Artykuł How to double conversion of a market leader’s campaign [case study] pochodzi z serwisu Adequate.

]]>The following case study shows how numerous improvements in our Client’s campaigns, made it possible to increase sales without sacrificing profitability. It was nominated to European Search Awards in the “Best PPC Campaign” category.

Vola.ro Performance Marketing Case Study

An important part of this success was the video campaign, with hard evidence of effectiveness measured by conversion lift study:

Vola.ro YouTube Campaign – Conversion Lift Study

This TrueView for Action campaign was also published as a case study by Think With Google.

Artykuł How to double conversion of a market leader’s campaign [case study] pochodzi z serwisu Adequate.

]]>Artykuł Markov Chain Attribution Modeling [Complete Guide] pochodzi z serwisu Adequate.

]]>The Markov chain is a model describing a sequence of possible events in which the probability of each event depends only on the current state.^{}

An example of a Markov chain may be the following process:

I am going for a week’s holiday. The likelihood whether I will do sport or just relax there, depends on where I spend my vacation (mountains or beach). The risk of injury during relaxation is negligible, while sport involves 1/10 probability of an accident:

The graph shows that an accident can happen when going to the beach as well as and in the mountains.

The probability that during given vacation I will go to the mountains AND I will have an accident there (i.e. START > Mountains > Injury), is:

**P _{(injury, mountains) }**= 7/10 × 8/10 × 1/10 = 56/1000 = 5.6%

The likelihood that the given holiday I will spend on the beach AND I will have an accident there (i.e, START > Beach > Sport > Injury) is:

**P**_{(injury, beach) } = 3/10 × 1/10 × 1/10 = 3/1000 = 0.3%

In this process, there is no other way to have an accident. The total probability of an accident is thus:

**P**_{(injury) } = P_{(injury, mountains) } + P_{(injury, beach)} = 5.6% + 0.3% = 5.9%

It explains why I return home with a leg in a cast every 5-6 years if I travel three times a year.

How can we use Markov chains to analyse multi-channel conversion paths?

Imagine that we have only four paths, two of which are converting. Conversion rate is therefore 50% (yes, it is high, but on such numbers it will be easier to understand the calculations).

These paths can be presented as the following graph:

For better readability, instead of multiple arrows, we will use single arrows (arcs) labelled with the number of conversion paths included in the given arc.

This number, divided by the count of all arcs outgoing of the given node, represents the probability of transition between the graph nodes:

Let’s see now what is the probability of conversion in this graph.

In order to calculate it, we have to sum up the probabilities of conversion for all possible paths in the graph that lead from the START node to the CONVERSION node.

In this case, there are three such paths, marked in different colours:

The total probability of conversion in this graph is 1/2 (i.e. 50%), which is the same as the conversion rate of the analysed paths (four paths, two conversions).

Please note that the paths of transitions in the Markov chain graph leading from START to CONVERSION or NULL (no conversion), are not the same as the conversion paths.

The Markov chain attribution modeling is based on the analysis of how the removal of a given node (a given touchpoint) from the graph affects the likelihood of conversion.

Let’s see what happens if we remove Facebook.

Arcs (arrows) outgoing of this node will cease to exist. Therefore, in such a truncated graph there is only one path from START to CONVERSION (marked in red). Its probability is 1/9:

Similar calculations made for Google, give a result that after removing this node, the probability of conversion is 1/6:

After removal of remarketing, there is no path from START to CONVERSION in this graph. Therefore, the probability of conversion without remarketing is zero:

The next step is to calculate the *Removal Effect* for each node of the graph. It is calculated according to the formula:

The removal effect represents the percentage of all conversions that will likely to be lost if you remove the given channel (touchpoint).

The removal effects of all nodes do not sum up to 100%, therefore in order to calculate the shares in the result, the removal effects must be normalised, i.e. reduced proportionally so that they total up to 1 (100%).

Of course, in order to calculate the conversions attributed to a given channel, we have to multiply the share in the total result by the total number of conversions. In our case, we have 2 conversions (two paths with one conversion each). The attribution in the Markov chain model in this case will look like this:

There are no instant reports including non-converting paths in Google Analytics

*You can try to obtain the non-converting paths data by creating a ‘pageview’ goal (any visit to the website is a conversion) or extract data about visits using segments in Google Analytics, or, preferably, use another tool to track interactions and conversions, e.g. Campaign Manager (Google Marketing Platform).*

The *Top Conversion Paths* report in Google Analytics contains only converting paths:

It is possible to build Markov chains using only converting paths. The converting paths in our example campaign are shown below:

The Markov chain graph will look like this:

In such a Markov chain the probability of conversion is 1, because all paths in the graph lead to CONVERSION.

