Anand's marketing blog

I watched an interview of April Dunford, an expert on Positioning for B2B technology companies. Here are my key takeaways:

Good Positioning requires understanding who your customer is, which customer use cases the product is solving, which other products are typically considered by the customer, how the product is different from others, and how the product adds value.

One quick way to understand whether a company has good Positioning is to tag along with sales people as part of "first calls" i.e. the first time the sales person visits the customer. If during the first call the customer doesn't understand the use cases the product is solving or doesn't understand how the product is adding value versus the status quo, chances are that the product isn't positioned well.

One of the key things I've learned about Marketing is that you need to test things all the time: Positioning in B2B tech is no different. April suggests testing out Positioning options by giving a script based on the suggested Positioning to the best sales rep, and see if the First call->Opportunity conversion rates improves. A side-benefit of this is that not only do you get to learn whether the suggested Positioning works, you also get to use this learning to get everyone — Sales, Product, and the management team — agree that the new Positioning works.

April views Positioning as being closely linked to Sales — she says this is likely due to her background in B2B tech, where CEOs tend to care about sales much more than they do about things like logos, fonts and branding. I think even in B2C businesses, Positioning well done ought to increase sales, but it may be more difficult to isolate its effect, for as far as I know, while consumer brands test any proposed re-positioning with customers for perception and mind metrics, they don't typically go the distance by, for instance, changing the Positioning in 1-2 cities and checking the effect on sales. And the reason they don't do that is: they want the brand to stand for one thing everywhere at any given time. However, I may be wrong on this and will need to check for counter-examples.

One useful suggestion April gives is to treat the inputs (and even what seem like complaints) from sales reps as a way to learn about what's happening with the customer. For instance, are they seeing new competitors emerging? etc. The key is to treat sales reps as a source of potentially vauable market information, and to frame questions in a way that elicits this information.

The Qualified Sales Leader by John McMahon is an excellent book for anyone looking to learn B2B sales. I haven’t read that many books on sales – in fact, I can recall having read only The Ultimate Sales Machine by Chet Holmes (a lot of which is in fact about Marketing) and a book by Zig Ziglar (the title of which I can’t recall), both of which were quite good. McMahon’s book, too, is really good. 

McMahon’s approach to sales is to break it down into its component steps (Discovery, Scoping etc.), have operational definitions of when each step is complete, define with exactitude the terms used in the process (e.g. Champion, Coach, Economic Buyer), ensure every sales rep and sales manager use the same vocabulary, and coach rather than goad sales reps to deliver their quotas (as McMahon urges sales managers: Don’t be a glorified scorekeeper.). This sounds simple enough, but I understand many if not most sales teams violate some or all of these in practice. “Go slow to go fast” (my words, not his) is another principle he insists on (Don’t skip steps is how McMahon describes it). This echoes what Bent Flyjberg writes about project management in How Big Things Get Done, and the saying from the Navy SEALS “Slow is smooth, and smooth is fast.” Knowing the fundamentals and practising them consistently isn’t the norm in most endeavours. Additionally, with a clear process and well-defined qualification method (i.e. a way to know where you are in the sales process), forecasting sales becomes much more scientific for the rep, the manager, and the CEO and avoids quarterly angst and handwringing from all parties involved.

An important point he makes is that Pipeline matters. If you have a pipeline of leads flowing freely all the time, that takes care of a lot of things and, in fact, allows you to be inefficient at the sales management process and still be fine. The reason this is easier said than done is that pipeline generation is painful: it's not easy to keep cold-calling, cold-emailing, networking tirelessly week after week and dealing with rejections from your fellow-men and -women every week. I am reminded of Albert Gray's article on the common denominator of success, where he points out that successful people are successful precisely because they are able to do the painful things that many others don't.

