Google's evolution into deep learning has created a whole new kind of algorithm; one that differs substantially from the type of ranking system SEOs & marketers have become used to over the past 17 years. In this presentation, Rand explores the changes Google's made, and how it impacts the actions necessary to be successful in 2016 and beyond.
8. By 2007, Link Spam Was Ubiquitous
This paper/presentation from
Yahoo’s spam team in 2007
predicted a lot of what Google
would launch in Penguin Oct,
2012 (including machine
learning)
9. Even in 2012, It Felt Like Google Was Making Liars Out
of the White Hat SEO World
Via Wil Reynolds
10. Google’s Last 3 Years of
Advancements Erased a
Decade of Old School
SEO Practices
11. They Finally Launched EffectiveAlgorithms to Fight
Manipulative Links & Content
Via Google
12. And They Leveraged Fear + Uncertainty of
Penalization to Keep Sites Inline
Via Moz Q+A
13. Google Figured Out Intent
Rand probably doesn’t
just want webpages
filled with the word
“beef”
14.
15. They Looked at Language, not Just Keywords
Oh… I totally
know this one!
16.
17. They Predicted When We Want Diverse Results
He probably
doesn’t just want a
bunch of lists.
18.
19. They Figured Out When We Wanted Freshness
Old pages on this
topic probably
aren’t relevant
anymore
29. In 2012, Google Published a PaperAbout How
they Use ML to Predict Ad CTRs:
Via Google
30. 2012
“Our SmartASS system is a
machine learning system. It
learns whether our users are
interested in that ad, and
whether users are going to click
on them.”
31. By 2013, It Was
Something Google’s
Search Folks Talked
About Publicly
Via SELand
32. As MLTakes Over More of Google’sAlgo, the
Underpinnings of the Rankings Change
Via Colossal
33. Google is PublicAbout How They Use MLin Image
Recognition & Classification
Potential ID Factors
(e.g.color,shapes,gradients,
perspective,interlacing,alttags,
surroundingtext,etc)
Training Data
(i.e.human-labeledimages)
Learning
Process
Best
Match
Algo
34. Google is PublicAbout How They Use MLin Image
Recognition & Classification
ViaJeffDean’sSlidesonDeepLearning;aMustReadforSEOs
35. Machine Learning in Search Could Work Like This:
Potential Ranking
Factors
(e.g.PageRank,TF*IDF,
TopicModeling,QDF,Clicks,
EntityAssociation,etc.)
Training Data
(i.e.good&badsearchresults)
Learning
Process
Best Fit
Algo
38. The Machines Learn to Emulate the Good Results & Try to Fix
orTweak the Bad Results
Potential Ranking
Factors
(e.g.PageRank,TF*IDF,
TopicModeling,QDF,Clicks,
EntityAssociation,etc.)
Training Data
(i.e.good&badsearchresults)
Learning
Process
Best Fit
Algo
39. Deep Learning is Even MoreAdvanced:
Dean says by using deep learning,
they don’t have to tell the system
what a cat is, the machines learn,
unsupervised, for themselves…
41. Googlers Don’t Feed in Ranking Factors… The Machines
Determine Those Themselves.
Potential Ranking
Factors
(e.g.PageRank,TF*IDF,
TopicModeling,QDF,Clicks,
EntityAssociation,etc.)
Training Data
(i.e.goodsearchresults)
Learning
Process
Best Fit
Algo
42. No wonder these guys are stressed about Google
unleashing the Terminators
Via CNET & Washington Post
44. Googlers Won’t Know Why Something Ranks or
Whether a Variable’s in theAlgo
He means other Googlers.
I’m Jeff Dean. I’ll know.
45. The Query Success Metrics Will BeAll That
Matters to the Machines
Long to Short Click Ratio Relative CTR vs. Other Results
Rate of Searchers Conducting
Additional, Related Searches
Metrics of User Engagement
on the Page
Metrics of User Engagement
Across the Domain
Sharing/Amplifcation Rate
vs. Other Results
46. The Query Success Metrics Will BeAll That
Matters to the Machines
Long to Short Click Ratio Relative CTR vs. Other Results
Rate of Searchers Conducting
Additional, Related Searches
Metrics of User Engagement
on the Page
Metrics of User Engagement
Across the Domain
Sharing/Amplifcation Rate
vs. Other Results
If lots of results on a SERP do
these well, and higher results
outperform lower results, our
deep learning algo will consider
it a success.
47. We’ll Be Optimizing Less
for Ranking Inputs
Unique Linking Domains
Keywords in Title
Anchor Text
Content Uniqueness
Page Load Speed
48. And Optimizing More for Searcher Outputs
High CTR for this position?
Good engagement?
High amplification rate?
Low bounce rate?
Strong pages/visit after
landing on this URL?These are likely to be the criteria of
on-site SEO’s future… People return to the site
after an initial search visit
49. OK… Maybe in the future.
But, do those kinds of
metrics really affect SEO
today?
55. 40 Minutes & ~400
Interactions Later
Moved up 2 positions after 2+ weeks
of the top 5 staying static.
56. 70 Minutes & ~500
Interactions Total
Moved up to #1.
57. Stayed ~12 hours, when it
fell to #13+ for ~8 hours, then
back to #4.
Google? You
messing with us?
58. Via Google Trends, we can see the relative impact
of the test on query volume
~5-10X normal volume over
3-4 hours
59. BTW –This is hard to replicate.600+
real searchersusinga varietyof
devices,browsers,accounts,geos,etc.
willnot lookthe same to Googleas a
Fiverr buy,a clickfarm,or a bot.And
note how G penalizedthe page after the
test…They might not put it back if they
thoughtthe site itselfwas to blame for
the clickmanipulation.
