Posts By :

    Apurv Lungade

    Machine Learning; 7 Algorithms In Digital Advertising Platform [Part 2/2]

    1024 632 Apurv Lungade

    Click Here to read part 1 of this blog

    Few points which can help Ad-Network do more business with same setup [macro-optimization]

    Dear marketers and brands, please ignore whatever you read in this section 🙂 Helping brands reach & interact with their consumers is the most important thing; similarly, it is equally important for an ad-server to make money out of it & meet the ROI. There must be a sweet spot between these two goals – and when an ad-server system achieves this, it is a win-win situation for both – marketers/brands and the ad-server agency/network.

    5. Dynamic Floor Price

    80-20 rule says, 80 percent of revenue is generated from 20% of the clients (brands). In an ad-network ecosystem, the rule is a bit different – 90% of the revenue is generated from 10% of the SSPs. It is a game of demand and supply where ad-server system is the referee. When all demand players land the battlefield, they all want to hunt THAT user (most relevant & likely to interact with the brand) & ad unit. This is the opportunity for a referee to change the rules of the game and make it expensive to yield maximum.

    Yeah, it is very cool thing to have an any RTB based ad-serving system; but there are two BIG challenges over here:

    1. It is not merely the site content that makes brands attract and bid higher, but it is majorly the quality and relevancy of the user that plays the vital role. So, defining a set of discrete rules to raise or lower down the floor price won’t help here. Again, Machine learning algorithms which continuously track demands at user and ad unit level would only help tackle the problem.
    2. Second challenge here is – if your ad-server keeps raising floor price then at one point in time, no/very few bidders will bid for the ad slot and most of your inventory will get unsold. And once your system gets a “Expensive” label, it becomes difficult to retain and gain more demand. So, the system should be intelligent enough to know if demand is consistent and the floor price is just below the optimal point beyond which if it is increased, they are not going to be satisfied with.

    6. Know Your Inventory Treasure

    What if your ad-server is not RTB based and still want to take benefit of variable pricing? Well, there are ways.


    ML algorithms can keep a watch on inventory parameters like:

    • Site/Brand Popularity – based on trending/viral content being published
    • Monthly Traffic
    • Alexa Rank
    • Content Quality – based continuous sentiment analysis, reader engagements & sessions durations
    • Audience Quality – it is completely based on user’s response towards ads being served to it – it is measured in KPIs like ads visibility, user interaction with ads, clicks, leads and brand engagement
    • Inventory Type – it can be anything – a social media platform, SEM, websites, mobile app, push notifications etc. but the behaviour of each one of these is different – ML tracks the changes in the behaviour

    So, the ultimate mantra here is to make marketers spend more on inventory in demand.

    7. Platform Secrete Survey

    It is ad-servers responsibility to make brands happy with quality performance of ads, best picked inventory and reach the unreachable audience. Similarly, the usability and experience of platform plays an important role to make brands happy and helps retaining them for product lifetime. So, the goal here is to understand user’s behaviour on the platform and mould the platform accordingly.

    Make FAQs interactive as if a human is interacting with the user. This requires a very popular machine learning algorithm – Natural Language Processing. Again, a heavy piece of data required here to make system precise and accurate while answering the user’s questions.

    Another example would be to track user’s interactions through click events, time taken to complete a process – say setting up a campaign. Using this data, system can make inferences as in – which features are most favourite, which are very rarely being used, which processes are time consuming and which quick ones. All the inferences made by the system if it consolidates and conveys to the project/product manager, they can work on the pain areas and work towards the betterment of user experience.

    Machine Learning; 7 Algorithms In Digital Advertising Platform [Part 1/2]

    945 427 Apurv Lungade

    Machine learning as the term implies, is the process of making machine learn on its own and make the decisions the way human brain takes. The learning process includes collection of information, reasoning for conclusions and self-correction. These algorithms are not limited to any specific industry or nature of business or kind of a product/platform.

