Big Data, Big Problems: 4 Major Link Indexes Compared

Posted by russangular

Given this blog’s readership, chances are good you will spend some time this week looking at backlinks in one of the growing number of link data tools. We know backlinks continue to be one of, if not the most important
parts of Google’s ranking algorithm. We tend to take these link data sets at face value, though, in part because they are all we have. But when your rankings are on the line, is there a better way to get at which data set is the best? How should we go
about assessing these different link indexes like
Majestic, Ahrefs and SEMrush for quality? Historically, there have been 4 common approaches to this question of index quality…

  • Breadth: We might choose to look at the number of linking root domains any given service reports. We know
    that referring domains correlates strongly with search rankings, so it makes sense to judge a link index by how many unique domains it has
    discovered and indexed.
  • Depth: We also might choose to look at how deep the web has been crawled, looking more at the total number of URLs
    in the index, rather than the diversity of referring domains.
  • Link Overlap: A more sophisticated approach might count the number of links an index has in common with Google Webmaster
  • Freshness: Finally, we might choose to look at the freshness of the index. What percentage of links in the index are
    still live?

There are a number of really good studies (some newer than others) using these techniques that are worth checking out when you get a chance:

  • BuiltVisible analysis of Moz, Majestic, GWT, Ahrefs and Search Metrics
  • SEOBook comparison of Moz, Majestic, Ahrefs, and Ayima
  • MatthewWoodward
    study of Ahrefs, Majestic, Moz, Raven and SEO Spyglass
  • Marketing Signals analysis of Moz, Majestic, Ahrefs, and GWT
  • RankAbove comparison of Moz, Majestic, Ahrefs and Link Research Tools
  • StoneTemple study of Moz and Majestic

While these are all excellent at addressing the methodologies above, there is a particular limitation with all of them. They miss one of the
most important metrics we need to determine the value of a link index: proportional representation to Google’s link graph
. So here at Angular Marketing, we decided to take a closer look.

Proportional representation to Google Search Console data

So, why is it important to determine proportional representation? Many of the most important and valued metrics we use are built on proportional
models. PageRank, MozRank, CitationFlow and Ahrefs Rank are proportional in nature. The score of any one URL in the data set is relative to the
other URLs in the data set. If the data set is biased, the results are biased.

A Visualization

Link graphs are biased by their crawl prioritization. Because there is no full representation of the Internet, every link graph, even Google’s,
is a biased sample of the web. Imagine for a second that the picture below is of the web. Each dot represents a page on the Internet,
and the dots surrounded by green represent a fictitious index by Google of certain sections of the web.

Of course, Google isn’t the only organization that crawls the web. Other organizations like Moz,
Majestic, Ahrefs, and SEMrush
have their own crawl prioritizations which result in different link indexes.

In the example above, you can see different link providers trying to index the web like Google. Link data provider 1 (purple) does a good job
of building a model that is similar to Google. It isn’t very big, but it is proportional. Link data provider 2 (blue) has a much larger index,
and likely has more links in common with Google that link data provider 1, but it is highly disproportional. So, how would we go about measuring
this proportionality? And which data set is the most proportional to Google?


The first step is to determine a measurement of relativity for analysis. Google doesn’t give us very much information about their link graph.
All we have is what is in Google Search Console. The best source we can use is referring domain counts. In particular, we want to look at
what we call
referring domain link pairs. A referring domain link pair would be something like> 9,444 which means
that links to 9,444 times.


  1. Determine the root linking domain pairs and values to 100+ sites in Google Search Console
  2. Determine the same for Ahrefs, Moz, Majestic Fresh, Majestic Historic, SEMrush
  3. Compare the referring domain link pairs of each data set to Google, assuming a
    Poisson Distribution
  4. Run simulations of each data set’s performance against each other (ie: Moz vs Maj, Ahrefs vs SEMrush, Moz vs SEMrush, et al.)
  5. Analyze the results


When placed head-to-head, there seem to be some clear winners at first glance. In head-to-head, Moz edges out Ahrefs, but across the board, Moz and Ahrefs fare quite evenly. Moz, Ahrefs and SEMrush seem to be far better than Majestic Fresh and Majestic Historic. Is that really the case? And why?

It turns out there is an inversely proportional relationship between index size and proportional relevancy. This might seem counterintuitive,
shouldn’t the bigger indexes be closer to Google? Not Exactly.

What does this mean?

Each organization has to create a crawl prioritization strategy. When you discover millions of links, you have to prioritize which ones you
might crawl next. Google has a crawl prioritization, so does Moz, Majestic, Ahrefs and SEMrush. There are lots of different things you might
choose to prioritize…

  • You might prioritize link discovery. If you want to build a very large index, you could prioritize crawling pages on sites that
    have historically provided new links.
  • You might prioritize content uniqueness. If you want to build a search engine, you might prioritize finding pages that are unlike
    any you have seen before. You could choose to crawl domains that historically provide unique data and little duplicate content.
  • You might prioritize content freshness. If you want to keep your search engine recent, you might prioritize crawling pages that
    change frequently.
  • You might prioritize content value, crawling the most important URLs first based on the number of inbound links to that page.