McMahon has been the CRO at 5 publicly listed software companies, so his experience isn’t overly narrow or too context-specific. He has an engineer’s process orientation, a sales guy’s pragmatism and appreciation for the craft, and comes across as someone who truly cares for the welfare of his sales managers and reps: he cares, as he says, through creating competence, which in turn creates lots of good things for his teams. He understands that berating people to deliver results isn’t useful – I was in fact reminded of Dr. Deming when I read the book. McMahon calls for transformation of the sales approach and delivers it through leadership: of course it reminded me of Dr. Deming. 

I listened to an interesting podcast episode on Stratechery where Ben Thompson interviews Michael Morton of Moffett Nathanson. Some gems from the episode:

Prior to AI search (ChatGPT etc.), so-called "brand.com" websites (think d2c websites run on the likes of Shopify) saw 2x traffic growth from search (I am guessing on a YoY basis) compared to marketplaces (the likes of Amazon). After AI search, d2c websites have seen 4-8x traffic growth vs. marketplaces. This means websites will become increasingly important for d2c companies and will increase as a share of revenue relative to marketplaces. This also means planning for AI search becomes an important aspect of the Marketing mix for d2c companies.

Agentic commerce is, per Morton, unlikely to become a thing. By agentic commerce I mean something like: ask Claude to book a cab for you and it invokes the Uber app to do it for you. Morton's point is that the core functionality of the Uber app includes features like checking the ETA and seeing the cab arrive on the map, and there's no point in agentic flows if they need to replicate this functionality. On a somewhat related note, he makes the point that features like "subscribe and save" haven't really worked in e-commerce, with just 1-3% of Amazon's revenue coming from this feature. The reason is that it's difficult to know well in advance exactly when you need certain products at your doorstep. Morton makes the fascinating observation that even in the '60s when doorstep milk delivery was at its peak in the US, only ~30% of households would use the facility. I suppose it's different in other countries e.g. India -- wonder what accounts for these differences.

Morton points out that its all but impossible for a d2c merchant to vibe-code the equivalent of Shopify. Key reasons include the level of detailing and knowledge embedded in Shopify's flows; Shopify's ability to secure much better commercial terms with the likes of Stripe than any one merchant could on its own; Shopify's ability to for example work with Meta directly to implement CAPI; and the need to be able to troubleshoot problems on the shopping platform when they inevitably happen. Makes a lot of sense: in fact, makes me wonder if the same reasoning applies to SaaS companies as well.

Rather amusingly for me, Morton and Thompson were extoling same-day / next-day delivery as a driver of conversion rates on e-commerce services. Wonder if they know about India and quick commerce :)

I went through Alphabet’s earnings call transcript for Q1 CY ‘26 to understand how they are thinking of their ads business with AI. In general, AI helps ads platforms like Google and Meta in a number of different ways:

In Search, AI (LLMs in particular) helps ad platforms understand user intent better and also make sense of longer queries. This should help platforms serve more relevant ads. Here’s the relevant quote from the earnings call: And overall, just the understanding that we have with Gemini on intent has just significantly expanded our ability to deliver Ads on longer, more complex searches that were previously really difficult to monetize.

It helps in targeting customers better. Even before the release of ChatGPT and the LLM era, marketers were better off specifying targeting broadly rather than narrowly, and this only keeps getting truer with more data and better AI systems.

It helps in improving and making creatives. It can help generate creative variants, touch up images and videos, add subtitles to videos and so on. See here (Meta’s Advantage+) and here (for some of Google’s AI offerings for creatives). We can expect more features to get added with time.

Another general rule of the ad platforms is: get users and usage first; and then try out various ad formats and see which ones work, then double down on the ones that do. So, expect that they’ll try a lot of formats now, and then refine them over time. For instance, here’s what they said about ads in Gemini in the earnings call:

It's fair to say that we really believe the format that works well in AI Mode would transfer successfully to Gemini app. And so, today in the Gemini app, we're focused on the free tier and subscriptions and our AI plans were a sizable contributor to our Google One revenue growth. But let's also be clear, Ads have always been a big part of scaling products to reach billions of people, and if done well, Ads can be really valuable and really helpful commercial information.

