the key event occurred, but in Google Ads we can add a “by conversion time” conversion column, and in Analytics — choose the appropriate setting of the attribution modeling tool.
In practice we observe certain differences, resulting from the different behavioral modeling in Google Ads and Analytics (data on users who use multiple devices, did not consent to tracking, or had tracking blocked for other reasons). Differences can also arise from the different treatment of conversions after engaged views (which Analytics can also report, but they don’t make it into all reports — see the Google Analytics help article).
Regardless of the differences that occur, the last click Google paid channels model (Google Ads last click) will be the closest to what Google Ads shows, at least in total for the entire account or MCC, if the advertiser uses several accounts.
Last click vs. Google Ads last click
In Google Analytics terminology, the title of this section would read “Last click (paid and organic channels) vs. Last click (Google paid channels)”.
By comparing these two models in Analytics we can see that the difference between conversions in Google Analytics and those shown by Google Ads in its reports is the result of overwriting the traffic sources for those conversions where, between the click on a Google Ads ad and the user’s conversion, there was at least one more visit by that user from another non-direct source.
Above: a screenshot from Google Analytics (modified for readability). Both compared last click models are “non-direct”, i.e. they avoid attributing conversions to direct, while the second model prefers Google paid channels. You can see that google / cpc appeared on the conversion path 40.25% more often than the last click (paid and organic channels) report would suggest — in that report, conversions after a Google Ads click were “captured” by, among others, google / organic and instagram.com / cpc. This shows which channels users interact with after clicking Google Ads, and at what scale. Because both models are non-direct, direct is reported only in the absence of other interactions on the path, so its share remains unchanged regardless of the attribution model used.
Modifying models and additional rules for the fallback model
The remainder of this article (updated in January 2024) contains references to solutions used in Universal Analytics (the previous version of Google Analytics). Attribution currently works the same way in Campaign Manager 360 and Search Ads 360 (Google Marketing Platform tools).
Modified single-touch attribution models, which assign conversion credit to an interaction meeting specific criteria (e.g. Google Ads last click), have to be prepared for the situation where no such interaction is on the path. In that case, simply assigning the conversion to the last interaction won’t necessarily match the intent of the model — for example, it would attribute conversions to impressions or direct visits even though there are interactions on the path that would “deserve it more”.
That’s why an additional rule would come in handy — one that applies when the model’s preferred interaction is not found on the path. Without it, comparisons with other modified last touch models (e.g. last non-direct click) can be difficult.
This is exactly what the additional rules for the fallback model are for. Behind this name hides a feature that lets you create attribution models that are meaningful to compare with the last non-direct click model.
A non-direct model
Here is what the definition of the Google Ads last click model looks like:
If no interaction on the path meets the rule’s requirements (in this case, if no interaction comes from google / cpc), the conversion will be attributed to the last interaction.
Clicking “Edit” lets you create the fallback model rules. Credit assignment will then work like this: if no interaction on the path meets the requirements of the primary rule (here: none comes from AdWords), the conversion will be attributed to the last interaction that satisfies the fallback model rule. In our case, we want it not to be a direct visit:
This way, only if no interactions meet the fallback model rule either (in this case: if there is no interaction that is not a direct visit), will the conversion credit be assigned to the last interaction regardless of what it is (in this case, a direct visit).
We have thus created the Google Ads last click (non-direct) model. It will show the same number of conversions attributed to paid search as the Google Ads last click model, while at the same time preserving the principle of not attributing conversions to direct traffic if there is any other source on the path. As a result, the number of conversions attributed to direct traffic will be the same as in the Last non-direct click model.
A model like this illustrates much better which channels overlap with Google Ads. In the example below you can see that the only channel overlapping with Paid Search is Organic Search, to which 25 conversions moved in the last click model (for the remaining channels the change is 0%).
Currently (January 2024) all attribution models in Google Analytics are “non-direct”, i.e. they avoid attributing conversions to direct visits, so the fallback model excluding direct visits is built into every attribution model.
The fallback model is also available for models based on the first click. It seems that it should be used in essentially every model that modifies the rules of assigning conversions to the last click, since attributing conversion credit to direct visits not only provides no information about the attribution of individual marketing activities, but actually distorts that data.
Avoiding attributing conversions to impressions
The fallback model can also be applied to models that exclude interactions other than clicks (e.g. display and video ad impressions) from receiving conversion credit in Campaign Manager:
The Non-Big Direct model
In an earlier article we presented an example channel group in which direct traffic and brand-related searches were combined into a single “Big Direct” channel. This reflects the assumption that we don’t want to give this traffic conversion credit if we have other identified sources on the conversion path.