Chances are, an organization’s crawl priority will blend some of these features, but it’s difficult to design one exactly like Google. Imagine
for a moment that instead of crawling the web, you want to climb a tree. You have to come up with a tree climbing strategy.

  • You decide to climb the longest branch you see at each intersection.
  • One friend of yours decides to climb the first new branch he reaches, regardless of how long it is.
  • Your other friend decides to climb the first new branch she reaches only if she sees another branch coming off of it.

Despite having different climb strategies, everyone chooses the same first branch, and everyone chooses the same second branch. There are only
so many different options early on.

But as the climbers go further and further along, their choices eventually produce differing results. This is exactly the same for web crawlers
like Google, Moz, Majestic, Ahrefs and SEMrush. The bigger the crawl, the more the crawl prioritization will cause disparities. This is not a
deficiency; this is just the nature of the beast. However, we aren’t completely lost. Once we know how index size is related to disparity, we
can make some inferences about how similar a crawl priority may be to Google.

Unfortunately, we have to be careful in our conclusions. We only have a few data points with which to work, so it is very difficult to be
certain regarding this part of the analysis. In particular, it seems strange that Majestic would get better relative to its index size as it grows,
unless Google holds on to old data (which might be an important discovery in and of itself). It is most likely that at this point we can’t make
this level of conclusion.

So what do we do?

Let’s say you have a list of domains or URLs for which you would like to know their relative values. Your process might look something like

  • Check Open Site Explorer to see if all URLs are in their index. If so, you are looking metrics most likely to be proportional to Google’s link graph.
  • If any of the links do not occur in the index, move to Ahrefs and use their Ahrefs ranking if all you need is a single PageRank-like metric.
  • If any of the links are missing from Ahrefs’s index, or you need something related to trust, move on to Majestic Fresh.
  • Finally, use Majestic Historic for (by leaps and bounds) the largest coverage available.

It is important to point out that the likelihood that all the URLs you want to check are in a single index increases as the accuracy of the metric
decreases. Considering the size of Majestic’s data, you can’t ignore them because you are less likely to get null value answers from their data than
the others. If anything rings true, it is that once again it makes sense to get data
from as many sources as possible. You won’t
get the most proportional data without Moz, the broadest data without Majestic, or everything in-between without Ahrefs.

What about SEMrush? They are making progress, but they don’t publish any relative statistics that would be useful in this particular
case. Maybe we can hope to see more from them soon given their already promising index!

Recommendations for the link graphing industry

All we hear about these days is big data; we almost never hear about good data. I know that the teams at Moz,
Majestic, Ahrefs, SEMrush and others are interested in mimicking Google, but I would love to see some organization stand up against the
allure of
more data in favor of better data—data more like Google’s. It could begin with testing various crawl strategies to see if they produce
a result more similar to that of data shared in Google Search Console. Having the most Google-like data is certainly a crown worth winning.


Thanks to Diana Carter at Angular for assistance with data acquisition and Andrew Cron with statistical analysis. Thanks also to the representatives from Moz, Majestic, Ahrefs, and SEMrush for answering questions about their indices.

Sign up for The Moz Top 10, a semimonthly mailer updating you on the top ten hottest pieces of SEO news, tips, and rad links uncovered by the Moz team. Think of it as your exclusive digest of stuff you don’t have time to hunt down but want to read!


Deconstructing the App Store Rankings Formula with a Little Mad Science

Posted by AlexApptentive

After seeing Rand’s “Mad Science Experiments in SEO” presented at last year’s MozCon, I was inspired to put on the lab coat and goggles and do a few experiments of my own—not in SEO, but in SEO’s up-and-coming younger sister, ASO (app store optimization).

Working with Apptentive to guide enterprise apps and small startup apps alike to increase their discoverability in the app stores, I’ve learned a thing or two about app store optimization and what goes into an app’s ranking. It’s been my personal goal for some time now to pull back the curtains on Google and Apple. Yet, the deeper into the rabbit hole I go, the more untested assumptions I leave in my way.

Hence, I thought it was due time to put some longstanding hypotheses through the gauntlet.

As SEOs, we know how much of an impact a single ranking can mean on a SERP. One tiny rank up or down can make all the difference when it comes to your website’s traffic—and revenue.

In the world of apps, ranking is just as important when it comes to standing out in a sea of more than 1.3 million apps. Apptentive’s recent mobile consumer survey shed a little more light this claim, revealing that nearly half of all mobile app users identified browsing the app store charts and search results (the placement on either of which depends on rankings) as a preferred method for finding new apps in the app stores. Simply put, better rankings mean more downloads and easier discovery.

Like Google and Bing, the two leading app stores (the Apple App Store and Google Play) have a complex and highly guarded algorithms for determining rankings for both keyword-based app store searches and composite top charts.

Unlike SEO, however, very little research and theory has been conducted around what goes into these rankings.

Until now, that is.