    Have you ever wondered, how Flipkart knows your choice and recommends a list of products which becomes very difficult for you not to buy them? How does Uber Eats know exactly when the delivery boy going to meet you and shows estimated delivery time?

    All these are epic examples of machine learning algorithms. Be it Google, Uber or Flipkart, the systems are trained in such a way that they analyse all data points and come up with the most relevant results/suggestions. The process is continuous evolving and producing day by day better results.

    Few pointers which can help brands reach the unreachable [micro-optimization]

    1. User’s Browsing Journey

    The most effective way to know a user’s choice is to think the way they think. Machine learning algorithms can trace the user’s everchanging choices and connects the dots to make a pattern. The continuous process makes user see the content and ad of his/her choice and thus increase in upselling a product or a brand. It may sound like traditional user categorization technique but the moral difference here is, categorization is a discrete method and it does not follow a pattern what ML algorithms do. Some platforms have taken a step forward and tried to build advance categorization by introducing scoring logic to each category user falls under. But again, bucketing a user in many categories makes the data skewed and this in turn leads in less accurate results than ML produces.

    2. Audience Cloning

    Continuing point 1 wherein machine learning algorithms continuously track user behaviour and makes a patter out of it; the process is applicable to all users in the network of an ecosystem. Once the system has a substantial amount of data, it can create samples of users having majorly same choices, interests & browsing patterns. These patterns change continuously with the change in audience counts and choices.

    The best example of this algorithm is – Netflix

    You have watched Sacred Games, Riverdale, 13 Reasons why, and suddenly you get a notification saying “Top Pick For You: Little Things”. Now, if you notice, all for web series are NOT of same a genre – so, it is definitely not picked by tracing your past taste. Netflix has got a huge user base and sampling those made it possible. Its ML algorithms continuously sample these users of same taste and try to suggest the unmatched shows across users in that sample – hoping that having same taste among these users, may also like the suggestion made by Netflix’s ML algorithm

    3. System Suggestions

    Have you ever seen a system talking to you? Yes probably – Google, Alexa, Siri etc. What if your advertising platform suggests you how to optimize your ads? Yes, it is possible with Machine Learning algorithms. System slice and dice the big data and correlates content’s meta data like – keywords, urls etc. That’s why, you type a single keyword and system suggests multiple around it.

    If you notice here, there can be thousands of keywords relevant to Virat Kohli but, system filters out the recent ones – This only possible when system learns the publisher content on continuous basis and this is the beauty of ML algorithms over traditional keyword suggestion techniques.

    Another example would be – system automatically crunches the inventory and user behaviour data for last T hours and suggests you change targeting accordingly.

    4. Analytics with suggestions

    Sometimes it is very difficult to define KPIs for you campaigns and taking decisions out of it. For video & rich media ads it can be views, engagements, sentiments and share of voice whereas for Native or Emailer ads, it can be CTR, eCPAs etc. What if you are using a comprehensive system which provides all possible stats of all possible ad formats and dimensions around these? It will be a mess! A straight away solution over this would be to have separate systems/analytics dashboards for separate ad formats – but, this will only help you analyse data separately and join the dots manually

    It is highly possible that, you reach your consumer through more than one channel & for that matter, having different systems to analyse those will never tell you the common user specific insights. Machine learning probabilistic algorithms can predict and identify common users from different channels and their responses towards your brand. Moreover, by using these insights, you can re-target your consumers through different channels; as a flip side of it, at some point in time consumers may experience it intrusive if they feel it irrelevant or disturbing. So, it is very important to continuously slice the user specific data and fine tune your campaign settings accordingly. ML algorithms can make it happen in a single dashboard with suggestions in it.