As it does every year, the Pitch-Madison AdEx Report 2026 gives a detailed analysis of trends in India ad spend. Here are my key takeaways:

They’ve revised the AdEx computation (Rs. 1.5 lakh cr., or ~US$ 15.8bn) to include two additional components viz. MSME digital spends (Rs. 36k cr.), and qcom retail media (i.e. advertiser spends on quick commerce platforms - Rs. 4000 cr.) According to this definition, AdEx was +12% YoY in CY 2025.

AdEx roughly breaks down into:

Traditional media viz. TV, OOH, Radio, Print – decreasing at 1% YoY. Within Traditional, OOH is the only medium that’s increasing, at 4% YoY. This bears out what I’ve seen anecdotally myself; an interesting thing about OOH is that companies now use it in concert with social media, so it’s not just about OOH-specific reach but also about “amplifi-ability” on social.

Digital media – increasing at 18% YoY. No surprises here. Digital is now 60% of AdEx per the new definition.

Qcom retail media is ~Rs. 4000 cr., and growing at 202% YoY. A few points here:

Qcom itself is growing at ~100% YoY (based on commentary from the likes of Blinkit, Instamart etc.) 

Most Qcom platforms are ramping up their advertising offering as they try to drive profitability.

[my estimates from market data etc.] Qcom is ~US$ 10bn/yr in revenue, so Rs. 4000 cr. of AdEx here means ~US$ 421mn / 10bn = 4.2% of revenue, which makes sense to me.

TV AdEx is ~Rs. 33,000 cr., with ad volumes at -10% YoY but with increasing price/ad, AdEx at -5% YoY. Spend is shifting towards sports (~23-25% of TV AdEx vs. 18-20% in 2024) and premium GEC (25-27% – roughly stable share YoY) at the expense of commoditised GECs etc. This trend of premiumisation is true for regional channels too.

Setting budgets is a key task of Marketing. Too low or too high a budget, and profits may be sub-optimal. What are some ways to set budgets?

Image: generated via ChatGPT

% of expected revenue - in this, you just start with estimated revenue for the upcoming year, and set Marketing budgets as a % of that. The % could be +/- from the % in previous years. Simplistic as it sounds, I’d wager this approach is quite common in the industry, if not as the sole method, at least as one of the methods to arrive at budgets.

Set share of voice >= expected share of market - this approach requires us to (a) set target share of market for the upcoming year (b) estimate marketing spends (proxy for share of voice) for competition. (a) and (b) allow us to set budgets based on share of voice. See here for some background on this.

Bottom-up budgeting based on required revenue - in this method, we break down required revenue into its components, and derive budgets for each component. Taking the example of an e-commerce app:Required revenue = (New visitor sessions x new visitor session conversion rates x Average order value - New visitors) + (Old visitor sessions x old visitor session conversion rates x Average order value - Old visitors).

New visitor sessions = Paid traffic + Organic; Estimate spends for each.

Similarly for old visitors. 

In case we require budgets for conversion rate improvements (e.g. voucher spends, pricing), we could bake those in too.

We could break this down one more level by, for example, splitting old visitors into RFM (Reach, Frequency, Monetary) classes and estimating budgets for each RFM class. 

Sum-of-the-parts, which I haven’t seen used in practice:

For each marketing channel or in fact for each element of the marketing mix, estimate how much you can increase it such that marginal profit = 0.

Sum up the line items from (a) to arrive at the marketing budget.

Truth be told, elegant as this method sounds, it’s difficult to make such estimates at a line item level. Also, there may be interaction effects between channels (e.g. the marginal profit on performance marketing is a function of the brand spend), so I’m not sure if this method is useful in practice.