As in the previous examples, let’s calculate the removal effect for individual channels. The likelihood of conversion after removing Facebook is zero:

Without Google, the probability of conversion is 1/2:

After removal of the remarketing, the probability of conversion is zero:

We can calculate now the shares in the total result for each individual channel:

An interesting fact is that in this case the Markov chains model gives the same attribution as the linear model. It’s not a coincidence.

However, this is not a rule. The Markov chains and linear models will give identical results only in case of less complex graphs. In case of the conversion paths below, the result will be different:

The graph based on the paths above will only be a bit more complex than in the previous example, but what makes it different, it’s the presence of loops in the graph (marked in yellow):

Loops in Markov’s chains significantly complicate the calculation of the conversion probability. The complication results from the fact that there are infinitely many possible paths leading from START to CONVERSION in the graph, because each loop can be entered any number of times.

Let’s see how it looks like in case of simplest loops.

Say, we have two converting paths:

The Markov chain graph for these interactions is as follows:

We know that the total probability of conversion is 1 (all paths convert), but let’s see how we can calculate it in the graph. The simplest path from START to CONVERSION does not contain any loop and represents the START > Facebook > CONVERSION path in the graph. Its probability is 2/2 × 2/3:

You can also get from START to CONVERSION by making one loop through the Google node: START > Facebook > Google > Facebook > CONVERSION.

After adding it, the probability of conversion increases by another component:

A double loop (START > Facebook > Google > Facebook > Google > Facebook > CONVERSION) generates another component of the conversion probability:

After adding infinite number of loops, the total probability of conversion will approach 1 (100%) in the limit:

It becomes even more complicated if there is more than one loop in the graph. See the example below:

The Markov chain graph will look like this:

The calculations in this case will be more complex. The transitions on the paths may enter both loops (loops via Google or via Facebook node) in a different order and we have to take into account all possible permutations.

In this case, there are two paths including one loop, four paths including two loops, 8 paths including three loops and for n loops there are 2^{n} possible paths:

In the case of more complex graphs it will be even more complicated.

It is obvious that the calculation of Markov chains must be done using numerical methods. We have built a simple tool (beta) for Markov chains attribution calculations. However, before you start using it, it is worth to read this article to the end.

See what happens if in our example conversion paths…

… one repetition of interaction (Facebook) would appear on one of the paths:

This modification of the graph would look as shown on the picture below. At the top there is a graph without repeating interaction with Facebook, and below you can see the same graph including a loop representing the repetition of the interaction with Facebook:

Calculation of such a loop is trivial. See the picture below: The removal effect of the node without the repetition (at the top) will be exactly the same as the removal effect of the node with the repetition (below):

Therefore, for the purpose of calculating the probabilities of transition in the graph, and in consequence, for the purpose of attribution modeling, repetitions do not change anything. Graphs with and without repetitions are equivalent.

In other words, if there are repetitions on the conversion paths, e.g.:

Facebook > Facebook > Google > Remarketing > Remarketing

this path is equivalent to a path without these repetitions:

Facebook > Google > Remarketing

The repetitions can be simply ignored, and we can replace the repeating interactions with a single interaction.

According to the definition, the probability of transition to a given state in the Markov chain depends solely on the current state. We can say that the nodes of the Markov chain “have no memory”.

What does this mean in practice? Take a look at the graph below.

The probability of conversion after interaction with Remarketing is 2/3 – regardless of whether the previous visit was from Facebook or Google:

We know that this is not really true. The effectiveness of remarketing will be radically different depending on how the audience was built, whether it contains users who previously searched for your product on Google, or people who has just shown interest in your post on Facebook.