    So, in this episode we have seen how ML algorithms can help in betterment of user’s ad experience. We will see how ad-servers can make use of ML to get maximum of it to grow the business and make a responsive platform in the final part of this post. Click Here to access the final part


    Controlled approach the next big thing in campaign planning and optimization

    649 288 Apurv Lungade

    31st July, 05:30pm – “An amount of Rs. xxxxxx has been credited to your salary account…”  the most awaited SMS of the month! But what happens next? A lot of us face a similar challenge of maintaining a balance between our earnings and spends/expenses. The major hurdle is lack of planning. This scenario can be also observed in the case of digital advertising spends.

    55/5 rule of problem-solving suggests that often it is preferable to spend more time on identifying and properly framing the problem before trying to solve it. The proportion 55/5 comes from a quote attributed to Albert Einstein who supposedly said that if he had only one hour to save the world he would spend 55 minutes identifying and formulating the problem and only 5 minutes solving it. If we try to relate this to digital marketing industry, a media planner should spend more time in planning the campaign objectives rather than taking it LIVE in a hurry.

    At the time of campaign planning and optimization, media planner should try and make different categories of campaigns based on goals that he receives from the client.
    For example: Daily delivery goals: 10k to 1L impression, 1L to 5L impressions, 500 to 2000 clicks, 2000 to 5000 clicks; CTR goals: 0 to 0.1, 0.1 to 0.5. 0.5 to 1.0, 1 to 3 etc.
    Once you have these categories in place, try to fit the campaign in respective category.

    In digital advertising industry, media planners should try to follow 20-50-30 rule for efficient delivery of campaigns.
    What does it mean? It means the total spend should be divided into 3 phases – 20%, 50% and 30%.20-50-30First phase – Experiment
    The first phase “Experiment” is the most important aspect. The first 20% of the budget should be spent by tweaking the campaign attributes in such a way that it would meet the campaign goals. For example, if campaign goals are – daily delivery of 5L impressions and CTR min 0.5% then it is clear that the campaign need to be focused on large delivery and less on its performance. In order to achieve the goals by keeping maximum possible gross margin, one of the experiments I would perform on this campaign would be – target it on inexpensive site at high frequency.
    In addition to that, in this phase – try to target the campaigns on all verticals (sites and segments/audiences) which seem relevant to the campaign.

    Second phase – Blast
    Analyse results which you got in first phase and choose best combination of campaign attributes which are likely to give best results in terms of RIO and campaign goals. Take this combination and extrapolate it making a blast. 50% of the budget should be spent in this phase. Monitor the performance of the campaign and make sure it is giving expected ROI.
    How to calculate ROI? – in general terms, ROI = Revenue/expense but in digital advertising industry campaign performance is equally important. So here the ROI concept is – what is output of campaign performance (CTR/CR) when input is campaign different attributes. Example: There are two sites – S1 which is premium site and S2 is a normal site; when a campaign is targeted on these sites separately, and you notice that there is a minimal difference in CTR on these sites then ROI of serving campaign on S2 is more than serving it on S1.

    Third phase – Retarget
    Before moving to this phase, planner should plan in such a way that all the remaining backend goals are met in this stage. In order to make this happen, historical data acts as a treasure here. Say for example, you did and auto company campaign few months ago and you again receive a similar campaign but from a different client. Still the historical data of the previous campaign can be utilized as a learning material. This will help you in choosing the best remark audiences and also the ones that never performed earlier.
    Slice and dice the historical data and make inferences out of it which will be used for optimizing the current campaigns site wise – category wise – impressions, clicks, CTR
    Segment wise – counts

    Don’t spend the entire budget at once instead play safe by breaking your budget in these 3 phases and treat all the phase as a new campaign setting. Try not to mix earlier settings. Utilize the learnings from historical data wisely and effectively.