Adaptive control system approach, as proposed by John DC Little of MIT, where the idea is as follows:

a. Create a mathematical model of the expected profits at different levels of sales promotion.  b. This model has parameters which one needs to determine using an experimental setup wherein you spend different amounts in different markets and observe the sales response. The basic ideas were clear enough from the paper, but I’m afraid I didn’t have enough of an understanding of some of the advanced statistical concepts in it. I think though that something like this could be encoded, with the help of statisticians and programmers, into a tool that regular marketers could use. That said, experience suggests that constructing these experiments is easier said than done, for it involves adjusting for natural variations in the test markets (e.g. differences in operational metrics in each market; differences in consumer behaviour; etc.). If you’re interested, the paper is called A Model of Adaptive Control of Promotional Spending, by John DC Little.

I was also reading some research papers on the subject of marketing budgets, and the following paper had some interesting ideas: How Corporate Cultures Drive Advertising and Promotion Budgets: Best Practices Combine Heuristics and Algorithmic Tools, by West, Ford and Farris. The following is a NotebookLM summary of this paper:

"Marketing budgeting is rarely a purely rational process, frequently combining "System 1" heuristics (simple rules of thumb) with "System 2" algorithmic tools (logic and probability) to manage risk and improve performance. Heuristics serve as a vital "check and balance" for complex analytical models and act as a "tie-breaker" when multiple models project similar profit levels.

Organizational culture significantly influences budgeting choices, with adhocracy cultures—those emphasizing innovation and entrepreneurship—being the most inclined toward algorithmic methods. While other cultures like market, clan, or hierarchy traditionally favor heuristics, the sources indicate that the most reliable results come from triangulating two or more methods.

Knowledge and experience within the marketing department are strong predictors of a firm's shift toward sophisticated budgeting, as greater marketing involvement is positively associated with the use of algorithmic mechanisms. Rather than replacing heuristics, experienced marketers use these logic-based tools to augment and verify traditional rules of thumb, especially when decisions must be made in fast-changing environments."

This link provides a good summary by one of the authors of the above paper.

In summary, what we can say is:

A mix of multiple algorithms and heuristics will probably be the “best” way to determine marketing budgets. In other words, triangulation is the way to go.

Budget estimation is as much a matter of risk-taking propensity (of the manager and the organization) as it is a matter of mathematical formalism.

Any budget estimates are best treated as tentative and works-in-progress, requiring continual adaptation to market conditions: course correction is just as important as setting the budget.

A few years back I interviewed with a tech company. The interviewer asked me the following question: A company is about to open a fashion store on the high street; how would you go about marketing this? In trying to answer the question, I realized I didn't have a usable framework, a general framework for all Marketing problems. Somehow I've never found the "4P's" (Product, Price, Place, Promotion) framework all that helpful; it's a mnemonic device rather than a forcing function to clarify thinking.

What does a good framework look like? Hamilton Helmer's "7 Powers" framework for Strategy comes to mind. To paraphrase Helmer, a good framework must be Simple without being Simplistic. In other words, it must be practical without sacrificing too much rigour. With this background, let's look at a few general frameworks for marketing.

General Framework 1: The 6-point check-list

What's your product (or service)? What's your core offering, quality level, price relative to competition?

Who are your customers? What are the different types (segments) of customers?

What does your P&L look like? And what does that mean for your quality level, pricing, distribution, advertising?

Where would you like to be in terms of market position and profitability?

What are your distribution channels?

How do you ensure customers know about the value you are offering? (positioning, messaging, media planning)

Work iteratively through steps (1) to (6).

General Framework 2: General Diagram of Marketing

This is represented by the following diagram. I'll explain this in a future essay.

General framework 3: It's all about value

The idea here is that companies exist to create value. The questions then translate to:

Is the company creating value? For which segments of customers?

Is it creating enough value to sustain the business?

How can the company keep creating value as customer expectations and technologies change?