This issue is solved by higher-order Markov chains. In the case of 2^{nd} order chains, the probability of transition to a subsequent state depends not only on the current state but also on the previous state. Therefore, instead of individual interactions, we will analyse pairs of interactions:

The Markov chains graph will look like this:

Calculations of conversion probability are made the same way as in the case of the first order Markov chains. The probability of conversion (sum of three possible paths in the graph, marked with colours) calculated using this graph is still 1/2:

We calculate the removal effects similarly as in the case of 1^{st }order Markov chains.

The removal of Facebook will erase all pairs of interaction containing Facebook and the probability of conversion decreases to 1/8:

Similarly, we can calculate the probability of conversion without Google:

… and without remarketing:

Calculations of the *removal effects* and *shares in the result* are the same as in the previous examples:

You can also create Markov chains of the 3^{rd}, 4^{th} and higher orders. Their nodes will have an even longer “memory” and the probability of transition between nodes will depend on the current state, as well as on two, three etc. previous states respectively.

If you will use higher-order Markov chains, you’ll notice that using higher and higher order at certain point does not significantly change the attribution. In practice, Markov chains of a higher order than 4 are rarely used.

Markov chains attribution calculations use the normalised removal effect. It causes some problems with single touchpoint paths, i.e. those with only one interaction.

Let’s see the example below. We have two paths here, one of which contains only one interaction, i.e. *Mail*:

The Markov chain graph will look as follows (probabilities add up to 100%):

Let’s calculate the conversion probability after removal of Facebook…

… Google …

… and Mail:

Calculation of the removal effect, share in the result and attributed conversions is shown below:

Note that in this attribution model, the *Mail* channel contributed to 0.67 conversions, which means that the third part of the conversion that occurred after interaction with *Mail* has been somehow attributed to Facebook and Google.

Since there was no other interaction on the conversion path containing *Mail*, why do we attribute any share in this conversion to other channels? It does not make sense.

It actually shows that the Markov chains attribution even as a theoretical model is a kind of approximation.

One of the solutions used is to remove single-channel paths from the Markov chains analysis. Since only one channel is involved in the conversion, the attribution is obvious (we attribute it to this one channel only) and there is no need for advanced modeling. Thus, we should only calculate Markov chains for conversion paths with two or more interactions, and then we should add conversions from single-channel paths to the attribution of multiple channel paths.

We have built a simple tool that allows you to calculate the Markov chains attribution. This tool has following options:

- inclusion of only converting paths OR both converting and non-converting paths
- Markov chains of the 1
^{st}, 2^{nd}, 3^{rd}and 4^{th }order - possibility of separate calculation of single-channel paths

The tool (beta) is available at tools.adequate.pl.

Markov chain Attribution is an alternative to attribution based on the Shapley value.

It is burdened with an embedded error resulting from the use of the *removal effect*. As result of normalisation, the attributed value of conversions is unfairly shifted from single-channel paths and other short paths to longer conversion paths.

However, this model has many useful features. First of all, this model is much less sensitive to random data with less statistical significance, which can completely distort, for example, the calculation of Shapley value. Thanks to this, Markov chain attribution can be used for smaller data sets and for larger channel granulation.

Despite certain complexity associated with graph loops, Markov chains require less computing power than Shapley value, which makes Markov chains attribution suitable for analysing higher numbers of touchpoints.

Markov chains often give results oscillating around the results of the linear model.

Markov chains are not so much sensitive in terms of detecting click spam as the Shapley value is (click spam are numerous interactions with actually no incremental effect on the total number of conversions).

You should also remember that regardless of the attribution method used, any algorithm that use conversion paths as input data is based on the analysis of correlations. Correlation, however, may indicate causation between interaction and conversion, but it is not proof of it, and therefore these models may sometimes misinterpret signals. For this reason, these algorithms do not detect conversion hijacking (like brand bidding, discount coupons etc.), as conversion lift methodology does.

Nevertheless, the comparison of Markov chain attribution to the linear model and other models can provide valuable signals for further analysis.

Artykuł Markov Chain Attribution Modeling [Complete Guide] pochodzi z serwisu Adequate.

]]>Artykuł AdWords Case Studies: FRU.PL and Sellektor pochodzi z serwisu Adequate.

]]>Change of attribution model allowed additional investment within the required ROAS goal. Portfolio biding strategy made it possible to deliver more conversions without changes in the total ROAS, despite the growth of the total AdWords spend by 64%. Automated optimization helped to react faster to changes in trends.

