Back in September last year, I was lucky enough to see Rand speak at MozCon. His talk was about link building and the main types of strategy that he saw as still being relevant and effective today. During his introduction, he said something that really got me thinking, about how the whole purpose of links and PageRank had been to approximate traffic.
Essentially, back in the late '90s, links were a much bigger part of how we experienced the web — think of hubs like Excite, AOL, and Yahoo. Google’s big innovation was to realize that, because people navigated the web by clicking on links, they could approximate the relative popularity of pages by looking at those links.
So many links, such little time.
Rand pointed out that, given all the information at their disposal in the present day — as an Internet Service Provider, a search engine, a browser, an operating system, and so on — Google could now far more accurately model whether a link drives traffic, so you shouldn’t aim to build links that don’t drive traffic. This is a pretty big step forward from the link-building tactics of old, but it occurred to me that it probably doesn’t go far enough.
If Google has enough data to figure out which links are genuinely driving traffic, why bother with links at all? The whole point was to figure out which sites and pages were popular, and they can now answer that question directly. (It’s worth noting that there’s a dichotomy between “popular” and “trustworthy” that I don’t want to get too stuck into, but which isn’t too big a deal here given that both can be inferred from either link-based data sources, or from non-link-based data sources — for example, SERP click-through rate might correlate well with “trustworthy,” while “search volume” might correlate well with “popular”).
However, there’s plenty of evidence out there suggesting that Google is in fact still making significant use of links as a ranking factor, so I decided to set out to challenge the data on both sides of that argument. The end result of that research is this post.
The horse's mouth
One reasonably authoritative source on matters relating to Google is Google themselves. Google has been fairly unequivocal, even in recent times, that links are still a big deal. For example:
- March 2016: Google Senior Search Quality Strategist Andrey Lipattsev confirms that content and links are the first and second greatest ranking factors. (The full quote is: “Yes; I can tell you what they [the number 1 and 2 ranking factors] are. It’s content, and links pointing to your site.”)
- April 2014: Matt Cutts confirms that Google has tested search quality without links, and found it to be inferior.
- October 2016: Gary Illyes implies that text links continue to be valuable while playing down the concept of Domain Authority.
Then, of course, there’s their continued focus on unnatural backlinks and so on — none of which would be necessary in a world where links are not a ranking factor.
However, I’d argue that this doesn’t indicate the end of our discussion before it’s even begun. Firstly, Google has a great track record of giving out dodgy SEO advice. Consider HTTPS migrations pre-2016. Will Critchlow talked at SearchLove San Diego about how Google’s algorithms are at a level of complexity and opaqueness where they’re no longer even trying to understand them themselves — and of course there are numerous stories of unintentional behaviors from machine learning algorithms out in the wild.
Third-party correlation studies
It’s not difficult to put together your own data and show a correlation between link-based metrics and rankings. Take, for example:
- Moz’s most recent study in 2015, showing strong relationships between link-based factors and rankings across the board.
- This more recent study by Stone Temple Consulting.
However, these studies fall into significant issues with correlation vs. causation.
There are three main mechanisms which could explain the relationships that they show:
- Getting more links causes sites to rank higher (yay!)
- Ranking higher causes sites to get more links
- Some third factor, such as brand awareness, is related to both links and rankings, causing them to be correlated with each other despite the absence of a direct causal relationship
I’ve yet to see any correlation study that addresses these very serious shortcomings, or even particularly acknowledges them. Indeed, I’m not sure that it would even be possible to do so given the available data, but this does show that as an industry we need to apply some critical thinking to the advice that we’re consuming.
However, earlier this year I did write up some research of my own here on the Moz Blog, demonstrating that brand awareness could in fact be a more useful factor than links for predicting rankings.
The problem with this study was that it showed a relationship that was concrete (i.e. extremely statistically significant), but that was surprisingly lacking in explanatory power. Indeed, I discussed in that post how I’d ended up with a correlation that was far lower than Moz’s for Domain Authority.
Fortunately, Malcolm Slade recently discussed some of his very similar research at BrightonSEO, in which he finds similar broad correlations to myself between brand factors and rankings, but far, far stronger correlations for certain types of query, and especially big, high-volume, highly competitive head terms.
So what can we conclude overall from these third-party studies? Two main things:
- We should take with a large pinch of salt any study that does not address the possibilities of reverse causation, or a jointly-causing third factor.
- Links can add very little explanatory power to a rankings prediction model based on branded search volume, at least at a domain level.
The real world: Why do rankings change?
