Manage attribution models

About data-driven attribution models in DoubleClick Search

Gain efficiency with a more accurate view of the interactions that lead to conversions

By default, DoubleClick Search (DS) gives all credit for a conversion to the last click that leads a customer to your site. But a significant number of conversions may involve several interactions—clicks on display ads as well as paid search clicks driven by shopping campaigns, generic or brand keywords, and other biddable items. Often these interactions follow patterns, and by analyzing these patterns you may learn that some types of keywords or other biddable items tend to play a larger role in driving conversions than simple last-click attribution can reveal.

Data-driven attribution (DDA) analyzes your campaigns for interaction patterns and uses this analysis to model the contribution that each interaction usually makes to a conversion. Then DDA applies the model to the conversions in your advertiser, so you can see a more accurate picture of how clicks on keywords and other biddable items lead to conversions.

For example, DDA may learn that customers who click on your generic keywords and then later click on your brand keywords are significantly likely to convert. Last-click attribution wouldn't give any credit to the generic keywords in this pattern, but a DDA model would give some amount of credit to the generic keyword based on the patterns it has observed. In this case, if you shift some of your spend to generic keywords you can increase the likelihood of conversions while keeping overall costs flat.
A data-driven model calculates distribution of credit in a funnel.


View reports

After DS generates a DDA model, you can use it for reporting insight. For example, by comparing conversions reported by a DDA model with the last click model, you may discover that your upper funnel keywords play a larger role in driving conversions than you assumed. Not only does this demonstrate the value of the keywords, you may also be able to use this insight for more efficient bidding.

Optimize bids

While you can manually adjust bids based on performance data from a DDA model, a DoubleClick Search bid strategy can automate this process for you. DS bid strategies can leverage both Smart Bidding technology as well as a DDA model's assessment of a biddable item's value to achieve the most efficient spend for a specific ROI goal.

As with all bid strategies, comparing performance with manual bidding (or with other bid strategies that use last-click attribution) requires careful preparation and a timeline of at least a few weeks. Learn more about comparing bid strategies.

Identifying interaction patterns

DDA can automatically use interactions with your keywords to generate a model. If you want to model specific types of interactions or interactions with product groups as well as keywords and other biddable items, you can define custom channel groupings

For example, in a simple DDA model you may be interested in understanding how keywords in a "Footwear - generic" campaign work together with keywords in a "ComfyWalkers - brand" campaign to drive purchases and newsletter signups. In a more complex model, you might want to know how your generic, brand, competitive, and promotion keywords interact together to drive purchases and newsletter signups. To create this more complex model, you can create custom channel groupings.

We recommend that you define custom channel groupings if you are a heavy user of shopping campaigns or dynamic search ad campaigns because the automatic channel groupings train DDA models on keywords only.

Learn more about channel groupings and see examples for various business verticals.

Floodlight conversions only

Each DS DDA model can analyze and report on interactions leading to as many as 50 different Floodlight activities. DS DDA models cannot analyze or report on conversions tracked by AdWords conversion tracking, Google Analytics, or other conversion tracking systems (AdWords and Google Analytics have their own attribution models that you can apply to AdWords or Google Analytics data).

Depending on what you're trying to learn, you may want to create a model that includes only Floodlight activities related to a specific conversion funnel. However, make sure you include enough Floodlight activities to provide DS with sufficient data to generate the model

For example if you have two Floodlight activities that are not in the same funnel, such as "Retail purchases" and "Wholesale inquiries", you could do either of the following:

  • Create two separate DDA models, one for the "Retail purchases" funnel, and the other for the "Wholesale inquiries" funnel.
    This is the best practice if you think the conversion funnels are different enough from each other that credit would be distributed differently. For example, create separate models if you think the two funnels mostly look like the following:
    Generic -> Brand -> "Retail purchases"
    Generic -> Geo -> Competitor -> Brand -> "Wholesale inquiries"
  • Create one DDA model for both funnels.
    This is the best practice if you think the conversion funnels are fairly similar or close enough for your purposes.

Or you can create three models, one for each funnel and one for the combined funnels, and compare them.

Accounting for unattributed conversions

If you use Floodlight iframe or image tags instead of the global site tag or Google Tag Manager, DoubleClick Search is not able to observe conversions when the website cookies that store information about your ad clicks aren't available due to factors including browser settings. In these cases, Floodlight uses scaling to account for the conversion data that can't be directly measured. That is, Floodlight uses machine learning and historical data to model the number of conversions and amount of conversion revenue that cannot be measured. Floodlight can then add the resulting estimate to its conversion metrics to provide a more complete picture of how your advertising drives conversions. 

DDA models include these scaled conversions both when generating a model and when applying the model to a Floodlight column. 

Which advertising channels are considered part of a conversion path?

When examining conversion paths, DS DDA analyzes clicks from the following channels:

DS only reports on clicks that are redirected through DS. That is, a DS DDA model might attribute partial credit to a display click managed by Campaign Manager and the rest of the credit to a paid search click. But because DS only reports on paid search clicks, when you view a report, you'll only see the partial credit attributed to the paid search click.

Note that just like other attribution models used in DS, a DS DDA model ignores impressions (both search and display). DS doesn't have access to data about display and natural search impressions, so only clicks can be considered as interactions that potentially lead to conversions.

Data requirements

As a general guideline, DDA needs 15,000 clicks and 600 online Floodlight conversions during the last 30 days to successfully generate a model.