See: FRU.PL Case Study

Officially started in 2015, fashion aggregator Sellektor.com offers over 5000 brands in hundreds of categories of products. Today, in 2017, the website has 600k+ unique users per quarter and produces 100k+ redirects to partner e-commerce websites every month. After over 2 years of traffic and revenue growth, the company is still expanding with the best yet to come.

See: Sellektor Case Study

Artykuł AdWords Case Studies: FRU.PL and Sellektor pochodzi z serwisu Adequate.

]]>Artykuł Fixed ROI Profit-Driven Optimisation pochodzi z serwisu Adequate.

]]>The growth (G) of the total conversion value after the change is:

– where V is the value of click and Clk is the number of clicks.

Before the changes all parts of the campaign have the same ROI. Therefore:

After the changes, the average ROI remains unchanged i.e. the total value of clicks is equal to the total cost multiplied by (ROI+1):

At the maximum growth (G):

The discriminant of the quadratic equation is:

and the solution:

Changing CPC by ΔCPC will maximise the income without changing the total ROI:

The elasticity data are available in AdWords interface (Bid Simulator) or via AdWords API (Bid Landscapes). The “1” and “2” may be any part/segment of the campaign: keyword, adgroup, segment. For example, “1” can be “mobile” and “2” – “desktop”, or “1” can be a keyword and “2” – all the other keywords in the campaign (in this case we should calculate total elasticity of all the other keywords).

How to calculate total elasticity of two campaigns?

From definition:

If in all campaigns the relative change of CPC is the same:

The formula for total elasticity is:

This model assumes that price elasticity does not change within the range of changes in campaigns. However, in reality, price elasticity decreases as CPC increase. Therefore significant changes of bids may be inaccurate and sometimes may have no meaning (e.g. change of bid below minus 100%). In practice, if the formula implies very significant changes of bids, you should not modify bids more than by 20-25% and after this modification, measure the elasticity again and continue optimisation at the next iteration.

Artykuł Fixed ROI Profit-Driven Optimisation pochodzi z serwisu Adequate.

]]>Artykuł How to remove unwanted content from Google pochodzi z serwisu Adequate.

]]>Search Engines (in most countries – mainly Google) are one of the most important sources of information and reviews about businesses and products. Users are searching for:

- official information on the company website;
- online media publications about the company or product;
- reviews on blogs, discussion forums and social media portals.

Nowadays, what Internet users find on websites listed on the first page of Google search results, is one of the most important factors building product and company reputation online.

- Usually, there is a number of websites with large amount of content with negative views on the product and company. Sometimes it is random content, displayed in search results, as Google is not able to find other valuable content relevant to search query, but it can be also an intentional result of malicious activity.
- Publications being result of the PR activities not always appear on top positions in Google search results.
- Attempts to remove unwanted content by, for example, legal actions or offering material compensation or by other forms of influence, most likely will bring no results, among others because of existing copies. Such actions are risky and, if revealed to public, potentially may cause PR crisis.

People searching for reviews in the Internet usually focus on negative reviews. We acknowledge positive reviews, while we carefully read the negative ones. Even one or couple of negative reviews can have stronger impact than numerous positive reviews.

In the example below, there are many negative and neutral reviews on the first page of Google search results:

SEO PR activities promote positive reviews. Negative reviews are moved to lower positions and their impact decreases:

*The top 3 results are most important as they are clicked by 60% of visitors. The top 7 search results are responsible for 85% of clicks . The 10th search result link is clicked only once per 50 search result page impressions. Less than 6% users visit second page of search results.*

Ultimately, negative content disappears from the first page of Google search results.

Further SEO PR activities may remove negative reviews also from the second and following pages of Google search results.

Sponsored links help to immediately react to PR crisis situations. Using AdWords, Google advertising system, desired communication is promoted to people searching for relevant key phrases. This is recommended to promote official information placed on the product or company website.

**Example**

- In a TV show, an influential person says that our product, X-Example causes cancer. As result, number of searches in Google for the relevant product rapidly increases.
- Google AdWords campaign allows immediate reaction in this case.

Contact us for more information regarding this matter.

Artykuł How to remove unwanted content from Google pochodzi z serwisu Adequate.

]]>Artykuł 20% CTR in Google Display Campaigns pochodzi z serwisu Adequate.

]]>It is, however, also possible to achieve over 20% CTR in Google Display Network.