69. Optimizing the Title, Meta Description, & URL
a Little for KWs, but a Lot for Clicks
If you rank #3, but have a higher-than-
average CTR for that position, you might
get moved up.
Via Philip Petrescu on Moz
70. Every Element Counts
Does the title match what
searchers want?
Does the URLseem
compelling?
Do searchers recognize
& want to click your
domain?
Is your result fresh? Do
searchers want a newer
result?
Does the description
create curiosity & entice
a click?
Do you get the brand
dropdown?
71. Given Google Often Tests New Results Briefly on Page One…
ItMayBeWorthRepeatedPublicationonaTopictoEarnthatHighCTR
Shoot! My post only made it to #15…
Perhaps I’ll try again in a few months.
72. Driving Up CTR Through Branding Or Branded
Searches May GiveAn Extra Boost
78. Speed, Speed, and More Speed
Delivers the Best UX on Every Browser
Compels Visitors to Go Deeper Into Your Site
Avoids Features thatAnnoy or Dissuade Visitors
Content that Fulfills the Searcher’s Conscious &
Unconscious Needs
An SEO’s Checklist for Better Engagement:
79. Via NY Times
e.g. this interactive graph
that asks visitors to draw
their best guess likely gets
remarkable engagement
80. e.g. Poor Norbert does
a terrible job at SEO,
but the simplicity
compels visitors to go
deeper and to return
time and again
Via VoilaNorbert
83. Google’s looking for
content signals that a
page will fulfill ALL of a
searcher’s needs.
I think I know a few
ways to figure that
out.
84. ML models may note that
the presence of certain
words, phrases, & topics
predict more successful
searches
85. e.g. a page about New York that doesn’t
mention Brooklyn or Long Island may not be
very comprehensive
86. If Your Content Doesn’t Fill the Gaps in Searcher’s Needs…
e.g. for this query, Google might
seek content that includes topics
like “text classification,”
“tokenization,” “parsing,” and
“question answering”
Those Rankings Go to Pages/Sites That Do.
87. Moz’s Data Science Team
is Working on Something to
Help With This
The (alpha) tool extracts likely
focal topics from a given
page, which can then be
compared vs. an engines top
10 results
90. Pages that get lots of
social activity &
engagement, but few
links, seem to
overperform…
91. Google says they
don’t use social
signals directly, but
examples like these
make SEOs
suspicious
92. Even for insanely competitive
keywords, we see this type of
behavior when a URLgets
authentically “hot” in the
social world.
93. Data from Buzzsumo & Moz
show that very few articles
earn sharesAND that links &
shares have almost no
correlation.
Via Buzzsumo & Moz
94. I suspect Google doesn’t
use raw social shares as
a ranking input, because
we share a lot of content
with which we don’t
engage:
Via Chartbeat
95. Google Could Be Using a Lot of Other Metrics/Sources to Get
Data That Mimics Social Shares:
Clickstream (from Chrome/Android)
Engagement (from Chrome/Android)
Branded Queries (from Search)
Navigational Queries (from Search)
Rate of Link Growth (from Crawl)
96. But I Don’t Care if It’s Correlation or Causation;
I Want to Rank Like These Guys!
97. BTW – GoogleAlmost Certainly Classifies SERPs
Differently & Optimizes to Different Goals
These URLs have loads of shares & may have high loyalty, but
for medical queries, Google has different priorities
98. Raw Shares & LinksAre Fine Metrics…
Via Buzzsumo
99. But If the Competition Naturally Earns
Them Faster, You’re Outta Luck
4 new shares/day
2 new shares/day
3 new shares/day
10 new shares/day
100. And Google Probably Wants to See Shares that
Result in Loyalty & Returning Visits
101. New KPI #1: Shares & Links Per 1,000 Visits
Unique Visits
÷
Shares + Links
Via Moz’s 1Metric
102. New KPI #2: Return Visitor Ratio Over Time
Total Visitor Sessions
÷
# of Returning Visitors
103. Knowing What Makes OurAudience (and their
influencers) Share is Essential
From an analysis of the
10,000 pieces of content
receiving the most social
shares on the web by
Buzzsumo.
104. Knowing What Makes them Return (or prevents
them from doing so) Is, Too.
105. We Don’t Need “Better” Content… We Need “10X” Content.
Via Whiteboard Friday
Wrong Question:
“How do we make something as
good as this?”
Right Question:
“How do we make something 10X
better than any of these?”
106. 10X Content is the Future, Because It’s the Only Way to Stand
Out from the Increasingly-Noisy Crowd
http://www.simplereach.com/blog/facebook-continues-to-be-the-
biggest-driver-of-social-traffic/
The top 10% of content
gets all the social shares
and traffic.
107. Old School On-Site Old School Off-Site
Keyword Targeting Link Diversity
Anchor Text
Brand Mentions
3rd Party Reviews
Reputation Management
Quality & Uniqueness
Crawl/Bot Friendly
Snippet Optimization
UX / Multi-Device
None of our old school tactics will get this done.
109. Broad search Narrower search
Even narrower search
Website visit
Website visit Brand search
Social validation Highly-specific search
Type-in/direct visit Completion of Task
Google Wants to Get SearchersAccomplishing
Their Tasks Faster
110. Broad search
All the sites (or answers) you probably would
have visited/sought along that path
Completion of Task
This is Their Ultimate Goal:
115. APage ThatAnswers the Searcher’s Initial Query
May Not Be Enough
Searchers performing this query
are likely to have the goal of
completing a transaction
116. Google Wants to Send Searchers
to Websites that Resolve their
Mission
This is the only site where
you can reliably find the
back issues and collector
covers