At the end of the day, we’re interested in whether links are a ranking factor because we’re interested in whether we should be trying to use them to improve the rankings of our sites, or our clients’ sites.
Fluctuation
The first example I want to look at here is this graph, showing UK rankings for the keyword “flowers” from May to December last year:
The fact is that our traditional understanding of ranking changes — which breaks down into links, on-site, and algorithm changes — cannot explain this degree of rapid fluctuation. If you don’t believe me, the above data is available publicly through platforms like SEMRush and Searchmetrics, so try to dig into it yourself and see if there’s any external explanation.
This level and frequency of fluctuation is increasingly common for hotly contested terms, and it shows a tendency by Google to continuously iterate and optimize — just as marketers do when they’re optimizing a paid search advert, or a landing page, or an email campaign.
What is Google optimizing for?
The above slide is from Larry Kim’s presentation at SearchLove San Diego, and it shows how the highest SERP positions are gaining click-through rate over time, despite all the changes in Google Search (such as increased non-organic results) that ought to drive the opposite.
Larry’s suggestion is that this is a symptom of Google’s procedural optimization — not of the algorithm, but by the algorithm and of results. This certainly fits in with everything we’ve seen.
Successful link building
However, at the other end of the scale, we get examples like this:
The above graph (courtesy of STAT) shows rankings for the commercial keywords for Fleximize.com during a Distilled creative campaign. This is a particularly interesting example for two reasons:
- Fleximize started off as a domain with relatively little equity, meaning that changes were measurable, and that there were fairly easy gains to be made
- Nothing happened with the first two pieces (1, 2), even though they scored high-quality coverage and were seemingly very comparable to the third (3).
It seems that links did eventually move the needle here, and massively so, but the mechanisms at work are highly opaque.
The above two examples — “Flowers” and Fleximize — are just two real-world examples of ranking changes. I’ve picked one that seems obviously link-driven but a little strange, and one that shows how volatile things are for more competitive terms. I’m sure there are countless massive folders out there full of case studies that show links moving rankings — but the point is that it can happen, yet it isn’t always as simple as it seems.
How do we explain all of this?
A lot of the evidence I’ve gone through above is contradictory. Links are correlated with rankings, and Google says they’re important, and sometimes they clearly move the needle, but on the other hand brand awareness seems to explain away most of their statistical usefulness, and Google’s operating with more subtle methods in the data-rich top end.
My favored explanation right now to explain how this fits together is this:
- There are two tiers — probably fuzzily separated.
- At the top end, user signals — and factors that Google’s algorithms associate with user signals — are everything. For competitive queries with lots of search volume, links don’t tell Google anything it couldn’t figure out anyway, and links don’t help with the final refinement of fine-grained ordering.
- However, links may still be a big part of how you qualify for that competition in the top end.
This is very much a work in progress, however, and I’d love to see other people’s thoughts, and especially their fresh research. Let me know what you think in the comments below.
First off, great article and great topic for discussion. It is refreshing to read thoughtful content like this. I thought I'd drop in some ideas for consideration.
> We should take with a large pinch of salt any study that does not address the possibilities of reverse causation, or a jointly-causing third factor.
We need to be careful to do this for all factors. Just as much as rankings might cause links, brand might cause links too. Massive advertising campaigns, newsworthiness, and other real-world interactions via brands can influence the link graph.
Moreover, brands can have spurious results with ranking correlations in other ways. For example, you could make the assumption that brands spend more on on-site SEO or faster websites which could influence rankings. Brands might already have a relationship with the customer creating higher engagement from the SERPs which might influence rankings.
I think we have to proceed with extreme caution before moving to any conclusion that "Google uses brand measurements as a ranking factor", even if it seems to have a great deal of explanatory value.
Finally, we do have a weapon against these types of issues - experimental studies. There are dozens of studies available on the web (and more conducted privately) which show a clear, causal relationship between link acquisition and improved ranking.
> Flux and CTR
I actually think that Google's addition of SERP features will increase the CTR for top results, rather than "drive the opposite". As a searcher makes a quick observation of the results revealed to them, if only 1 organic result appears (or the first 2 or 3), they may be more inclined than ever to click on those results rather than scroll and find the remainder. Moreover, as Google limits the number of total organic results on the initial results page (often 7 or 8), one would expect a steeper curve.
> brand awareness seems to explain away most of their statistical usefulness
Or, brand awareness just happens to correlate with a number of other actual ranking factors (links, content quality, site speed, user engagement).