Over the course of five studies analyzing various publicly available data points for a cross-section of the top 500 iOS (U.S. Apple App Store) and the top 500 Android (U.S. Google Play) apps, I’ll attempt to set the record straight with a little myth-busting around ASO. In the process, I hope to assess and quantify any perceived correlations between app store ranks, ranking volatility, and a few of the factors commonly thought of as influential to an app’s ranking.

But first, a little context

Image credit: Josh Tuininga, Apptentive

Both the Apple App Store and Google Play have roughly 1.3 million apps each, and both stores feature a similar breakdown by app category. Apps ranking in the two stores should, theoretically, be on a fairly level playing field in terms of search volume and competition.

Of these apps, nearly two-thirds have not received a single rating and 99% are considered unprofitable. These studies, therefore, single out the rare exceptions to the rule—the top 500 ranked apps in each store.

While neither Apple nor Google have revealed specifics about how they calculate search rankings, it is generally accepted that both app store algorithms factor in:

  • Average app store rating
  • Rating/review volume
  • Download and install counts
  • Uninstalls (what retention and churn look like for the app)
  • App usage statistics (how engaged an app’s users are and how frequently they launch the app)
  • Growth trends weighted toward recency (how daily download counts changed over time and how today’s ratings compare to last week’s)
  • Keyword density of the app’s landing page (Ian did a great job covering this factor in a previous Moz post)

I’ve simplified this formula to a function highlighting the four elements with sufficient data (or at least proxy data) for our analysis:

Ranking = fn(Rating, Rating Count, Installs, Trends)

Of course, right now, this generalized function doesn’t say much. Over the next five studies, however, we’ll revisit this function before ultimately attempting to compare the weights of each of these four variables on app store rankings.

(For the purpose of brevity, I’ll stop here with the assumptions, but I’ve gone into far greater depth into how I’ve reached these conclusions in a 55-page report on app store rankings.)

Now, for the Mad Science.

Study #1: App-les to app-les app store ranking volatility

The first, and most straight forward of the five studies involves tracking daily movement in app store rankings across iOS and Android versions of the same apps to determine any trends of differences between ranking volatility in the two stores.

I went with a small sample of five apps for this study, the only criteria for which were that:

  • They were all apps I actively use (a criterion for coming up with the five apps but not one that influences rank in the U.S. app stores)
  • They were ranked in the top 500 (but not the top 25, as I assumed app store rankings would be stickier at the top—an assumption I’ll test in study #2)
  • They had an almost identical version of the app in both Google Play and the App Store, meaning they should (theoretically) rank similarly
  • They covered a spectrum of app categories

The apps I ultimately chose were Lyft, Venmo, Duolingo, Chase Mobile, and LinkedIn. These five apps represent the travel, finance, education banking, and social networking categories.


Going into this analysis, I predicted slightly more volatility in Apple App Store rankings, based on two statistics:

Both of these assumptions will be tested in later analysis.


7-Day App Store Ranking Volatility in the App Store and Google Play

Among these five apps, Google Play rankings were, indeed, significantly less volatile than App Store rankings. Among the 35 data points recorded, rankings within Google Play moved by as much as 23 positions/ranks per day while App Store rankings moved up to 89 positions/ranks. The standard deviation of ranking volatility in the App Store was, furthermore, 4.45 times greater than that of Google Play.

Of course, the same apps varied fairly dramatically in their rankings in the two app stores, so I then standardized the ranking volatility in terms of percent change to control for the effect of numeric rank on volatility. When cast in this light, App Store rankings changed by as much as 72% within a 24-hour period while Google Play rankings changed by no more than 9%.

Also of note, daily rankings tended to move in the same direction across the two app stores approximately two-thirds of the time, suggesting that the two stores, and their customers, may have more in common than we think.

Study #2: App store ranking volatility across the top charts

Testing the assumption implicit in standardizing the data in study No. 1, this one was designed to see if app store ranking volatility is correlated with an app’s current rank. The sample for this study consisted of the top 500 ranked apps in both Google Play and the App Store, with special attention given to those on both ends of the spectrum (ranks 1–100 and 401–500).


I anticipated rankings to be more volatile the higher an app is ranked—meaning an app ranked No. 450 should be able to move more ranks in any given day than an app ranked No. 50. This hypothesis is based on the assumption that higher ranked apps have more installs, active users, and ratings, and that it would take a large margin to produce a noticeable shift in any of these factors.


App Store Ranking Volatility of Top 500 Apps

One look at the chart above shows that apps in both stores have increasingly more volatile rankings (based on how many ranks they moved in the last 24 hours) the lower on the list they’re ranked.

This is particularly true when comparing either end of the spectrum—with a seemingly straight volatility line among Google Play’s Top 100 apps and very few blips within the App Store’s Top 100. Compare this section to the lower end, ranks 401–)500, where both stores experience much more turbulence in their rankings. Across the gamut, I found a 24% correlation between rank and ranking volatility in the Play Store and 28% correlation in the App Store.