This campaign had only placement targeting. We have used domain names (usually parked domains) being misspellings or variants of the advertiser’s name, names of competitors and offered products.

This campaign had not only high CTR, but also good conversion rates, comparable with search engine campaigns. The creatives were mainly text units, with only small share of image ads.

Artykuł 20% CTR in Google Display Campaigns pochodzi z serwisu Adequate.

]]>Artykuł Facebook Case Studies (Poland) pochodzi z serwisu Adequate.

]]>Polish Stem Cell Bank (Polski Bank Komórek Macierzystych PBKM) is Poland’s largest cord blood bank. Stem cells, obtained from the umbilical cord blood at the birth of the child are cryopreserved, stored and may be used for future transplants. This Facebook promotion was a part of a sales and product awareness campaign targeting expecting parents (download PDF file).

FRU.PL is an Online Travel Agency, specialising in flights, one of leaders in Poland, with sister companies in Romania, Vietnam and Ukraine. The “Don’t clap when plane lands” campaign was a part of Facebook promotion for FRU.PL and affiliate website “Pan Lotek” (download PDF file).

Artykuł Facebook Case Studies (Poland) pochodzi z serwisu Adequate.

]]>Artykuł Real attribution of retargeting ads (ultimate A/B test) pochodzi z serwisu Adequate.

]]>The target audience is narrowed to small group of potential customers, and can be run on low-cost placements. For this reason, retargeting ads usually have very good post-click conversion ROI, much better than in any other type of display advertising.

Retargeting providers (including Google, who name it *remarketing*) use this argument to persuade advertisers to spend more money on retargeting ads, especially that for many advertisers it is the only form of display ads that brings positive ROI measured by post-click, last-click conversions.

Retargering providers very often claim that retargeting has positive impact on conversions even without clicks, and persuade advertisers to look at post-view conversions. However, this is a manipulation.

Usually, certain time elapses from the click to conversion (see the example below).

The retargeting ads are served to visitors who started the purchase process. Some users will click the retargeting ad and buy, but a significant part of them would buy this product without seeing the ads too. Many users who convert may not even view the ads served, but the system will still show the post-view, or precisely: post-impression conversion.

At the moment, the most popular web analytics tools are not ready to measure the actual attribution of the last-interaction and assisting clicks and impressions/views. The fact, that interaction (a click or impression) appeared in conversion funnel, does not prove that it had positive impact.

*Numerous studies show that during the conversion process, customers interact with advertisements served by competitors. However, it is hard to believe that these interactions could increase sale of our product, and most likely, competitor ads have negative impact on our conversions*.

Of course, we should rather not expect that the impact of our own ads can be negative, but what is the actual attribution to the conversion, is it 1%, 10%, 30% or more? Maybe the ads have negative ROI, because majority of ad spending goes to target visitors who would buy this product anyway?

For sure, retargeting does not have 100% attribution in conversion. This is because we have previously paid for each user who converted from retargeting: in SEM, display, social media, e-mailing. Even if the original visit was a direct visit, it was result of brand awareness investments. For this reason, the post-click conversion ROI of retargeting ads can’t be compared with, for example, search engine sponsored link for non-brand keywords.

In case of remarketing, we can, however, run an experiment that will show the exact ROI of retargeting campaigns.

We assign to each user who visited the website, on a random basis, a Google Analytics custom variable with a value of “A” or “B”. It is a user-level variable (not session-level or page-level), and re-visit does not overwrite the value.

Then we use Google Analytics remarketing list (custom variable “A” and “B”) and display the remarketing campaign only to the users from the “A” list.

In other words, the potentially retargeted users are divided in two statistically identical populations, “A” and “B”, of which only the population of “A” is subject to remarketing. Comparing conversions in both populations, we will be able to determine to what extent the remarketing has increased the number of conversions.

We measured visits and e-Commerce transactions (purchases). The experiment, run on an e-commerce website, was divided into stages:

1. **A = B** – Both “A” and “B” users see remarketing ads.

2. **Remarketing only in the “A” group ** – Remarketing in the “B” group has been turned off.

3. **A = B** – Remarketing for the “B” group has been re-enabled. This stage is split into two parts:

(3a) **interim stage (P)** – the first three days after “B” remarketing was re-enabled;

(3b) **A = B** – fourth day and subsequent days.

