FLoC (Federated Learning of Cohorts) is an interesting bird: mysterious and on the verge of extinction. If it disappears without a replacement, then along with the end of 3rd party cookies, advertisers may see a dramatic shift in the effectiveness of interest-based audience targeting. So despite recently announced delays in the retirement of 3rd party cookies in Chrome, it’s worth knowing what FLoC is all about. That’s exactly why I hosted a session on the topic at SMX Advanced earlier this year. Here are a few of the takeaways of the session.
What is FLoC
FLoC is one of Google’s proposed solutions that would allow advertisers to target audiences in a more privacy-sensitive manner. As most advertisers already understand, a lot of audience targeting technologies depend on the existence of 3rd party cookies and those are notoriously bad for privacy.
While FLoC doesn’t address remarketing, there is a separate proposal for that called Fledge but that’s for another post to dive into.
FLoC is wrapping up its trial run through the Privacy Sandbox effort led by the Chromium team at Google. Chromium is the open-source code base for browsers maintained by Google and the basis for the Chrome browser which commands over 64% of the worldwide browser market.
Why we need an alternative to 3rd party cookies
Apple, Firefox, and others without a stake in online advertising have already started to block 3rd party cookies. Google, with its interests in online advertising, is taking a more measured approach and proposing solutions like FLoC before they also turn off 3rd party cookie support in their browser.
In a 2019 study by Google, news sites would lose 62% of ad revenue if 3rd party cookies were retired without a replacement for audience targeting. Google obviously has a horse in this race, but the contract between users and sites of free access in exchange for seeing targeted ads is also at stake. And if that contract is broken, publishers and consumers will suffer along with ad platforms like Google.
How FLoC protects user privacy
There is a lot of discussion about the ways FLoC may not be a good enough solution but let’s get to some of the basics first.
In traditional audience targeting, advertisers rely on 3rd party cookies to share data about what individuals do on the web. For example, visit any site on Apple Safari and look at how many 3rd party trackers are blocked. These trackers are cookies not owned by the website you’re visiting. For example, while visiting cnn.com, adnxs.com is using a cookie to track me even though I have no relationship with adnxs or any of the other trackers like it.
What FLoC proposes is a way to anonymize users by adding them into cohorts using a federated learning methodology. This grouping of users into cohorts happens in the browser so no personal data leaves the device. Then the cohort is made available for publishers who wish to continue showing targeted ads to their users.
In the simplest terms, instead of the website asking the ad network to serve an ad to Fred who spends way too much on workout gear, the website can ask the ad network to show an ad to a cohort (with at least 1000 anonymous users) who share some similarities to one another, like a passion for fitness.
So no user data leaves a person’s device under the FLoC proposal. That is unquestionably more private than what happens with 3rd party cookies and fingerprinting that is prevalent today.
Will FLoC work for advertisers
In experiments done by Google Ads where FLoCs were simulated and used in place of 3rd party cookies, they found FLoC delivered around 95% efficiency compared to the audiences using 3rd party cookies.
So it’s pretty good, but even if we wanted, we cannot use FLoC for targeting. As of right now, we can only observe. It’s pretty similar to how audiences for observation or targeting works in Google Ads today. We just have the observation capability because we can ask our Chrome browser to identify its cohort but we can’t act on any performance-related findings by managing cohort placements in Google Ads.
So if FLoC cohorts aren’t of any practical use at this stage, why are we talking about it? My hope is that by learning some of the background and underlying technical details, the advertising community will be better positioned to quickly deploy new solutions and engage in conversations that will ultimately decide how audience targeting will work in a privacy-first world. Let me cover one reason I am excited about a future world in which we might be able to target cohorts.
New opportunities using cohorts
Until now, which audiences you can target has relied mostly on how ad platforms define those audiences. They build audiences by monitoring the behaviors of individuals across sites and grouping those individuals into the appropriate lists.
For example, here’s how Google thinks about me:
You can see how they’ve classified you at https://adssettings.google.com/.
Note that as an individual, I am part of many audiences, e.g. I care about 3d printing and athletic apparel.
But FLoC is not a Google Ads project but rather a Chromium project with open source code. This means any website can find out what cohort a visitor belongs to by asking the browser. It takes literally a single line of code, the first one in this screenshot:
The site can start to map users to behaviors and find interesting commonalities. For example, users in a particular cohort tend to convert at a higher than average rate. It is conceivable they could at some point ask the ad network to target these desirable cohorts. What’s also cool is that these observations can come from organic traffic since FLoC is independent of any of the ad platforms.
How are cohorts formed?
In Chrome’s first experiment (called origin trial) of FLoC, cohorts were generated by using a publicly known algorithm (Simhash). The input is a list of domains a user has visited and the output is a cohort. Users with a similar list of visited domains will be mapped to the same cohorts. The Simhash algorithm allows for controlling how many cohorts can exist and that’s currently capped at 34,000 in the current implementation.