To put this into perspective, the average app in Google Play’s 401–)500 ranks moved 12.1 ranks in the last 24 hours while the average app in the Top 100 moved a mere 1.4 ranks. For the App Store, these numbers were 64.28 and 11.26, making slightly lower-ranked apps more than five times as volatile as the highest ranked apps. (I say slightly as these “lower-ranked” apps are still ranked higher than 99.96% of all apps.)

The relationship between rank and volatility is pretty consistent across the App Store charts, while rank has a much greater impact on volatility at the lower end of Google Play charts (ranks 1-100 have a 35% correlation) than it does at the upper end (ranks 401-500 have a 1% correlation).

Study #3: App store rankings across the stars

The next study looks at the relationship between rank and star ratings to determine any trends that set the top chart apps apart from the rest and explore any ties to app store ranking volatility.


Ranking = fn(Rating, Rating Count, Installs, Trends)

As discussed in the introduction, this study relates directly to one of the factors commonly accepted as influential to app store rankings: average rating.

Getting started, I hypothesized that higher ranks generally correspond to higher ratings, cementing the role of star ratings in the ranking algorithm.

As far as volatility goes, I did not anticipate average rating to play a role in app store ranking volatility, as I saw no reason for higher rated apps to be less volatile than lower rated apps, or vice versa. Instead, I believed volatility to be tied to rating volume (as we’ll explore in our last study).


Average App Store Ratings of Top Apps

The chart above plots the top 100 ranked apps in either store with their average rating (both historic and current, for App Store apps). If it looks a little chaotic, it’s just one indicator of the complexity of ranking algorithm in Google Play and the App Store.

If our hypothesis was correct, we’d see a downward trend in ratings. We’d expect to see the No. 1 ranked app with a significantly higher rating than the No. 100 ranked app. Yet, in neither store is this the case. Instead, we get a seemingly random plot with no obvious trends that jump off the chart.

A closer examination, in tandem with what we already know about the app stores, reveals two other interesting points:

  1. The average star rating of the top 100 apps is significantly higher than that of the average app. Across the top charts, the average rating of a top 100 Android app was 4.319 and the average top iOS app was 3.935. These ratings are 0.32 and 0.27 points, respectively, above the average rating of all rated apps in either store. The averages across apps in the 401–)500 ranks approximately split the difference between the ratings of the top ranked apps and the ratings of the average app.
  2. The rating distribution of top apps in Google Play was considerably more compact than the distribution of top iOS apps. The standard deviation of ratings in the Apple App Store top chart was over 2.5 times greater than that of the Google Play top chart, likely meaning that ratings are more heavily weighted in Google Play’s algorithm.

App Store Ranking Volatility and Average Rating

Looking next at the relationship between ratings and app store ranking volatility reveals a -15% correlation that is consistent across both app stores; meaning the higher an app is rated, the less its rank it likely to move in a 24-hour period. The exception to this rule is the Apple App Store’s calculation of an app’s current rating, for which I did not find a statistically significant correlation.

Study #4: App store rankings across versions

This next study looks at the relationship between the age of an app’s current version, its rank and its ranking volatility.


Ranking = fn(Rating, Rating Count, Installs, Trends)

In alteration of the above function, I’m using the age of a current app’s version as a proxy (albeit not a very good one) for trends in app store ratings and app quality over time.

Making the assumptions that (a) apps that are updated more frequently are of higher quality and (b) each new update inspires a new wave of installs and ratings, I’m hypothesizing that the older the age of an app’s current version, the lower it will be ranked and the less volatile its rank will be.


How update frequency correlates with app store rank

The first and possibly most important finding is that apps across the top charts in both Google Play and the App Store are updated remarkably often as compared to the average app.

At the time of conducting the study, the current version of the average iOS app on the top chart was only 28 days old; the current version of the average Android app was 38 days old.

As hypothesized, the age of the current version is negatively correlated with the app’s rank, with a 13% correlation in Google Play and a 10% correlation in the App Store.

How update frequency correlates with app store ranking volatility

The next part of the study maps the age of the current app version to its app store ranking volatility, finding that recently updated Android apps have less volatile rankings (correlation: 8.7%) while recently updated iOS apps have more volatile rankings (correlation: -3%).

Study #5: App store rankings across monthly active users

In the final study, I wanted to examine the role of an app’s popularity on its ranking. In an ideal world, popularity would be measured by an app’s monthly active users (MAUs), but since few mobile app developers have released this information, I’ve settled for two publicly available proxies: Rating Count and Installs.


Ranking = fn(Rating, Rating Count, Installs, Trends)

For the same reasons indicated in the second study, I anticipated that more popular apps (e.g., apps with more ratings and more installs) would be higher ranked and less volatile in rank. This, again, takes into consideration that it takes more of a shift to produce a noticeable impact in average rating or any of the other commonly accepted influencers of an app’s ranking.


Apps with more ratings and reviews typically rank higher

The first finding leaps straight off of the chart above: Android apps have been rated more times than iOS apps, 15.8x more, in fact.

The average app in Google Play’s Top 100 had a whopping 3.1 million ratings while the average app in the Apple App Store’s Top 100 had 196,000 ratings. In contrast, apps in the 401–)500 ranks (still tremendously successful apps in the 99.96 percentile of all apps) tended to have between one-tenth (Android) and one-fifth (iOS) of the ratings count as that of those apps in the top 100 ranks.