**Step 1 (A=B).** The purpose of the first step was to ascertain that the two populations of “A” and “B” are statistically identical. Number of visits and transactions in groups “A” and “B” turned out to be almost identical (the difference was much smaller than ± 1%), and other metrics, such as visit duration and pages per visit were almost the same.

** Step 2 (only remarketing “A”).** After turning off remarketing in the “B” group, we have obviously recorded decrease of visits in the “B” group, and the number of visits in the “A” group was higher by 5.2%. At the same time the number of transactions in the “A” group was higher by 18.5%.

**Step 3a (P).** Having restored remarketing in the “B” group, the number of visits in “A” and “B” became identical again, but during the first three days, the “A” group had still about 15% more transactions.

**Step 3b.** On the subsequent days, the number of transactions in “A” and “B” became similar again.

How is this possible? This experiment confirmed that the impact of online advertising is not only click-based. Advertising banners work, even without being clicked. Nobody clicks on the classic TV ad, nobody clicks outdoor or radio ads either, but these ads have impact on sales.

Users who saw remarketing ads made 18.5% more transactions. In the same time, in Google Analytics we could see that the remarketing ads caused only 1.7% of all conversions in the “A” group. So, the actual impact was 11x higher than we could see by observing last non-direct click conversion in Google Analytics.

*Interesting observation: during the experiment (Remarketing “A” only), Google Analytics has still reported significant traffic and conversions from remarketing source in the “B” group:*

*The reason is that Google Analytics assigns a traffic source to the last non-direct visit. Therefore if the user came to website from remarketing, and then, in the course of the experiment, he re-entered the website directly (direct visits), the source of this visit will be still shown as “remarketing” in standard reports. This, actually, shows that the difference between post-click and “real” impact is even higher than 11x.*

In this case, the positive impact of remarketing ads was much higher than the cannibalization. The cost of remarketing ads was much lower than the income from additional 18% transactions. This, however, is not a general conclusion. The result of this experiment can be completely different, depending on:

- product type and the conversion time lag;
- competitive environment – with low competition retargeting may have smaller impact;
- seasonality – the remarketing impact can be different in case of last-minute deals;
- retargeting strategy – Non-optimized retargeting can more likely cannibalise other campaigns.

For this reason, we encourage our clients to conduct this type of experiment individually. Please contact us if you need assistance.

You can read this article in Polish language (extended version).

Artykuł Real attribution of retargeting ads (ultimate A/B test) pochodzi z serwisu Adequate.

]]>Artykuł List of all countries for Google AdWords pochodzi z serwisu Adequate.

]]>For this reason you may need to add the list of all countries and territories manually.

You can do it fast using AdWords Editor and country ID codes. Currently, the country ID are numbers between 2004 (Afghanistan) and 2900 (Kosovo). However, the list of countries can change, so you should produce in Excel a list of IDs, e.g. between 2000 and 3500.

Using “Add many locations”, paste this list to AdWords Editor (giving information that the nubers are country IDs).

Then, press the button “Check locations”.

The system will verify the locations, add the names and mark the correct ones. You should delete non-recognised IDs now, and the campaign will include the list of all countries and territories.

Artykuł List of all countries for Google AdWords pochodzi z serwisu Adequate.

]]>Artykuł Reach and frequency report in Google AdWords pochodzi z serwisu Adequate.

]]>Reach and frequency report shows only display network impressions, clicks and conversions. Impressions, clicks and conversions from search network campaigns are not included. However, video campaigns impressions, clicks and conversions are included here.

The caption name *Unique users* is a little misleading in the case of frequency 8 or more. It does not mean that 4 193 839 users saw the ad 8 or more times. This figure represents the sum of the numbers of users who saw the ad 8, 9, 10… times. As the number of users who saw the ad *n* times includes users who saw the ad *n+1* times, this figure does not have any real meaning in terms of users. *Impressions* would be a better name here.

So, how to interpret the data? What conclusions come from reach and frequency report? What is the optimum frequency? How to combine frequency capping with CPC bids management?

Our article explains how to interpret the reach and frequency data in Google AdWords display campaigns in and how to use it in practice.

**The article is available to download here**

Artykuł Reach and frequency report in Google AdWords pochodzi z serwisu Adequate.

]]>