> For competitive queries with lots of search volume, links don’t tell Google anything it couldn’t figure out anyway
But then again, we see sites like PolicyGenius ranking in the top 10 for "Life Insurance", a highly competitive, highly valuable term, despite having little brand presence relative to the big insurers.
> user signals
I have seen controlled studies of user signals like CTR give modest (1 position in a month, for example) rankings improvements, but I am not convinced these are sufficient to explain much about the top rankings. I don't mean to say they don't matter (I am almost certain that Google does measure these in one form or fashion), but call me skeptical as to their prominence.
> However, links may still be a big part of how you qualify for that competition in the top end.
I think this is a fair statement. I'm guessing Google has studied the relationship between each of its myriad ranking factors and user satisfaction. I imagine that links are not a highly sensitive factor in this regard (that is to say, a page with 100 backlinks is not significantly likely to please a user more than one with 99). We could even go so far as to imagine several types of relationships between links and rankings, such as a percent based approach where your backlink score is relative to other matching documents, thus the score for Document A1 and A2 with 1000 and 2000 links respectively for query A might be the same as Document B1 and B2 with 100 and 200 links respectively for query B. If a ratio of 1:2 is used, it is possible that other factors like content quality or user metrics might have an identical impact even though at first glance we would expect the site with 2000 links vs 1000 to have a wider advantage than the 200 vs 100 on raw count alone.
Ultimately, I think that Google's continued move towards machine-learned algorithms means teasing out individual ranking factors via correlation studies will become more and more difficult. The relationships between factors will can vary from query to query. Our best bet, in this regard, is to experiment and moderate. That is to say, we shouldn't rely on a single factor (link building for example) once we determine its effectiveness.
Hi Russ
I essentially agree with everything you've written here. Point by point:
Thanks for the considered reply!
Good points all around!
I am still quite skeptical of "Brand Awareness". I think most of what we could consider Brand Awareness is captured in other metrics and wouldn't add much more in the long run, but I'm not quite sure.
I think the problem is that there is too much generalisation about what works and what doesn't work in SEO. I think the main thesis that links are yesterdays ranking factors, is over simplistic as it depends on the sector. Even for the same sector in two different countries speaking the same basic language (e.g. USA vs UK). Some markets will get it via links and others by lots of other means such as online advertising. Some research:
When I used machine learning for snapshot SERPs data like I had for the Payday loans industry in the UK, "Do Follow Backlinks" crops up as 2nd most important factor that can explain the variation in Google rankings. This showed that in this sector, a site would gain 18 ranking positions on average by going from zero do follow backlinks to having do follow links from 200 domains.
Brand searches ended up as the 4th most important factor which could only forecast a 5 rank position gain in Google.
The overall ML model which was mainly onsite based achieved an r-squared of 42%, is considered predictive in the machine learning community. Anything less than 35% is not a consideration for predictive modelling, despite statistical significance. This makes sense as the model is missing data on the optimal taxonomy and keyword-URL mapping optimised for search intent. And that's before we even look at the content, IA and authority of the links.
In another case, I used ML (based on stochastic processes - the science of random events) to model SERPs over time for another industry (fresh fish) and found that it was potency that could explain the variation in rankings better than brand searches per se. The potency is defined as the distribution of domain authority across it's pages, so you may have two sites A and B with identical DA, but A has more potency because it takes less links to achieve the same DA.
The research showed that the potency of the site pages could most reliably explain the variation i.e. a high potency AHREFs rating for a site's pages could explain why 60% of the pages would gain 1 ranking position per day.
Looking deeper into the detail, the research showed that any site with their pages with less than a 50% potency rating of 40% would drop ranking positions anywhere between 0.1 to 0.2 per day.
The point is, brand is a ranking factor, but it's not the sole ranking factor and from what the research we've done across sites using machine learning models - it doesn't look like it will ever be.
I would show images, unfortunately this comment box doesn't allow me to insert images.
Are links yesterday? It's meaningless as a question. Authority/Brand is a major ranking factor and links are one expression of it as they are one (and probably the most efficient) way to increase brand awareness.
Any research needs to be sector specific and deploy statistical science (machine learning being the most effective) to ensure any knowledgeable understanding of the SERPs.
Hi Andreas
I'm not sure whether you're agreeing or disagreeing with me, or just adding commentary, but just to be clear: I'm not saying that brand is the sole ranking factor (I'm pretty confident there are thousands at this point), and I'm not saying that the picture is homogeneous across keywords or industries. Indeed, I'm saying the exact opposite - brand is only part of the picture, there is never a single explanatory factor, and you get very different results for different types of keyword.