Speaking of software engineering in his essay "No Silver Bullet" in 1987, Fred Brooks contrasted difficulties that are inherent ("essential") and accidental to its practice. Accidental difficulties may be removed by improving extant abstractions and programming languages; the essential ones, however, will persist and will not yield to a "silver bullet". The solution has two fronts: sure, don't stop improving programming languages. But know that your true adversary will not yield to quick fixes but rather to discipline and incremental improvements over decades. Thus it is with Marketing.

The Essential difficulty of Marketing lies in our inability to know exactly what people want: Mel Gibson's movie "What Women Want" (and its mirror image 'What Men Want") has an interesting premise precisely because it's so preposterous in real life. Look at it this way: we often don't ourselves know what we want; good luck trying to read other people's minds. Furthermore, people's tastes change: the Ambassador car, so popular in the 1980s would hardly find takers today. This is the result of both cultural trends and competition's tendency to not sit still. Lest one thinks data and rigorous forethought can help save the day, I would point you to the recent book "Working Backwards" by Colin Bryar and Bill Carr. Describing Amazon's approach to greenlighting project proposals, they concede that in many cases it was impossible to predict customer behaviour in advance. Now, if forethought worked anyplace, I'd bet it would work at Amazon.

All this would mean there's no single marketing mix that would work for all companies, all of the time. Indeed, Prof. Neil Borden of Harvard, who invented (more precisely, made popular) the term "Marketing mix" wrote in the 1960's of trying to discern underlying commonalities in the marketing mix of various companies. Finding none to speak of, he could only counsel patience and a methodical approach towards improving the practice of marketing. The very term "Marketing mix" was meant to evoke a masterchef's art rather than a changeless recipe, and Prof. Borden's metaphor has worn well to this day.

That, however, is different from saying that all of Marketing is beyond redemption. Indeed, there are several areas, falling in the domain of "accidental" difficulties, which are well served by a combination of existing research and what savvy marketers have long known. For instance, new customer acquisition, rather than improved retention or cross-sell, drives revenue growth. A combination of sustained reach and distinctive brand assets creates physical and mental availability, in turn enabling growth (Byron Sharp, How Brands Grow and other publications from the Ehrenberg-Bass Institute). Delivering Reach at low cost helps in acquiring said new customers (Sudhir Sitapati of Hindustan Unilever, in The CEO Factory). It's easier to reinforce what's on customers's minds rather than change their minds (Stephen King, JWT planning guide). Branding efforts drive long-term growth while activation efforts drive near-term growth; both are essential (Binet and Field, The Long and Short of It). More fundamental still, we know customers respond to greater convenience, lower prices, lesser friction and (generally speaking) more variety.

Likewise, in the digital era, the tenets of growth hacking are ignored at one's own peril. For instance, Facebook's success at growth marketing tells us, above all, the importance of creating experiments and fast feedback loops to figure out what customers will respond to best. The digital era also brings with it new vectors for growth: for instance, DropBox's referral programme, or the network effects wrought by social networks, or unusual content->traffic->transaction pathways such as the ones created by Zillow or TripAdvisor.

People say high bounce rates are bad. At first glance, this makes sense. After all, who’d want customers to visit a (web or app) page and leave without seeing any more pages? We might decide, therefore, that digital advertising may be made more efficient by eliminating channels with high bounce rates.

This is a mistake. Let me explain why.

Consider the following:

1. Salesperson A stands just outside a retail store handing out pamphlets. 90% of the people who get pamphlets visit the store. Of the people who visit the store, all of them end up shopping at the store. 10% Bounce rate, 100% Conversion rate (=customers/visitors).

2. Salesperson B expends shoe-leather, visits customers’s homes and hands out pamphlets, even speaks to a few customers. 10% of the people who’ve seen the pamphlets end up visiting the store; of these, 50% end up shopping. 90% of people “bounce” i.e. they see the pamphlet but don’t visit the store. 90% Bounce rate, 50% Conversion rate.