To ensure no cohorts are too small which would reduce the privacy of the system, there is a separate system that ensures k-anonymity. Basically, whenever an individual is added to a cohort, the Chromium team keeps track of how many users are in that cohort. And until there are 1,000 individuals in a cohort, it cannot be used for targeting.
To ensure cohorts are not based on visits to sensitive sites, Chromium excludes certain sites from being part of the input to the algorithm. And site owners can also set a flag in their website code to be excluded from the inputs into the hashing algorithm. This, by the way, is what one of WordPress contributors suggests the popular CMS tool should do by default. If implemented, this would be one of the issues threatening the viability of the FLoC proposal.
A big difference is that a user can only be in a single cohort unlike in the past where an individual was on a lot of lists based on a multitude of interests. But the ad platforms could map one cohort to many interests. For example, if my cohort tends to read content about baseball and cooking, my cohort might be included in ad tech’s mapping for sports enthusiasts and cooking enthusiasts). These cohorts are updated weekly, and can only change at that time.
Cohorts are specific to one web browser. Even if a user is logged into Chrome, the data is not shared across devices so your work computer may be in a different cohort than your home computer. This could actually be a great thing for B2B advertisers who would benefit from being able to target a person based on their “work cohort.”
How to use cohorts
If someday we can target cohorts in Google Ads or elsewhere, how do we get those cohorts? You could look to adtech platforms to help with generating probably cohort IDs based on behaviors like what domains a user visited.
While this is not a full implementation, Optmyzr (my company) built a proof of concept to show how Simhash works. You add a list of domains and get back the simhash key: https://swiy.io/get-FLoC-id
It’s similar to custom audiences in Google where you pick from a list of behaviors to create an audience. You could do the same to create a cohort ID for targeting.
The other way to find audiences to target (once targeting is allowed) is to use observation to build an understanding of cohort behaviors. Adtech platforms may monitor what cohorts over-index on visits to particular types of sites. Remember any site can ask for the cohort of a user. So any company working with many publishers can start to map cohorts to interests.
For example, an ad network could observe that home improvement sites over-index on visits from users in cohort 123. They can then let advertisers specify they want to advertise to users interested in home improvement and know that they should show ads to cohort 123.
By the way, this is how it will likely also work in Google Ads. Google will do the mapping of cohorts to interests for its advertisers so they can continue to target the same audiences they do today. In the backend, Google will simply swap out individual users for cohorts. This is exactly what they did in the experiment mentioned above that had 95% efficiency.
Google delays 3rd party cookie sunset to 2023
So what does Google’s announcement that the sunset of 3rd party cookies will be delayed until late 2023 mean for FLoC? They said, “We received substantial feedback from the web community during the origin trial for the first version of FLoC. We plan to conclude this origin trial in the coming weeks and incorporate input, before advancing to further ecosystem testing”. Is FLoC dead?
Any answer to that question would be speculation but what I do know is that the current iteration will not be the one that launches. How much it will change is unknown. In all fairness, that was my assumption and a frequent disclaimer at my SMX session all along.
From the ads perspective, the best we can do is continue monitoring FLoC or its new replacement and experiment with it so that we are ready to make the most of it the day 3rd party cookies stop working. I have a feeling even Google’s Ads teams are continuing experimentation as they are not necessarily more clued in to what the W3C and Chrome teams will do than anyone else who attends their public meetings or reads the minutes. In reading the W3C meeting minutes, the Chrome team states they’ll announce in those meetings when they intend to make any change, so it’s a good place to keep an eye on.
FLoC is an evolving topic
Hopefully you now better understand FLoC and what it’s trying to achieve. If you still have lots of questions, you’re not alone. This is an evolving topic but one worth keeping a close eye on.
But rest easy, even if there is no replacement for 3rd party cookies, this won’t necessarily be the end of interest-based audience targeting because ad tech companies like Google already use signals beyond 3rd party cookies to create audiences. It’s just that they would lose one powerful signal and as a result, advertisers would have to more closely monitor any resulting shifts in performance.
And remember when 3rd party cookies are phased out, even in the absence of a replacement, advertisers will still be able to target their own first-party data on Google Search and YouTube, target audiences based on Google users’ consented activity on Search and YouTube, measure click and view conversions from Search and YouTube by linking advertiser first-party data to Google and measure click conversions from Search, YouTube and partner inventory by linking clicks to advertisers’ sites.
To give credit to the moderator of my session at SMX, the always affable Matt Van Wagner said, “Personally, I don’t think FLoC will thrive and may die in the nest before it fully hatches due to regulatory and competitive pressures from Microsoft, Apple, and objections from WordPress and others. Too Google-centric, IMHO, to become a standard.” And less than two weeks after my session, Matt may have been proven right about FLoC. But the problem it is trying to solve remains unsolved so I bet there will be many more sessions like mine at future conferences.
Opinions expressed in this article are those of the guest author and not necessarily Search Engine Land. Staff authors are listed here.