Considering that almost two-thirds of apps don’t have a single rating, reaching rating counts this high is a huge feat, and a very strong indicator of the influence of rating count in the app store ranking algorithms.

To even out the playing field a bit and help us visualize any correlation between ratings and rankings (and to give more credit to the still-staggering 196k ratings for the average top ranked iOS app), I’ve applied a logarithmic scale to the chart above:

The relationship between app store ratings and rankings in the top 100 apps

From this chart, we can see a correlation between ratings and rankings, such that apps with more ratings tend to rank higher. This equates to a 29% correlation in the App Store and a 40% correlation in Google Play.

Apps with more ratings typically experience less app store ranking volatility

Next up, I looked at how ratings count influenced app store ranking volatility, finding that apps with more ratings had less volatile rankings in the Apple App Store (correlation: 17%). No conclusive evidence was found within the Top 100 Google Play apps.

Apps with more installs and active users tend to rank higher in the app stores

And last but not least, I looked at install counts as an additional proxy for MAUs. (Sadly, this is a statistic only listed in Google Play. so any resulting conclusions are applicable only to Android apps.)

Among the top 100 Android apps, this last study found that installs were heavily correlated with ranks (correlation: -35.5%), meaning that apps with more installs are likely to rank higher in Google Play. Android apps with more installs also tended to have less volatile app store rankings, with a correlation of -16.5%.

Unfortunately, these numbers are slightly skewed as Google Play only provides install counts in broad ranges (e.g., 500k–)1M). For each app, I took the low end of the range, meaning we can likely expect the correlation to be a little stronger since the low end was further away from the midpoint for apps with more installs.


To make a long post ever so slightly shorter, here are the nuts and bolts unearthed in these five mad science studies in app store optimization:

  1. Across the top charts, Apple App Store rankings are 4.45x more volatile than those of Google Play
  2. Rankings become increasingly volatile the lower an app is ranked. This is particularly true across the Apple App Store’s top charts.
  3. In both stores, higher ranked apps tend to have an app store ratings count that far exceeds that of the average app.
  4. Ratings appear to matter more to the Google Play algorithm, especially as the Apple App Store top charts experience a much wider ratings distribution than that of Google Play’s top charts.
  5. The higher an app is rated, the less volatile its rankings are.
  6. The 100 highest ranked apps in either store are updated much more frequently than the average app, and apps with older current versions are correlated with lower ratings.
  7. An app’s update frequency is negatively correlated with Google Play’s ranking volatility but positively correlated with ranking volatility in the App Store. This likely due to how Apple weighs an app’s most recent ratings and reviews.
  8. The highest ranked Google Play apps receive, on average, 15.8x more ratings than the highest ranked App Store apps.
  9. In both stores, apps that fall under the 401–500 ranks receive, on average, 10–20% of the rating volume seen by apps in the top 100.
  10. Rating volume and, by extension, installs or MAUs, is perhaps the best indicator of ranks, with a 29–40% correlation between the two.

Revisiting our first (albeit oversimplified) guess at the app stores’ ranking algorithm gives us this loosely defined function:

Ranking = fn(Rating, Rating Count, Installs, Trends)

I’d now re-write the function into a formula by weighing each of these four factors, where a, b, c, & d are unknown multipliers, or weights:

Ranking = (Rating * a) + (Rating Count * b) + (Installs * c) + (Trends * d)

These five studies on ASO shed a little more light on these multipliers, showing Rating Count to have the strongest correlation with rank, followed closely by Installs, in either app store.

It’s with the other two factors—rating and trends—that the two stores show the greatest discrepancy. I’d hazard a guess to say that the App Store prioritizes growth trends over ratings, given the importance it places on an app’s current version and the wide distribution of ratings across the top charts. Google Play, on the other hand, seems to favor ratings, with an unwritten rule that apps just about have to have at least four stars to make the top 100 ranks.

Thus, we conclude our mad science with this final glimpse into what it takes to make the top charts in either store:

Weight of factors in the Apple App Store ranking algorithm

Rating Count > Installs > Trends > Rating

Weight of factors in the Google Play ranking algorithm

Rating Count > Installs > Rating > Trends

Again, we’re oversimplifying for the sake of keeping this post to a mere 3,000 words, but additional factors including keyword density and in-app engagement statistics continue to be strong indicators of ranks. They simply lie outside the scope of these studies.

I hope you found this deep-dive both helpful and interesting. Moving forward, I also hope to see ASOs conducting the same experiments that have brought SEO to the center stage, and encourage you to enhance or refute these findings with your own ASO mad science experiments.

Please share your thoughts in the comments below, and let’s deconstruct the ranking formula together, one experiment at a time.

Sign up for The Moz Top 10, a semimonthly mailer updating you on the top ten hottest pieces of SEO news, tips, and rad links uncovered by the Moz team. Think of it as your exclusive digest of stuff you don’t have time to hunt down but want to read!