I very much don't think it's a meaningless question to ask whether "links are yesterday". I do think it's one with an obvious answer - namely that they're certainly not gone entirely, and won't be for some time. But just because it seems obvious doesn't mean we shouldn't challenge it from time to time.
I also don't agree that research needs to be sector specific - this can be useful for sure, but sector boundaries are fuzzy, and when we're just looking to establish a bigger picture, it seems very unlikely that there will be completely different paradigms per sector in terms of whether Google uses a certain ranking factor whatsoever. That said, sector specific research is very valuable, and I'd love to see more of it.
Incidentally, I think treating machine learning as a special or specially category of statistics is fairly dangerous, especially for these sorts of questions where "traditional" statistics are equally applicable, which is very much the case here as you're essentially talking about multivariate regression - the same caveats and pitfalls do apply.
I'd also caution against claims like "an r-squared of 35% is considered predictive" - any r-squared of above zero is by definition predictive. It doesn't predict a huge amount of the variation in the data, sure, but we know there's a whole lot going on here beyond off-site factors. The more interesting question is how likely it is that our chosen variable is predictive purely by coincidence, and that's where statistical significance comes in. This is especially true when we look at many variables in many industries - we're bound to come up with some highly predictive variables sooner or later, and you can always get a very high R-squared just by adding in more and more random variables. Similarly so when we start to compare static variables with movements in rankings over time - clearly the stronger pages cannot go upwards forever, so we're at risk of introducing a great deal of noise.
Thanks for the response - it'd be great to see more research like you describe published here (or indeed elsewhere!)
Tom
Dear Tom,
Thank you for your feedback. A greater understanding of stochastics and machine learning would help you understand my feedback.
So I'm adding analysis based on robust statistical analysis. I appreciate your clarification that brand is not the sole ranking factor. So we're in agreement there.
I take the point back about the the question being meaningless. All questions are good, including the point about being ready to challenge. The question is nevertheless over simplistic because the better question in my view is are links the most effective source of authority? If so what types of links and why? How many tweets per URL or other metric would equate to an authoritative link?
You may disagree that research needs to be sector specific. You did write that the evidence you've found is contradictory which supports the findings of my research. My research quite often shows factors do vary not just across different industries - but also across countries for the same industry. Happy to share examples in more detail.
Machine learning is special and exciting, because right now, we're entering an age where statistical analysis on large datasets can be automated and become more sophisticated. ML brings alive many of the mathematical abstractions (beyond regression modelling) that were dreamt up by the likes of Bayes, Maclaurin and countless others centuries ago, which are now made possible by cost effective cloud computing, to be applied to the data rich field of SEO.
Why should search engines have all the fun?!
ML is for SEOs too.
To clarify, I'm not talking about multivariate regression. I'm using other abstractions that can deal with nonlinearities much better than any regression based model. R squared is simply one measure of seeing how the model performs.
A factor/feature that can only explain 1% of a dataset is not predictive, even from a common sense perspective.
My claims of "an r-squared of 35% is considered predictive" for a model as that is what is the general consensus is of any serious data science professional I've worked with. The error variable would account for the other inputs not explained or fed into the model, including those beyond offsite factors.
I'm glad you agree "the more interesting question is how likely it is that our chosen variable is predictive purely by coincidence". However, I'm afraid that isn't necessarily "where statistical significance comes in" it again depends on the model being used and how the test was being set up.
Also, you can't "always get a very high R-squared just by adding in more and more random variables", because you could be just adding more noise to the model. It's the addition of intelligent features that improves the model's predictive accuracy, not the mere number of features.
To infer the stronger pages go up forever isn't correct either and nor is that implied by the model. The model is describing what happens on average in the SERPs and is adding a predictive element of time, which is a far less risky and less noisy proposition than a non stochastic ML model.
I appreciate nobody is born with a knowledge of statistics let alone machine learning and stochastics. I hope the above clarifies as I think my feedback would be better understood with more knowledge. And hopefully provoke more questions on what ML can do for SEO practitioners. ML is an opportunity for the SEO community and it's discussions like these that will further our knowledge
We don't need to know the inner workings of your model algorithms as that is your IP, but we should speak a common accepted language when evaluating models or communicating the predictive value of SEO insight. Wouldn't you agree?
I'd be happy to share more research in this blog, when invited by Moz staff. Or you could read my blog.
Andreas
"Thank you for your feedback. A greater understanding of stochastics and machine learning would help you understand my feedback."
