Google Discover is an important step in the evolution of Google from a Search to an “Answer Engine”. It highlights the transition from intent-driven pull to behavior-driven push results. I ran a little n=1 experiment on Discover to understand how following entities in Search changes the results. I learned more than I expected: how Google creates a profile of users based on the Topic Layer together, how Google tracks users across its ecosystem, and what brands might do to increase Discover traffic.
Discover is part of 3 big shifts that Google announced in 2018 and that mark the transition from search to answer engine. It’s is the embodiment of the second shift: from queries to recommendations.
One thing that’s important to understand is that Discover sits on top of the Topic Layer, Google’s evolution of the Knowledge Graph. The Topic Layer understands people’s interests and the relationship between topics like opinions, trends, news, evergreen content, etc.
It learns from your “web activity”: everything you do across all 53 Google products.
This little experiment is more anecdotal than scientific. Here’s how. I set it up:
- I started taking screenshots of my Discover Feed on 7/22
- On 7/26, I followed a couple of entities:
- League of Legends
- Gary Vee
- The Verge
- Joe Rogan
- Then, I monitored my Discover Feed for another 7 days
- The whole time didn’t click on any Discover card
- I took screenshots every day
The original idea was to measure the impact of following an entity in Search. How focused would Discover cards be around topics and publishers I follow? I knew it does have some impact but I wanted to learn more.
“While we’ve been getting better at understanding your interests, it hasn’t always been easy for you to choose new topics for your feed. To help you keep up with exactly what you care about, you’ll now be able to follow topics, right from Search results. Look out for a new “Follow” button next to certain types of search results—including movies, sports teams, your favorite bands or music artists, famous people, and more. A quick tap of the the follow button and you’ll start getting updates and stories about that topic in your feed.“
See if following entities in Google Search had an impact on Discover. In the process, I learned a lot more about how Discover works and how Google creates a profile of me. In return, this helps me understand how to get more traffic from Discover.
Side note: I did mess up one screenshot. On 7/24, I only captured the upper part of the discover feed but I reconstructed it with my Web Activity (more in a second).
I then tracked the Discover cards in a spreadsheet to see which ones are recurring (see the colored ones).
I was surprised that you cannot follow every publisher. For example, you can follow The Verge but not The New York Times.
Apart from that observation, I spotted a couple of patterns.
A month before the experiment, I did a lot of research around Elon Musk for an article I wrote. I specifically googled queries like “Elon musk outrageous tweets” and visited related search results.
During the experiment, Discover showed me content specifically about “Elon outrage”, for example how he says that Warren Buffet isn’t the kind grandfather he plays to be in public. It seems that Discover understands not just what entity I’m interested in, but also the context.
Google really wants me to read a Forbes article about the new MacBook. It teased me the same Forbes article 6 times over the 12 days! Just one other topic was teased as often as this one (coming next). But it makes sense.
On July 19th, I googled “Apple Michigan Avenue” and “apple store Chicago”. A day before, I googled several queries around Apple’s new MacBook, and a month before, I watched a video on Youtube about Apple’s design philosophy. No wonder Google thought I really wanted to see this.
FAZ – Riots in Frankfurt
The only other topic next to “Apple” that was shown to me 6x were riots in Frankfurt, Germany (“Krawalle in Frankfurt” in German) by the FAZ. Part of my family lives there and I visited this year, so maybe that’s why.
Digging through my Search feed, I did in fact increase my search activity around the entity “Frankfurt” around July 12th. A week later, Discover started showing the card for the article and did so persistently throughout the experiment.
One interesting point is that the article was in German. Google announced that Discover might show cards or results from different languages.
At the beginning of the experiment, I saw a lot of content around the Microsoft Surface Duo, which also fits to my search history. 2 weeks before seeing several Discover cards over the course of the 12 days, I searched for queries around Microsoft’s acquisition of Beam. On July 14th, I send 5 queries around “Microsoft” within a couple of minutes.
But I noticed an interesting switch in topics on 7/30 from the Microsoft Surface to Microsoft Edge. Looking at Google trends (see screenshot below), you can see that “microsoft edge” had a higher interest than “Microsoft surface” for a few days before it reversed.
I think Discover knew I’m interested in Microsoft but wanted to see what happens when they tease me a different subject within the same topic to get me to click.
I certainly saw more cards from The Verge after following their entity. I only had them in my feed once before, then 5x in the 7 following days. To be fair, I am a regular The Verge reader (you might have noticed) but the data shows that I did see a clear increase after following their entity.
Is Discover’s shelf-life really 3-5 days? I don’t think so. To be very fair here, we’re all very early in Discover research and I myself observed a shelf-life of 3-5 das a couple of times before. However, I also found counter-evidence in the Frankfurt Riot article and Forbes’ Apple article. Both were shown to me over 12 days! The open question is whether they would’ve shown if I clicked on them or not.