For example, if you create a model with the following data:

  • Two channel groupings: one for Generic keywords and another for Brand keywords
  • All of the Floodlight activities in your retail funnel

You'll need a total of 15,000 clicks on the Generic and Brand keywords in your channel groupings and a total of 600 conversions from the Floodlight activities in your retail funnel, all during the last 30 days.

DS can start preparing the model as soon as you receive the minimum number of clicks and conversions over a period of 30 consecutive days. The initial learning period lasts about 24 hours, and the model will be applied to the previous 60 days of performance data (if available), as well as to all performance data going forward.

Offline conversions

If you include offline Floodlight activities when you create a model:

  • If you use DS bulksheets or the DS API to upload the conversions, DDA ignores the offline conversions when generating or updating the model. That is, offline conversions uploaded from bulksheets or the DS API do not count towards the data requirements.
  • If you use the DCM API to upload the conversions, DDA includes the conversions when generating or updating the model. That is, offline conversions uploaded from the DCM API count towards the data requirements.

Once the model has been generated, you can apply the model to some of the offline conversions (regardless of how you uploaded them) by including the offline Floodlight activities in your DS reports or in a bid strategy. Note that the DDA Floodlight column will only show credit for offline conversions that you have uploaded no later than 24 hours after occurrence.

The DDA model learns continuously

The DDA model continues to learn and updates the model each week. Any new campaigns, keywords or other items you add to your channel-grouping labels will be reviewed and incorporated into the model. That is, the attribution can continue to change. Note that updates to the model only apply to new performance data going forward. DS does not apply the updates to historical data (conversions that occurred before the model was updated). 

If the weekly number of clicks and conversions falls below the data requirements, DS uses a previously generated model. If DS hasn't been able to generate a model within the previous 30 days, the DDA model applies the basic linear model and will retry the update next week.

Model updates are typically subtle, and the effect on reporting data is gradual, since the updates apply only as new conversions are reported.

If you create a bid strategy that uses a DDA model, during the bid strategy's learning period you might see a bit of volatility in the DDA model as both the bid strategy and DDA model adjust to each other. Usually, you'll see fewer updates to the DDA model after the bid strategy's learning period completes.

The model is applied once a day

Once a day, DDA applies its current model to the data reported during the pevious 24 hours. Depending on when you view a report, data in a DDA column from today or even yesterday may not reflect the DDA model yet. Similary, bid strategies using a DDA model take a day before seeing the model's effects on the latest activity.

Create and compare multiple models

You can create up to five DDA models in DoubleClick Search. Consider starting out with a couple of different DDA models with different channel groupings and compare the insights you gain from each.

After you create the model, compare data from the model with other models. For example, create the following Floodlight activity columns that report on the activities in your model:

  • One column that uses the DDA model you created (If you created multiple DDA models, create one column for each model) 
  • Once column that uses the default last-click attribution model

Add the columns to a reporting table and compare the data in each. You may learn that some of your long-tail keywords are playing a larger role in driving conversions than you thought.

If you find a model that you think accurately reflects your business and business goals, use the model in a bid strategy to achieve your goals with the optimal spend.

Include cross-environment conversions

To get the most complete view of conversion paths, include cross-environment conversions when you report on a data-driven attribution (DDA) model or use a DDA model in a new or existing bid strategy that adjusts mobile bids.

Cross-environment conversions start on one device or browser and end in another. For example, a customer may click a search ad on a mobile phone,  then later use a desktop device to directly access an advertiser's site and convert. By default, DS only reports conversions that occur on a single device or browser. But by leaving out cross-environment conversions, you may be undervaluing the role some of your ads play in driving conversions. A default report in DS—even a report that contains a column with a DDA model—wouldn't count the conversion in the example above because the paid search click and the conversion occurred on separate devices.

The best practice is to include cross-environment conversions in your DDA Floodlight column. You can also create two DDA Floodlight columns: one with cross-environment conversions and another without. Then add both columns to a report and compare the difference. See more detail about which DS features support DDA models and cross-environment conversions.

Cross-environment reporting is available only at the ad group level or higher. So if you include cross-environment conversions in your DDA Floodlight column, the column only contains data when you view it on the Ad Groups, Campaigns, or Engine Accounts tab. If you add the column to the Ads or Keyword tab, the column contains zeros. 

How soon do cross-environment conversions appear in a DDA Floodlight column?

Within 24 hours of creating a DDA model, a DDA Floodlight column that has enabled cross-environment conversions will start showing the effects of cross-environment conversions. The column does not include cross-environment conversions that occurred before you created the model.

How do DDA models in DS differ from DCM?

If you use DoubleClick Campaign Manager (DCM) to manage display ads, you can also create a data-driven attribution model in DCM. The model you create in DCM is completely separate from DS models. It cannot be imported into DS, so it can only be applied to performance data in DCM.  

Both DS and DCM models analyze the following types of events:

DCM models also analyze the following types of events:

  • Paid search impressions
  • Display impressions

DS doesn't analyze impressions.

While the DCM DDA model can show the number of whole or partial conversions attributed to paid search and display activity, DS only reports the number of whole or partial conversions attributed to paid search. For example, if a conversion path starts with a paid search click but also includes display activity, DS will report partial credit for the paid search click.

Ready to get started?

  1. Apply labels at least 12 hours before you create a DDA model. Otherwise, the model might distribute credit to the "Unknown" channel grouping. 

  2. Create a data-driven attribution model.
  3. Use the model by doing any of the following:
Was this article helpful?
How can we improve it?