Of course! That said, based on what knowledge do have, I'm confident in standing by everything I said previously.
Just to address this final point:
"We don't need to know the inner workings of your model algorithms as that is your IP, but we should speak a common accepted language when evaluating models or communicating the predictive value of SEO insight. Wouldn't you agree?"
I would agree! But metrics like R-squared and statistical significance are easy to manipulate. The validity of the core logic is the key area where we need to avoid losing people, both in machine learning and in more traditional statistics, whether it be SEO or any other discipline.
Incidentally, there are no complex inner workings - my correlations come from mean spearman's correlations. The regression model is ordinary least squares.
Thanks,
Tom
Thanks for your feedback. We can respectfully disagree or chat in person next time we meet in London.
On r-squareds, I'd be very concerned if someone is manipulating their model validation metrics, r squareds or otherwise. Why they'd do that I don't know as they'd be cheating themselves. It doesn't make sense.
BTW Thanks for your questions at yesterday's STAT event which were both challenging and thought provoking. Next time, we must have a drink together.
Agreed on all fronts :)
"...it seems very unlikely that there will be completely different paradigms per sector in terms of whether Google uses a certain ranking factor whatsoever."
I'm inclined to disagree with this slightly. The value prop for each sector can and does differ in varying degrees, and how that value prop is scored from an algorithmic standpoint would thusly vary as well. Stone Temple's case study showed a good example (imo) of Google using different ranking factors for different types of content. It indicates to me that not all things are scored and ranked equally, but are on a case-by-case bases. This could easily extend to the overarching industry/sector situation.
I agree, I'm just saying it won't be binary. It might be that between ranking factors A and B, there is a 90/10 weighting in one situation and a 10/90 rating in another, but it will never quite reach 100/0 and 0/100.
I also think "sector" might be a pretty clumsy dividing line - "competitiveness", "intent", "scope" etc. might be interesting comparisons.
I agree with you here on the weighting... one of those "60% of the time, it works every time" scenarios. :D
We already know that intent has a LOT to do with SERP structure, and I think that it plays a large role in what factors are considered for SERP competitiveness.
Great discussion going on in these comments, well done Tom. Looking forward to learning more from you.
Hey Andreas,
I checked your site and case studies and did not see any methodology on your ML work in SEO (assuming you used stochastic gradient descent?). When you say your blog is there another area besdies the place you work at?
Thank you,
JP
Hi JP, I generally don't go into the specifics of my models (different ML for different types of SEO predictions) but I do in my Artios blog eg "Making Google Predictable" describe the general ML approaches I take to making SEO predictions. You could also check the deep crawl case study also.
Ahh looking for specifics. Recently finished up some course work at Stanford and working through some ideas, but no one in this field seems to believe in transparency (I come from a more traditional science). Appreciate the reply!
I'm still waiting for the day when search engines get so good, they will no longer have to rely on links. Chasing natural link opportunities for small companies/websites is becoming increasingly challenging. It can be a headache. Links being such a seemingly integral factor for ranking seems so old hat.
I'm with you. I am in a industry that links seem impossible for me to get. I've seo my site for a year and just built less then 20 links. My rankings are terrible, but still managed to get some traffics.
I agree. I work with clients with small budgets. They are paying for 5 to 8 hours. I could easily use up all that time chasing links that may or may not have a positive impact on ranking.
My tactic is to focus on getting link parity with the competition and then focus on a handful of high value links.
Rand spoke of turning link building into a flywheel and I still haven't figured out an efficient way of doing that for clients essentially paying for one day of SEO a month.
Just remember that it is link Quality, not Quantity that is important. This can save many hours of link building.
You only need a few good quality links as a small business to build enough authority. The rest is in your control, like solid on-page optimisation and proper keyword analysis and of course content. Good luck!
That's kind of like asking for democracy to get so good it doesn't need votes.
That's an interesting thought experiment - what would the equivalent replacement metric be?
What about if we replaced links with votes? Or votes with links?
Dear Tom,
Thank you for your feedback. A greater understanding of stochastics and machine learning would help you understand my feedback.
So I'm adding analysis based on robust statistical analysis.
The question is nevertheless over simplistic because the better question in my view is are links the most effective source of authority?
You may disagree that research needs to be sector specific. You did write that the evidence you've found is contradictory which supports the findings of my research. Happy to share examples in more detail.