    The curious case of AD blocking

    The curious case of ad blocking

    599 400 Apurv Lungade

    Ad blocking to become a serious issue, should you be worried

    This year ad blocking created a lot of buzz in the ad tech and digital advertising industry after Apple rolled out its iOS9 update in September, enabling ad blocking on mobile devices. Advertisers and publishers around the world are concerned and are trying to figure out the solution for ad blocking. Though a lot of advertisers and publishers are still not clear what’s at stake, let’s go through some data to analyze the situation at hand.


    There are about 144 million active ad block users around the world, which is 5% of all Internet users.The number of people actively using ad blockers in India increased from 2 million (second quarter of 2014) to 4 million (198 million globally) in the three months ended June, only about 2% of the total number of Internet users, but this number has been steadily climbing.

    What is an ad blocker?

    Ad Block is a program (software/hardware) that prevents ads from being displayed to the user on a web page. In most cases, ad blockers are available as browser extensions.

    An ad blocker can block display ads, video ads, mobile ads. Native advertising is still untouched by ad blockers. This advantage can make native advertising more popular among advertisers.

    How does and ad blocker work?

    Ad blocker uses three simple mechanisms to block ads.

    • Blocking Requests – As explained above, it checks for the 3rd party domains and blocks all incoming/outgoing requests to serve ads. If there is any mismatch between domains, the ad blocker will not run the script.
    • Element Hiding – Hides all foreign elements which are embedded in an ad
    • Filters – User can explicitly add filter rules or URLs to block ads from a set of specific ad networks

    Ad block uses source URL of images, iframes, scripts and flash files to identify advertisements in the page. And then it blocks (HTTP and HTTPS requests) and hides (CSS display, visibility, and height) those advertisements. It identifies the domain names of ad networks and also uses built-in filters and keywords to identify advertising links.

    The curious case of ad blocking

    Linux users have a staggering 29.04% blocking rate, compared with 12.95% for Mac users, and 9.25% for Windows users. Mobile blocking is gaining popularity: Android shows 2.24% blocking rate and iOS 1.33%.

    How to disable an ad blocker?

    Option 1: Get the domain whitelisted

    An ad network /server can directly approach an ad blocking company to whitelist their domain. But, it is solely at the discretion of the ad blocking company to go ahead with the whitelisting or not based on some parameters.

    Eyeo (the company behind ad block plus) is dealing with the aftermath of an article by the Financial Times, which reported that “not only Google but also Microsoft, Amazon and advertising network Taboola are among the companies paying to stop having their ads blocked. In response, users have blasted the company and have vented frustration with the software on social media and elsewhere. According to the Financial Times, companies pay 30 percent of additional ad revenue that they would make from being unblocked“.


    1. No animations, no sound, no video – they call it non-intrusive
    2. User can explicitly restrict whitelisted ads too

    Option 2: Manipulate ad-blocker

    Step 1 – check if ad-blocker is present or not

    Step 2 –
    Secret Media developed a technology (patent pending), based on polymorphic encryption to pass ads through Ad Blockers, and to make sure that the ad Blockers are not able to identify these patterns code.

    Limitations –
    The technique explained above is generally used in viruses or malicious software to hide its presence in the machine or website. In short, the method is not legally approved.
    Advanced options of ads blocker plus are capable of detecting few of such encrypted scripts thus it blocks those too.


    1. Till now there is no concrete solution to manipulate ad blocker.
    2. Keeping business loss of ad-networks/publishers into consideration, ad blocker supports non-intrusive and standard ads
    3. If we focus on the Indian digital advertising market, ad blocking has not become very popular yet if we compare it to the US and UK market.

    As advertisers, you should try and diversify your advertising delivery channels and make sure you’re using all the ad targeting options available to you. Do A/B testing to check which channel delivers the highest ROI. Try your hands at some innovative high impact ad units and native advertising.  Moreover, social media advertisement is still unaffected by ad blockers. You can also give users the option of AdChoices. So Indian agencies/ad networks can breathe a sigh of relief as we still have time at hand.