What Deep Learning and Machine Learning Mean For the Future of SEO – Whiteboard Friday

Posted by randfish

Imagine a world where even the high-up Google engineers don’t know what’s in the ranking algorithm. We may be moving in that direction. In today’s Whiteboard Friday, Rand explores and explains the concepts of deep learning and machine learning, drawing us a picture of how they could impact our work as SEOs.

For reference, here’s a still of this week’s whiteboard!

Video transcription

Howdy, Moz fans, and welcome to another edition of Whiteboard Friday. This week we are going to take a peek into Google’s future and look at what it could mean as Google advances their machine learning and deep learning capabilities. I know these sound like big, fancy, important words. They’re not actually that tough of topics to understand. In fact, they’re simplistic enough that even a lot of technology firms like Moz do some level of machine learning. We don’t do anything with deep learning and a lot of neural networks. We might be going that direction.

But I found an article that was published in January, absolutely fascinating and I think really worth reading, and I wanted to extract some of the contents here for Whiteboard Friday because I do think this is tactically and strategically important to understand for SEOs and really important for us to understand so that we can explain to our bosses, our teams, our clients how SEO works and will work in the future.

The article is called “Google Search Will Be Your Next Brain.” It’s by Steve Levy. It’s over on Medium. I do encourage you to read it. It’s a relatively lengthy read, but just a fascinating one if you’re interested in search. It starts with a profile of Geoff Hinton, who was a professor in Canada and worked on neural networks for a long time and then came over to Google and is now a distinguished engineer there. As the article says, a quote from the article: “He is versed in the black art of organizing several layers of artificial neurons so that the entire system, the system of neurons, could be trained or even train itself to divine coherence from random inputs.”

This sounds complex, but basically what we’re saying is we’re trying to get machines to come up with outcomes on their own rather than us having to tell them all the inputs to consider and how to process those incomes and the outcome to spit out. So this is essentially machine learning. Google has used this, for example, to figure out when you give it a bunch of photos and it can say, “Oh, this is a landscape photo. Oh, this is an outdoor photo. Oh, this is a photo of a person.” Have you ever had that creepy experience where you upload a photo to Facebook or to Google+ and they say, “Is this your friend so and so?” And you’re like, “God, that’s a terrible shot of my friend. You can barely see most of his face, and he’s wearing glasses which he usually never wears. How in the world could Google+ or Facebook figure out that this is this person?”

That’s what they use, these neural networks, these deep machine learning processes for. So I’ll give you a simple example. Here at MOZ, we do machine learning very simplistically for page authority and domain authority. We take all the inputs — numbers of links, number of linking root domains, every single metric that you could get from MOZ on the page level, on the sub-domain level, on the root-domain level, all these metrics — and then we combine them together and we say, “Hey machine, we want you to build us the algorithm that best correlates with how Google ranks pages, and here’s a bunch of pages that Google has ranked.” I think we use a base set of 10,000, and we do it about quarterly or every 6 months, feed that back into the system and the system pumps out the little algorithm that says, “Here you go. This will give you the best correlating metric with how Google ranks pages.” That’s how you get page authority domain authority.

Cool, really useful, helpful for us to say like, “Okay, this page is probably considered a little more important than this page by Google, and this one a lot more important.” Very cool. But it’s not a particularly advanced system. The more advanced system is to have these kinds of neural nets in layers. So you have a set of networks, and these neural networks, by the way, they’re designed to replicate nodes in the human brain, which is in my opinion a little creepy, but don’t worry. The article does talk about how there’s a board of scientists who make sure Terminator 2 doesn’t happen, or Terminator 1 for that matter. Apparently, no one’s stopping Terminator 4 from happening? That’s the new one that’s coming out.

So one layer of the neural net will identify features. Another layer of the neural net might classify the types of features that are coming in. Imagine this for search results. Search results are coming in, and Google’s looking at the features of all the websites and web pages, your websites and pages, to try and consider like, “What are the elements I could pull out from there?”

Well, there’s the link data about it, and there are things that happen on the page. There are user interactions and all sorts of stuff. Then we’re going to classify types of pages, types of searches, and then we’re going to extract the features or metrics that predict the desired result, that a user gets a search result they really like. We have an algorithm that can consistently produce those, and then neural networks are hopefully designed — that’s what Geoff Hinton has been working on — to train themselves to get better. So it’s not like with PA and DA, our data scientist Matt Peters and his team looking at it and going, “I bet we could make this better by doing this.”

This is standing back and the guys at Google just going, “All right machine, you learn.” They figure it out. It’s kind of creepy, right?

In the original system, you needed those people, these individuals here to feed the inputs, to say like, “This is what you can consider, system, and the features that we want you to extract from it.”

Then unsupervised learning, which is kind of this next step, the system figures it out. So this takes us to some interesting places. Imagine the Google algorithm, circa 2005. You had basically a bunch of things in here. Maybe you’d have anchor text, PageRank and you’d have some measure of authority on a domain level. Maybe there are people who are tossing new stuff in there like, “Hey algorithm, let’s consider the location of the searcher. Hey algorithm, let’s consider some user and usage data.” They’re tossing new things into the bucket that the algorithm might consider, and then they’re measuring it, seeing if it improves.