There was only one day on which I saw no repetitive result. Every other day had at least one, most of the time between 3 and 4.
I also noticed that when I search a bout of queries around the same topic within a short time frame, I see more Discover cards roughly 2 weeks later. That turned out to be the case with Frankfurt, Joe Rogan, and Microsoft.
Lastly, following entities did make a difference but not in every case. As I mentioned, I did see more Joe Rogan and The Verge content after following their entities in Search. But I didn’t notice anything for the other four 4 entities I followed.
That wasn’t good enough for me. I wanted to learn more, so I exported my Web Activity to see if I could find any signals that would give Google hints.
What Web Activity tells about Discover
First, it’s not so easy to export all the data Google has about you. For my first try (exporting everything) resulted in a package of 210gb.
After filtering the 53 Google products to the ones that likely made an impact, I stuck to mostly Search, Youtube, and News data.
Within Search, you also see what Discover Cards Google shows you (see screenshot below). That was helpful!
That allowed me to get an insight into how Google classifies the entity for each article. VERY interesting! As you can see from the screenshot above about my Discover feed on 7/25, Topic Layer entities range from specific brands (Zoom) to countries (China) or people (Elon Musk) to activities (cooking).
That made me wonder how different classifications might repeat over the course of the experiment.
I exported the Discover cards from my Activity feed for each day and looked at the repeating ones.
As you can see in the result below, Google cycled different articles through the same types of entities. It would show me a Technology-related card every day (except for one day) but different articles.
Now I was curious about what topic Google thought I was most interested in. Turns out it’s technology (accurate). Here’s the list of entities Google showed me articles for over the course of the 12 days:
|Search engine optimization||2|
|Software as a service||1|
When realizing that Google most often shows me technology-related Discover cards, I checked my search activity around that topic and found that I had visited articles in the /technology/ subdirectory of publishers like CNN or The NY Times a couple of days before. So, again, a simple signal that fits right in.
The last piece was figuring out how following the 6 entities changed my feed. As you see, Joe Rogan did increase and so did Technology, Business, and Coronavirus.
|Interest||Occurences||Before 7/26||After 7/26|
|Search engine optimization||2||1||1|
|Software as a service||1||1||0|
Before following, I saw The Verge once in my Discover feed. Afterward, 6 times (pretty much every day). On one day, even two times!
What I learned
Without clickstream data or Search Console exports, it’s impossible for us to get Discover data at scale. What a shame!
Aside of that, here is what I learned
First, Google has an accurate profile of my interests at the moment. Looking at your Web Activity data shows you how many signals you’re giving Google. It’s crazy. When I looked at the data, I wasn’t surprised at all by how accurately some of Discover cards piqued my interest. Just what I search for on Google Search and Youtube and my location would be enough.
Second, following entities in Search might impact Discover but it’s still not clear by how much. Some entities might see a big impact, other don’t. I think it still makes sense to activate your audience and ask them to follow you if you have a knowledge panel.
Third, some cards might be shown longer than 12 days on Discover, at least when you don’t click them.
Fourth, the more signals you give Google, the more precisely it understands your interests. That makes sense intuitively but it what fascinating to see that the entity classification for some cards changed from “Entertainment” to “Joe Rogan”. It almost seemed to me that Google’s understanding of what I want to see in a specific vertical (Technology, Entertainment, Politics, Economy) got better over time. It might even hold spots for those verticals in Discover but I wasn’t able to prove that.
Finds of the week
The Markup Google’s Top Search Result? Surprise! It’s Google
As someone who has spend the better half of 2020 to talk about exactly this problem, of course, I want to share this article. It comes with a few small flaws in the logic but overall present a good case interesting about The Markup: they collect their own data for any article they write.
Animalz Everybody Wants Thought Leadership Content. But How Do You Do It, Exactly?
The thing everybody wants but nobody knows how to do. Not any longer.
a16z Growth+Sales: The New Era of Enterprise Go-to-Market
The Enterprise segment is changing rapidly and attacked by bottoms-up products. Part of the reason is that old-school sales cycle are not in the interest of the user anymore.
The Google Cache On mathematics, experimentation and value
Rebuttal from Russ Jones for a pieces I linked a couple of weeks ago that claims we don’t have the tools to do correlation studies around ranking factors. I personally very much agree to what Russ here says because I prefer having sub-par information over no information at all.
Screaming Frog How Many Sites Pass the Core Web Vitals Assessment?
Richard from Screaming Frog looked at 22K keywords across mobile and desktop and checked whether the ranking pages would pass the Core Web Vitals assessment.