Machine learning is special and exciting, because right now, we're entering an age where statistical analysis on large datasets can be automated and become more sophisticated. ML brings alive many of the mathematical abstractions (beyond regression modelling) that were dreamt up by the likes of Bayes, Maclaurin and countless others centuries ago, which are now made possible by cost effective cloud computing, to be applied to the data rich field of SEO.
Why should search engines have all the fun?!
ML is for SEOs too.
To clarify, I'm not talking about multivariate regression. I'm using other abstractions that can deal with nonlinearities much better than any regression based model.
A factor/feature that can only explain 1% of a dataset is not predictive, even from a common sense perspective.
My claims of "an r-squared of 35% is considered predictive" for a model as that is what is the general consensus is of any serious data science professional I've worked with. The error variable would account for the other inputs not explained or fed into the model, including those beyond offsite factors.
Re coincidence, I'm afraid that isn't necessarily "where statistical significance comes in" it again depends on the model being used and how the test was being set up.
Also, you can't "always get a very high R-squared just by adding in more and more random variables", because you could be just adding more noise to the model. It's the addition of intelligent features that improves the model's predictive accuracy, not the mere number of features.
To infer the stronger pages go up forever isn't correct either and nor is that implied by the model. The model is describing what happens on average in the SERPs and is adding a predictive element of time, which is a far less risky and less noisy proposition than a non stochastic ML model.
I appreciate nobody is born with a knowledge of statistics let alone machine learning and stochastics. I hope the above clarifies as I think my feedback would be better understood with more knowledge. And hopefully provoke more questions on what ML can do for SEO practitioners. ML is an opportunity for the SEO community and it's discussions like these that will further our knowledge
We don't need to know the inner workings of your model algorithms as that is your IP, but we should speak a common accepted language when evaluating models or communicating the predictive value of SEO insight. Wouldn't you agree?
I'd be happy to share more research in this blog, when invited by Moz staff. Or you could read my blog.
Andreas
Andreas, I believe you and Tom are getting at the same question - what's the equivalent metric/are links the most effective source of authority - are both essentially asking how to gauge popular opinion across the web in a statistically significant way. The free internet is a democracy, and search engines are trying to determine what the people want and provide it to them based on popular opinion aka the "Searcher Satisfaction" that Rand speaks of in your linked MozCon video.
The common analogy has always been, "links are votes." But after reading yours and Rand's articles and understanding a bit more of the mechanics, I think that analogy is more accurately expanded to "clicks are votes." With that in mind, it seems to me that we could one day move beyond looking at links for understandings of "authority" and user preference, and look more to performance metrics to gauge resonance with users and thus actual value. Even with that line of thought however, you run into circumstances that skew whatever thresholds you might want to establish (high bounce rates and low time on site for easily digestable or short-form content for example could easily be interpreted as a value "fail" without understanding the context, which just adds more variables to the mix).
All-in-all, Google is learning how to move away from this (Stone Temple's case study talks about this with it's point about valuable content) but a full transition is going to require Google innovating an effective way to utilize its behavioral data to accurately represent an understanding of what people want and what they're resonating with.
Great Post about the diminishing influence of links. I have noticed that for several medium competition keywords , it is possible to outrank high DA websites with more relevant content even if you have a fraction of the links or even no links at all. Recently Gary Illyes stated on Twitter (on 14th March) that it possible to rank medium competition keywords without backlinks even if the Domain age is just 1 month old.
Here in the Moz blog also several experts have agreed that it is possible for a small website to rank higher than a high DA website with more relevant content e.g. a hotel in a local area can out rank a travel agency operating nationally by creating high quality comprehensive content about their local area even though their DA is much less than that of the big travel agency.
This has been confirmed by me several times in the course of my work.
This post was great, Tom! After reading through the insightful posts AND the comments, I have a lot to take in and digest. But I have to say, the commentary on this post is some of the best I've seem on any Moz blog! Thank you for starting such an important discussion.
Thanks for the post. I love the fact that there are so many of us wondering what really affects rankings and what the role of links are these days. I find myself constantly monitoring rankings and noticing huge SERP volatility; asking all the same questions.
I literally limit my rankings checks because I am simply "too interested" (if you follow me) to the point of being emotionally dependent on my across the board rankings.
Rankings Up = I feel GOOD; Rankings down = Why did I ever get into SEO for a living?
SEO seems to be for the "meticulous." I keep a daily log (for nearly a year) of each site with everything I change/do.
When I see bumps in ranking I look back at what actions I took leading up to the rank spike.
Unfortunately it often leaves a lack of clarity... was it day 10 before the spike or what I did on day 7?