3. Salesperson A and B have the same fixed salary.

Would you, on the basis of the above data, conclude that Salesperson A is the better salesperson? Or, that Salesperson A is attracting “higher-quality traffic” than Salesperson B?

Consider the following question instead:

Suppose Salespersons A and B hadn’t handed out pamphlets. What percentage of people would have become customers anyway?

The conclusion is the very opposite of what we’d get if we went by bounce rates or conversion rates. 

Granted, it’s a lot easier to measure things based on last-click based attribution, and a lot harder to measure Incrementality. This would be alright if a sole emphasis on last-click attribution and bounce rates resulted merely in a suboptimal solution. The trouble is, it can lead you in exactly the wrong direction.

Notes

1. The title of this essay is inspired by the computer scientist Edsger Dijkstra’s essay “Go to statement considered Harmful”, see http://www.u.arizona.edu/~rubinson/copyright_violations/Go_To_Considered_Harmful.html

2. I’ve taken some liberties in extending the concept of bounce rates to the offline world. Specifically, bounce rate is defined as the % of total sessions that begin and end at the landing page. In the Salesperson analogy, I’m considering a “view” of the pamphlet to be a visit to the landing page, and a visit to the retail store as the negation of bounce. Imperfect though the analogy is, I think it captures the essence of the problem well enough.

3. For an example of how an emphasis on different metrics, each seemingly sensible, leads to entirely opposite conclusions, see Jeff Bezos’s 2004 letter to shareholders, where he compares the effect of optimizing for free cash flows versus net profits. https://www.sec.gov/Archives/edgar/data/1018724/000119312505070440/dex991.htm. 

In this letter, Bezos says: “Though some may find it counterintuitive, a company can actually impair shareholder value in certain circumstances by growing earnings.” 

Analogously, marketing teams can actually impair their effectiveness in certain circumstances by trying to reduce bounce rates. 

There’s another great quote from the same letter: “Cash flow statements often don’t receive as much attention as they deserve. Discerning investors don’t stop with the income statement.” 

Of marketing, I’d say the following: Incrementality doesn’t receive as much attention as it deserves. Discerning marketers don’t stop at bounce rates and conversion rates, they also consider incrementality.

A few weeks back I attended a 2-day virtual workshop on Customer Centricity, conducted by Prof. Peter Fader and Sarah Toms from The Wharton School of the University of Pennsylvania. As Prof. Fader explains in this video film, Customer Centricity means recognising high-value customers and treating them differently. Needless to say, the belief is that doing so creates more value from marketing spend. Two questions arise. Is there basis for this belief? If yes, how do you revise your approach to marketing?

The first question, I think, is best approached by having an open mind that it might be true, and testing its applicability to your particular marketing situation. Now for the second question.

If you believe in Customer Centricity, there are a couple of things you’ll need to do. First, you’ll need to get a granular understanding of your customers. This will require a combination of tools (e.g. CRM software) and people (marketing or business analysts) to collect data, define segments and, within segments, identify high-value customers. 

Second, you’ll need a framework to organise marketing interventions. One such framework is a 2x2 matrix with Targeted Customers (High Value/Low Value) on one axis and Tactical Approach (Offensive/Defensive) on the other. To illustrate:

a. High-value/Defensive: customer relationship managers for high-value accounts

b. Low-value/Offensive: Loyalty programme

c. High-value/Offensive: Premium tier

Having got the definitions out of our way in Part 1, let’s now look at the policy changes introduced by Apple in June 2020.

1. In order to obtain IDFA, apps will now require an explicit opt-in from a user (earlier it was default opt-in). Apps will have to ask for this permission via something called the AppTrackingTransparency framework.

2. Attribution (i.e. finding out which ad network or campaign resulted in app installs) will be enabled by something called SKAdNetwork.

Point 2 is a topic for another day. Let’s consider point 1 and its impact on the various entities in the industry.