But you get to the algorithm today, and gosh there are going to be a lot of things in there that are driven by machine learning, if not deep learning yet. So there are derivatives of all of these metrics. There are conglomerations of them. There are extracted pieces like, “Hey, we only ant to look and measure anchor text on these types of results when we also see that the anchor text matches up to the search queries that have previously been performed by people who also search for this.” What does that even mean? But that’s what the algorithm is designed to do. The machine learning system figures out things that humans would never extract, metrics that we would never even create from the inputs that they can see.

Then, over time, the idea is that in the future even the inputs aren’t given by human beings. The machine is getting to figure this stuff out itself. That’s weird. That means that if you were to ask a Google engineer in a world where deep learning controls the ranking algorithm, if you were to ask the people who designed the ranking system, “Hey, does it matter if I get more links,” they might be like, “Well, maybe.” But they don’t know, because they don’t know what’s in this algorithm. Only the machine knows, and the machine can’t even really explain it. You could go take a snapshot and look at it, but (a) it’s constantly evolving, and (b) a lot of these metrics are going to be weird conglomerations and derivatives of a bunch of metrics mashed together and torn apart and considered only when certain criteria are fulfilled. Yikes.

So what does that mean for SEOs. Like what do we have to care about from all of these systems and this evolution and this move towards deep learning, which by the way that’s what Jeff Dean, who is, I think, a senior fellow over at Google, he’s the dude that everyone mocks for being the world’s smartest computer scientist over there, and Jeff Dean has basically said, “Hey, we want to put this into search. It’s not there yet, but we want to take these models, these things that Hinton has built, and we want to put them into search.” That for SEOs in the future is going to mean much less distinct universal ranking inputs, ranking factors. We won’t really have ranking factors in the way that we know them today. It won’t be like, “Well, they have more anchor text and so they rank higher.” That might be something we’d still look at and we’d say, “Hey, they have this anchor text. Maybe that’s correlated with what the machine is finding, the system is finding to be useful, and that’s still something I want to care about to a certain extent.”

But we’re going to have to consider those things a lot more seriously. We’re going to have to take another look at them and decide and determine whether the things that we thought were ranking factors still are when the neural network system takes over. It also is going to mean something that I think many, many SEOs have been predicting for a long time and have been working towards, which is more success for websites that satisfy searchers. If the output is successful searches, and that’ s what the system is looking for, and that’s what it’s trying to correlate all its metrics to, if you produce something that means more successful searches for Google searchers when they get to your site, and you ranking in the top means Google searchers are happier, well you know what? The algorithm will catch up to you. That’s kind of a nice thing. It does mean a lot less info from Google about how they rank results.

So today you might hear from someone at Google, “Well, page speed is a very small ranking factor.” In the future they might be, “Well, page speed is like all ranking factors, totally unknown to us.” Because the machine might say, “Well yeah, page speed as a distinct metric, one that a Google engineer could actually look at, looks very small.” But derivatives of things that are connected to page speed may be huge inputs. Maybe page speed is something, that across all of these, is very well connected with happier searchers and successful search results. Weird things that we never thought of before might be connected with them as the machine learning system tries to build all those correlations, and that means potentially many more inputs into the ranking algorithm, things that we would never consider today, things we might consider wholly illogical, like, “What servers do you run on?” Well, that seems ridiculous. Why would Google ever grade you on that?

If human beings are putting factors into the algorithm, they never would. But the neural network doesn’t care. It doesn’t care. It’s a honey badger. It doesn’t care what inputs it collects. It only cares about successful searches, and so if it turns out that Ubuntu is poorly correlated with successful search results, too bad.

This world is not here yet today, but certainly there are elements of it. Google has talked about how Panda and Penguin are based off of machine learning systems like this. I think, given what Geoff Hinton and Jeff Dean are working on at Google, it sounds like this will be making its way more seriously into search and therefore it’s something that we’re really going to have to consider as search marketers.

All right everyone, I hope you’ll join me again next week for another edition of Whiteboard Friday. Take care.

Video transcription by

Sign up for The Moz Top 10, a semimonthly mailer updating you on the top ten hottest pieces of SEO news, tips, and rad links uncovered by the Moz team. Think of it as your exclusive digest of stuff you don’t have time to hunt down but want to read!


Is It Possible to Have Good SEO Simply by Having Great Content – Whiteboard Friday

Posted by randfish

This question, posed by Alex Moravek in our Q&A section, has a somewhat complicated answer. In today’s Whiteboard Friday, Rand discusses how organizations might perform well in search rankings without doing any link building at all, relying instead on the strength of their content to be deemed relevant and important by Google.

For reference, here’s a still of this week’s whiteboard!

Video transcription

Howdy Moz fans, and welcome to another edition of Whiteboard Friday. This week we’re chatting about is it possible to have good SEO simply by focusing on great content to the exclusion of link building.