--Or... some other unknown factor such as how a group of visitors interacted with the site. I am seeing "branding" as one access point to better rankings, but it could simply be the nature of the brand signals coming back to the site in the form of links.
Crazy... just happy to know I'm not the only out there looking for all those undetected "x" factors that produce results, and I think most of us noticed that the ol' "chuck a few links" scenario doesn't work like it used to.
I recently had a solid sustained spike in my rankings and I noticed that about 2 weeks before that I simply updated all of the plugins on the site (among a couple of other factors). That made me think => Could it have been that? I figure if Google released an algo to monitor site related updates, that would single out a lot of pbns and junk sites. Most of us would ignore something so trivial, BUT what if?
I am running a case study on another site where that is the only change I have made. Now I have to wait.
Yeah - it's always difficult to separate it from the context of what changes your competitors have, or have not, made. That's why we're left dealing with correlations and wondering whether we're measuring the relevant factor, or something related to it!
Great insight Tom! I have to say I am still seeing a strong correlation between links in rankings on the local SEO side of things, I do however believe user signals will be come more and more a factor in the Google algorithm as time goes on.
Hi Tom - great write up overall. One thing I want to point out though, and this is why I still believe that links remain a very big deal, is that in our study that you linked to above (https://www.stonetemple.com/links-remain-a-very-po...) we provided more than just a correlation analysis. We also included a section called "Cementing the Point With Case Studies", in which we shared our experiences with a number of highly competitive keywords.
For client confidentiality reasons, we did not identify the specific keywords, but the case studies are real, and have been repeated many times over. We continue to see that a process of driving links to existing pages surely and steadily drives up their ranking. We've done this across several different marketplaces.
That does not mean, of course, that it works for all marketplaces.
Agreed! My caution is on correlation studies in general, not Stone Temple specifically. I also included a case study above which does show links moving the needle - albeit with some apparent lag or non-linearity.
No top 10 without links. Sometimes it's impossible to get some links and improve rankings. Sometimes we are forced to buy them, with the riskt it implies.
It's a great article to read. I also believe backlinks should not be consider as ranking factor.
Great post Tom, I believe both branding and links work together to help with SEO, and often working on quality links and brand mentions or NAP listings to your business will help with your local brand awareness as well. Very interesting topic and well covered sir.
Hi Tom,
Thank you for the above examples and stats. I think the complete scenario depends on the in which industry you are. We all know that there are many websites ranking top with black-hat techniques, but it could be just temporary. Links are still one of the major criteria Google still consider for ranking.
I think SEO it is not only links, altough of course they are very important at this moment.
Great Article Tom, thanks for sharing!
I always think talks about links are interesting, but I always make sure to keep the user in mind. If an article does not have too many referral links that were created, I would still be curious about the quality of those referral links. Overall getting links is great (and seems to have a correlation with rank), but being referenced in the right areas for your niche and helping the business would be my bottom line.
I am a doer (but a thinker) in a digital marketing agency. I do not develop strategies, but, I read research and share my opinion. I shared your conclusions with our agency's principals as we are discussing modifying existing strategies re LINKS. My own conclusion (not solely based on your post) is that if it ain't broke, don't fix it. Spend time and energy going after some of the other actionable factors effecting organic rankings recognizing that even if you think that Google is changing emphasis on other variables, don't change what has worked; don't change what has historically ranked your clients in that 'top-end'. Links still matter. "Don't complicate the ham sandwich" (or the banger n' mash). Thank you for your articles' insights.
Hey Tom, thanks for the article. I think that nowadays the social media are taking over the purpose of links which they had some years ago. In my opinion, links were sign of relevance and social proof in some way which is more and more the matter of social media nowadays. At least, it seems there is certainly a raising correlation. Looking forward the next article! Cheers, Martin
I'm still in the same boat as you, Tom.
I've been checking competitor analysis on a certain keyword phrase and almost 80% of the top ranked sites do not have a high volume of links pointing to them or at least to a particular page. I can only attribute their SERP positions to their DA and overall site traffic.
Working for a smaller company, it leaves an opening to try and obtain strong value links to see if we can compete against the bigger fish as, in my sector anyway, the bigger guys don't seem to worry so much on SEO: they simply use their size to their advantage.
Very interesting research and read here, thank you Tom. I think links will continue to be a major ranking signal for some time. I think that this will change as the web innovates and changes more but for the time being I don't see a better signal that could replace links.
Links and content are still king. However, what I've recently noticed is how much 'better' Google has become interpreting links and being able to identify good links. For example, I find directories to still have some worth today but nowhere near how much they used to be worth whilst links on relevant sites, written well with a wide variety of anchor texts are now worth more.
