Uncovering 'not provided' keyword data

How does the Hero do it?
Extended Version

Table of Content
1. How it Doesn’t Work
2. Understanding the Domain
3. Generating Keywords
4. Clustering and Classifying Keywords
5. Teaching the Hero
6. The First Computation
7. The Final Output

1. How it does not work
For Clarification Purposes

There are no unique keys that would allow you to map a keyword with
a session. For this very reason we use several data sources to reduce the
number of possible keywords for a specific session to the minimum.
Key Keyword Key Session

2. Understanding the Domain
First Things First

Once a user signs up, the Hero analyzes the history of
the GA account, looking for:
1. patterns e.g. site structure, URLs that pull traffic; clusters among
organic traffic; device categories; time / date; locations; …
2. type and content e.g. # of direct sessions attributed to
organic; news or ecomm; one-pager or many product sites;
semantical topics; …

3. Generating Keywords
What could it be???

In the first step we only generate a large set of
possible keywords per URL.
There is no classification or attribution in this step.
There are two types of data sources we use for this:
1. 3rd party data sources e.g. rank monitoring services;
browser extensions data; Bing search API; ....
2. 1st party user data Search Console; remaining keywords in
GA

→ the result is one big bucket with
possible keywords per URL
Keyword 1
Keyword 2
Keyword 3
Keyword 4
...
YOURDOMAIN.COM/FOLDER

4. Clustering and Classifying
Keywords
It’s getting tricky...

Using external APIs we look for anormal behaviour in
a keyword’s history
were there changes in
traffic in the last 12
months
(per country / device)
is this a new keyword?
did spikes occur?
is there a reference
article on Wikipedia and
how does it behave?
→ cool link

Semantics
After the Hero created a keyword set and checked for
potential spikes and anormalities, he needs to
“understand” the keyword.
is this a brand or a
non-brand keyword?
is it a person / brand
name? (important for
ngram analysis)
is the keyword
transactional /
navigational or
informational?
category clusters

→ the result is one big dataframe
where certain parameters are
attributed to the keywords
(looks sort of like this thing above)

In the next step we analyze differences in the
performance of organic traffic due to keyword
fluctuation in the SERPs.
suppose we have a site that ranked on position 13 - 15 in calendar week X for the term
“shoes”. At this point we have the following metrics for this landing page:
in calendar week X + 1 the site now ranks on position 5 for shoes and metrics changed to:
in this simple example we assume ceteris paribus, that nothing else has changed, so almost
all new sessions can be attributed to “shoes”.

these differences tell us a couple of things:
1. “shoes” on position 5 brings in 1000 more daily sessions compared to position 13 - 15
2. the bounce rate of the keyword is identical to other keywords that rank for this site
3. the conversion rate of this keyword seems to be significantly lower than the CR of the
others. It has brought only 5 more conversions with 1000 sessions, compared to the 10
conversions with the 500 sessions we had before.
→ when analyzing this, we take a couple of factors into consideration, most importantly:
- seasonal fluctuations
- overall site performance during this time frame
- keyword performance → spike due to trends?
Luckily the SERPs are changing quite a lot, otherwise this would not be
possible.

5. Training Classifications
This is where his brains kick in...

A first step towards matching keywords
and sessions
Using the aforementioned generated data, the Hero tries to calculate a
first probability of a certain keyword matching a cluster of sessions.
Sessions that were captured thru extensions and those where the
keyword is still visibly in GA (= “hard data”) can be matched with 100%
certainty, they will be left out.
→ after the first probability is calculated, the data will again be
compared against the “hard data”. This happens on the cluster level to
adjust the classifications.

The second step towards matching
After the first iteration the Hero constructs strings from the GA data.
These strings contain between 5 and 25 dimensions.
(It's interesting how few sessions are left if you query a big string - even
if you have huge sites. You can check this for yourself using the Core
Reporting API, where you get up to 7 dimensions in one string)
Now: the keywords have been clustered really well and you get only a
handful of sessions back from one string, leaving very few potential
keywords per cluster of session (the cluster has gotten smaller, too)

An unfair visual representation
Keyword A
Keyword A might now be matched with one or more of those sessions.
This is a session
cluster, containing
four sessions.
This results from the
Hero requesting all
sessions with a certain
string
Session 1
Session
2
Session
3
Session
4

Training the classifications
What happened now: the Hero created an adjusted unique algorithm
for your account, that tries to match sessions with the keywords.
He applies this algorithm to the entire history of the GA account and
then matches the results with the “hard data”, the remaining visible
keywords and the keywords captured by browser extensions.
After this last step the algo is ready to go for each page.
In most cases the historic data is not enough as some data sources are
not available retrospectively, which is why the Hero works far better
after a month.

The learning curve
MatchingCertainty
Time
Entire algorithm
Your algorithm

6. The First Computation
The magic happens...

After all that learning, the Hero can start
The process is essentially the same as the training
process:
1. Computing possible keywords
2. Checking the results against the “hard data”
3. Adjust classifications based on results
Actually, quite straightforward!

7. The Final Output
What are you giving me here?

What data do you get?
Obviously the matching does not happen thru CIDs.
Looking in that mirrored account, you’ll see that
some dimensions are missing. This is because the
matching happens based on dimension strings and
not sessions.

Session vs. string-based matching
If the data was session based, you’d get all info:
but what it looks like is this:
{{Keyword}} GA-CID =
All Dimensions and
Metrics for this CID
{{Keyword}}
GA-CID =
GA-CID =
similar
metrics
same cluster
of
dimensions

What do the 83% certainty mean?
This is just the average threshold but easier to
communicate.
Really, the certainty threshold for a keyword to
match with a session varies between 80% and 85%.
The certainty level is based on the assumption that
95% of all possible keywords have been found in the
first bucket in page 9.

This keeps the Hero from top shape:
- you had a major redesign / relaunch / big change
of your website resulting in very different metrics.
- the Hero can’t recognize your brand or all of your
(misspelled) brand terms. We adjust this manually.
- your GA account is in bad shape.

We hope that helped. We are certain about the
validity of the mathematical concepts involved here.
We have to remain quite top level mostly, as the Hero
tries to keep some secrets :-)
However, if you have any further questions about
some of the concepts involved, we try our best to
answer them. Just let us know @
tech-questions@keyword-hero.com

How does the semantic clustering
work?
After all possible keywords that rank were clustered
and analyzed, the classification will be adjusted to
what the Hero knows from analyzing the browser
extension data and their users.
→

An example:
Search phrase: “How tall is the Eiffel Tower?”
Is the data set from the browser extensions large
enough to attribute metrics to this keyword?
YES: how deep can you go? can you even attribute
on country / device level?
NO: can we get an approximation of the
performance from the columns in the dataframe?

Some stats about the browser
extensions
We buy data from some browser extension conglomerates across the
globe. Depending on vertical and country the ratio of desktop searches
that we catch is between 0.5% and 8%.
For many URLs, however, we don’t have a matching search phrase
captured by a plugin. Still the data is very important as it trains the
overall algorithm keyword related stats, such as:
“clicks on paid terms vs. organic”
“intention vs. vertical based click behaviour”

Uncovering 'not provided' keyword data

Related slideshows

Recommended for you

Recommended for you

Recommended for you

Recommended for you

Recommended for you

Recommended for you

Recommended for you

Recommended for you

More Related Content

What's hot

What's hot (19)

Similar to Uncovering 'not provided' keyword data

Similar to Uncovering 'not provided' keyword data (20)

Recently uploaded

Recently uploaded (20)

Uncovering 'not provided' keyword data