Following this direction, we might want to create models that won’t assign conversion credit not only to direct visits, but also to visits coming from searches for brand-related terms. Below is an example configuration of the Last Non-Big Direct model:
In this model, the conversion will be attributed to the last non-direct visit that is not a visit from paid brand search (in this case the brand is, for example, “adequate”). This makes it easier to determine which channels actually deliver customers, rather than merely helping people who are already looking for our store to reach it.
If brand searches account for the majority of conversions in organic search results, it makes sense to include medium = organic in Big Direct as well (see the article on channel groups). In that case, a condition excluding medium = organic should also be added to the attribution model.
If the path contains only direct visits, the fallback model will attribute the conversion to the last (non-direct) click, if there is one on the path. This condition avoids unnecessarily inflating direct visits in source / medium reporting.
If the attribution model comparison is carried out on a channel group containing Big Direct, this won’t matter, because a conversion that doesn’t meet the Last Non-Big Direct model’s condition will be counted towards Big Direct anyway.
Analyzing the cannibalization of non-brand and Google Ads campaigns by brand campaigns
Google Ads accounts often run campaigns on keywords related to the advertiser’s own brand (brand campaigns). Their role on the conversion path is decidedly different from that of the other campaigns (non-brand campaigns), and if a generic keyword appears on the path followed by a brand keyword, we will probably want to assign the larger share of the conversion value to the generic keyword — and we certainly wouldn’t want the brand keyword to take all the credit.
Using custom models, we can easily assess in what share of Google Ads paths a brand keyword follows non-brand keywords and “takes away” their conversion in the last click model. We will see how much conversion tracking in Google Ads based on the Google Ads last click model would change the distribution of conversions if brand keywords were skipped and all of the credit went to searches for non-brand keywords and to other campaigns, e.g. display.
To do this, we will compare the Last Google Ads model
— with the Last Google Ads Non-Brand model, in which the fallback model will first try to attribute the conversion to Google Ads (in this case, brand campaigns are what remains) — this way we ensure the same number of conversions is attributed to Google Ads in both models:
Comparing these two models on a channel group that splits Google Ads campaigns into brand and the rest, we can see significant differences:
In this example, brand campaigns significantly “strip” conversions from the other campaigns. The value of generic keywords would be 64.43% higher in the Google Ads last click model if it weren’t for the brand keywords that follow them, and display, including remarketing, also showed significant differences.
In this situation it might be justified to separate generic campaigns into a dedicated account with separate conversion tracking and use those conversions for goal optimization, since the current configuration significantly understates their value and may thus prevent them from reaching their full potential.
Our campaigns are most likely too conservative, and increasing spend on non-brand keywords may allow sales to grow while maintaining the right campaign profitability.
Currently the default attribution model in Google Ads is the data-driven model. Experience shows that it doesn’t always properly handle the role of brand keywords on the path, and an analogous analysis should lead to similar conclusions.
Remarketing vs. Prospecting
Like brand, remarketing plays a different role on the conversion path than prospecting activities. Yes, remarketing does increase conversions, and its effectiveness can be very high. Nevertheless, remember that without the other marketing activities that originally bring users to the site (prospecting), this channel would have no reason to exist.
As with paid brand keyword campaigns, this budget can be treated separately from the budget devoted to acquiring new customers. Excluding remarketing campaigns from conversion credit (not just AdWords remarketing) allows an even better assessment of customer acquisition costs.
To do this, the Last Google Ads Non-Brand model needs to be modified to exclude remarketing campaigns alongside brand campaigns (the Last Google Ads Non-Brand -Rem model). In our example we will see that the share of generic keywords is even higher (+77.74% vs. +64.43). The share of prospecting display campaigns also grows (+46.38% vs. +38.13%).
If mailings sent to registered users (i.e. people who have already visited the site and left their e-mail address for marketing purposes) account for a significant share of our traffic sources, we can also consider creating a model in which this traffic is likewise excluded from conversion credit — for the same reason we excluded remarketing.
With some marketing automation activities we are de facto dealing with remarketing, for example when an e-mail is sent to a user after they abandon their shopping cart.
Comparisons like these give us more information about which sources close the customer acquisition process and how profitable they are.
Next article (part 8): Multi-touch models
Worth reading: A guide to attribution in Google Analytics