    As publishers, you should always focus on user experience. Allowing non-intrusive ads and innovative ad formats will not hurt user experience. You should also keep checking CTR’s of these ads to check how much your ads are being liked by your website visitor. Conduct periodic surveys to get valuable feedback from your end user.

    Why cross device targeting is the new black

    Why cross device targeting is the new black

    599 400 Apurv Lungade

    Scenario – (as if it is a magic!)

    User was searching for a t-shirt on a shopping site on his desktop but didn’t buy it. Now he searches for more t-shirts on his mobile and gets to see the same t-shirt which he had searched on the desktop previously. An instant question that comes to your mind is – how does mobile get to know that the user on desktop and mobile is the same? Is it cookie or device identifier? But the cookie is present on the desktop. Is it something different which links these two devices with a single identifier? Was he logged in chrome/shopping site on both desktop and mobile? Yes, maybe! We don’t know…let’s explore!

    What was the scenario about?

    We all can see the shift in content consumption. More and more people are consuming content on the go i.e. on their mobiles or tablets. The above scenario considers the case of a typical cross-device targeting which many brands/advertisers are adopting in order to engage with their users/consumers hopping on multiple devices. Peechha naa chhodenge! (Wherever you go, our ads will follow you). The basic understanding about cross device targeting is “Identify a user or its behavior on more than one device and show him relevant ads accordingly”. It can be any two devices (see the image % share by eMarketer)

    Why cross device targeting is the new black

    A recent TNS study in Australia showed that for those owning at least three Internet devices, share of time was split as follows: PC/laptop (52 percent), smartphone (29 percent), and tablet (19 percent) (Source: TNS Mobile Life 2013).

    There are two approaches to cross-device targeting.

    Deterministic – This approach is executed using PII (Personal Identified information) of the user. This process is more accurate since the user is either logged in or he has shared his information with the publishers. The user data/behavior is tracked (stored) against this PII (email id, mobile no., physical address) instead of storing it in a cookie. The reason being cookies and virtual IDs do differ device to device but PII remains same. Now once a new user logs in, it checks if the PII is available – if it is, then stores the behavior info as well as based on previous data it shows ad; else creates a new entry in the database and starts studying the user’s behavior (fingerprinting). The above example is based on the deterministic approach.

    Benefits: Since user’s behavioral pattern is studied across devices in this model the ads served are more accurate and precise. Digital advertisers prefer this approach as here the user is already identified and segmented for personalized targeting.

    Challenges: In the deterministic model it is difficult to scale. We can say “Scale is inversely proportional to accuracy”. Privacy of the user is the biggest concern in this model. Another challenge is browser incapability when serving ads on different devices with different default browsers.

    Probabilistic – This approach involves using data points and attributes from different devices, operating systems, location data associated with bid requests, time of day and a host of other such details to predict statistical, aka likely, matches between devices. These probabilities take time to result in accurate predictions. The accuracy level in this model is 60 to 90%. Example 1- if a phone, a tablet, and a laptop connect to the same networks or Wi-Fi hotspots in the same places every weekday, it’s safe to summarize that all three devices belong to a specific user.

    Example 2- SilverPush did it for one of their clients. When a user is watching the brand’s ad, an ultrasonic sound (Beacons) is emitted from TV which triggers the user’s mobile to show the offer on it.

    Benefits: In this model user privacy is not at risk at all. Users can be paired across devices, operating systems, browsers which increase the scalability. Advertisers get to reach to their fragmented audiences in a better way.

    Challenges: This model is based on probability, and continuously recursive algorithms. Precise predictions take time to deliver desired results.

    Conclusion: The future lies in cross-device/screen targeting. Today’s marketers and advertisers should be ready for the big change. Both the approaches require efforts to overcome huge technical feats. They require huge infra, database (Big data), strong analytics tool (multidimensional reporting through slicing and dicing) and self-learning complex algorithms. Cross-device targeting plays a vital role in campaign performance which should be kept in mind before setting up a campaign.