Whether or not they appear to be worth more because other links are now worth less I am not so sure, but from what I can say building the right links is as important today as it has ever been post Penguin 1.0.
Educated guess based on personal results - links and content primary ranking factors until page 1, then user signals (CTR, dwell time, etc) weighted heaviest. Google takes an algorithmic stab at determining best content for given query but ultimately leaves to users to judge. This is where brand awareness and affinity come into play. If a known brand ranks, it likely gets clicked
That's true. Still links are the biggest factors for ranking. But it should be well balanced with on-page and content marketing. If you have done awesome on-page optimisation, your keyword will start to appear in ranking only with content marketing activities.
Nice Article Tom!!!
I think wondering whether or not a link is a ranking factor, i think they are. There are many factors but if we "compress" these factors we get On-Page, Off-Page and Technical SEO. Now when you look at link building, obviously you need very good links. But making sure that the sites are relevant that rank to you is a key factor of this. I have noticed that people still pay for links on "link directories" if you will. There's one page with loads of links which has no point what so ever. Even if the site hat the best Trust Flows, Citation Flows, Page Authority or what ever, you are still losing a lot of this link juice in internal and external links. Google will also see these as spam and soon in my predictions the sites that are on these types of sites will all be penalised.
"Google will also see these as spam and soon in my predictions the sites that are on these types of sites will all be penalised."
It would be a bit risky and very out of character for them to ever take a blanket approach like that. After all, I could put your agency on one of these sites without you doing it yourself...
Exactly. The links would more likely be devalued rather than penalised.
Yes you are right. But what i meant was they would penalise the site hosting all of these links then manually check the sites backlinks profiles that are on there to see how common it is. In my tests i have noticed my sites have ranked a lot better with out these sorts of links. Personally i think this is already in use at the minute.
Someone did an analysis of sites with bad neighbors on the hosting?
Hi Tom,
I couldn't agree more about the importance, still, of link building. You'll get nowhere without good content and link building is kinda like crowd-surfing, you'll just fall flat on your face otherwise. Enjoyed reading the article very much!
Thanks,
Deepak
Links are outdated? Nope, just have a look at Googles's Serps; overall winners are those Sites with the strongest Link profiles. Megasites, well Google likes it big.
Bing on the other hand does again what Google announced. Sites with hardly any link have a chance to rank.
I understand its pleasant to talk about the right content and explain how much Google knows about all websites. To know that much Google needs to host all websites. Google knows a lot but it is far from being allmighty.
I've been focused on creating new contents, but now, after reading this awesome article, it seems to me that link building also plays a big role to rank up a content. But, there are lots of ways to build links. Among them, how to know what are the most successful one?
Do forum posting, blog commenting provide any value to link building? I know, guest posting is very effective, but it requires much time. So, what can be a little easier way to build successful links?
Thanks for your post Tom !!
It is always interesting to read them and I always learn new things.
Using an R-squared value presumes the underlying relationship is linear. I firmly believe that the the link-ranking relationship is monotonic, and positive but I could image several nature inflection points. Are links 1-10 equally as valuable as links 90-100? Or another way visualize the response space is do you believe position 1 and position 2 are have the same separation as position 9 and position 10? Strictly speaking, a low R-squared value indicates that the relationship between the two factors is not linear which is different from the factors not being correlated.
Hi Coleman, not necessarily. The r squared in my case is used as an evaluation metric, however the underlying ML model isn't a linear model, and therefore overcomes non linearities that are common in SEO.
I am a physicist and do stochastic modeling on a regular basis. When building a regression model it's critical to check to goodness of fit. You stated your ML model was non-linear by design. I refer to the paper title - "An evaluation of R2 as an inadequate measure for nonlinear models in pharmacological and biochemical research: a Monte Carlo approach" The paper opens -
"Background
It is long known within the mathematical literature that the coefficient of determination R2 is an inadequate measure for the goodness of fit in nonlinear models. Nevertheless, it is still frequently used within pharmacological and biochemical literature for the analysis and interpretation of nonlinear fitting to data.
ResultsThe intensive simulation approach undermines previous observations and emphasizes the extremely low performance of R2 as a basis for model validity and performance when applied to pharmacological/biochemical nonlinear data..."
Although it's a different field of pharmacology and not SEO, the fundamental mathematics associated with machine learning and model building should apply. R2 is a simple and useful tool but it is not a definitive test for nonlinear data.