This question was posed in the Moz Q&A Forum, which I deeply love, by Alex Moravek — I might not be saying your name right, Alex, and for that I apologize — from SEO Agencias in Madrid. My Spanish is poor, but my love for churros is so strong.

Alex, I think this is a great question. In fact, we get asked this all the time by all sorts of folks, particularly people in the blogging world and people with small and medium businesses who hear about SEO and go, “Okay, I think can make my website accessible, and yes, I can produce great content, but I just either don’t feel comfortable, don’t have time and energy, don’t understand, or just don’t feel okay with doing link building.” Link acquisition through an outreach and a manual process is beyond the scope of what they can fit into their marketing activities.

In fact, it is possible kind of, sort of. It is possible, but what you desperately need in order for this strategy to be possible are really two things. One is content exposure, and two you need time. I’ll explain why you need both of these things.

I’m going to dramatically simplify Google’s ranking algorithm. In fact, I’m going to simplify it so much that those of you who are SEO professionals are going to be like, “Oh God, Rand, you’re killing me.” I apologize in advance. Just bear with me a second.

We basically have keywords and on-page stuff, topical relevance, etc. All your topic modeling stuff might go in there. There’s content quality, all the factors that Google and Bing might measure around a content’s quality. There’s domain authority. There’s link-based authority based on the links that point to all the pages on a given domain that tell Google or Bing how important pages on this particular domain are.

There are probably some topical relevance elements in there, too. There’s page level authority. These could be all the algorithms you’ve heard of like PageRank and TrustRank, etc., and all the much more modern ones of those.

I’m not specifically talking about Moz scores here, the Moz scores DA and PA. Those are rough interpretations of these much more sophisticated formulas that the engines have.

There’s user and usage data, which we know the engines are using. They’ve talked about using that. There’s spam analysis.

Super simplistic. There are these six things, six broad categories of ranking elements. If you have just these four — keywords, on-page content quality, user and usage data, spam analysis, you’re not spammy — without these, without any domain authority or any page authority, it’s next to impossible to rank for competitive terms and very challenging and very unlikely to rank even for stuff in the chunky middle and long tail. Long tail you might rank for a few things if it’s very, very long tail. But these things taken together give you a sense of ranking ability.

Here’s what some marketers, some bloggers, some folks who invest in content nearly to the exclusion of links have found. They have had success with this strategy. They’ve basically elected to entirely ignore link building and let links come to them.

Instead of focusing on link building, they’re going to focus on product quality, press and public relations, social media, offline marketing, word of mouth, content strategy, email marketing, these other channels that can potentially earn them things. Advertising as well potentially could be in here.

What they rely on is that people find them through these other channels. They find them through social, through ads, through offline, through blogs, through very long tail search, through their content, maybe their email marketing list, word of mouth, press. All of these things are discovery mechanisms that are not search.

Once people get to the site, then these websites rely on the fact that, because of the experience people have, the quality of their products, of their content, because all of that stuff is so good, they’re going to earn links naturally.

This is a leap. In fact, for many SEOs, this is kind of a crazy leap to make, because there are so many things that you can do that will nudge people in this link earning direction. We’ve talked about a number of those at Moz. Of course, if you visit the link building section of our blog, there are hundreds if not thousands of great strategies around this.

These folks have elected to ignore all that link building stuff, let the links come to them, and these signals, these people who visit via other channels eventually lead to links which lead to DA, PA ranking ability. I don’t think this strategy is for everyone, but it is possible.

I think in the utopia that Larry Page and Sergey Brin from Google imagined when they were building their first search engine this is, in fact, how they hoped that the web would work. They hoped that people wouldn’t be out actively gaming and manipulating the web’s link graph, but rather that all the links would be earned naturally and editorially.

I think that’s a very, very optimistic and almost naive way of thinking about it. Remember, they were college students at the time. Maybe they were eating their granola, and dancing around, and hoping that everyone on the web would link only for editorial reasons. Not to make fun of granola. I love granola, especially, oh man, with those acai berries. Bowls of those things are great.

This is a potential strategy if you are very uncomfortable with link building and you feel like you can optimize this process. You have all of these channels going on.

For SEOs who are thinking, “Rand, I’m never going to ignore link building,” you can still get a tremendous amount out of thinking about how you optimize the return on investment and especially the exposure that you receive from these and how that might translate naturally into links.

I find looking at websites that accomplish SEO without active link building fascinating, because they have editorially earned those links through very little intentional effort on their own. I think there’s a tremendous amount that we can take away from that process and optimize around this.

Alex, yes, this is possible. Would I recommend it? Only in a very few instances. I think that there’s a ton that SEOs can do to optimize and nudge and create intelligent, non-manipulative ways of earning links that are a little more powerful than just sitting back and waiting, but it is possible.

All right, everyone. Thanks for joining us, and we’ll see you again next week for another edition of Whiteboard Friday. Take care.

Video transcription by

Sign up for The Moz Top 10, a semimonthly mailer updating you on the top ten hottest pieces of SEO news, tips, and rad links uncovered by the Moz team. Think of it as your exclusive digest of stuff you don’t have time to hunt down but want to read!