1. Ad networks: It’s worth analyzing this in 3 buckets: Google; Facebook; All Others.

a. Google: ~60% of Google’s revenues are from search (details here, pages 28 and 29), where users are quite directly expressing their intent. Therefore, I expect Google to be only marginally affected.

b. Facebook: Absent IDFA-based re-targeting signals, I’d expect Facebook’s ads to iOS users to be less relevant, and therefore worth lower CPMs. However, Facebook also has access to a lot more high-quality signals than other ad networks. To cite an example, with the IDFA rule change it’s true that Facebook will not know a user visited Patek Philippe’s website 3 times in the last week, but it will still know that the customer is interested in international vacations, The New York Times, gadgets, eating out, and sourdough bread. In other words, Facebook is most likely at a big advantage (as is Google) when you look at relative degradation in efficacy. 

c. All Others: From (b) above, it follows that all other ad networks will see their efficacy relative to Google and Facebook decrease. This means they will lose market share to Google and Facebook.

2. Advertisers: The more sophisticated advertisers pay for incremental conversions. This means that in the short run they’ll simply reduce the marginal spends that don’t deliver conversions at their acceptable cost/conversion. Indeed, this is what puts downward pressure on CPMs. However, in the long run, due to the worse relative degradation of the “All Others” category, they will become even more reliant on Google and Facebook.

3. Publishers: Lower CPMs will affect publishers the most, in my view. On the web, when third-party cookies were disabled, publisher earnings for the top 500 publishers decreased by 52% on average, per a study published by Google in Aug 2019. It’s not hard to imagine a similar effect from the IDFA change.

4. Users: Users will see less relevant ads.

News

Changes are afoot in the digital advertising universe. In June 2020, Apple announced changes to its IDFA policy. In January, Google announced “a path to make third-party cookies obsolete”. In this essay, I will:

Define Re-targeting.

Explain how customers are tracked for Re-targeting. This involves demystifying some terms namely IDFA, GAID, cookies and third-party cookies.

Walk you through the policy changes from Apple and Google.

Predict the implications for the 4 entities affected by these changes, namely Advertisers, Advertising platforms, Publishers and Users.

I’ll cover points 1 and 2 in this essay, and parts 3 and 4 in another essay.

What is Re-targeting?

Re-targeting means:

Using information gathered from a user’s activity on a website or app, to show her more relevant messages. These relevant messages may be shown via the website or app, push notifications, emails or external advertising channels.

As an example, say you’ve seen the product detail page of The Grapes of Wrath on an e-commerce website,  but haven’t bought it yet. Through Re-targeting, the  e-commerce website could show you an ad for The Grapes of Wrath while you’re scrolling through your Facebook feed. Through Re-targeting again, they could send you an email with a discount on the book.

What’s required for re-targeting?

For Re-targeting, an identifier is required that tracks customers across platforms. The identifier depends on the re-targeting scenario:

In the above chart, let’s start from the lower-left corner and go clockwise.

1. Logged-in and Internal channels: We have the email ID or phone number of the customer. That’s all we need to reach the customer via PNs or emails or the app.

2. Logged-in and External channels: We have the email ID or phone number of the customer. Both Google and Facebook permit the matching of your internal customer list with audiences (Google’s Customer Match and FB’s Custom Audience). This matching is based on the fact that most people (~70%, in my experience) use the same email ID when they log into a website and when they use Facebook or Google.

3. Non-logged in and External channels: Here we must rely on cookies, IDFA or GAID.

4. Non-logged in and Internal channels: We have cookies and the device ID (or push tokens, in the case of Push Notifications) to work with. These may be used for personalization on the website or app or via Push Notifications (but not emails).

Let’s now demystify cookies, IDFA and GAID:

1. Cookies: Here’s how Brian Kernighan describes cookies in his book “D is for Digital”. He starts with the observation that the web’s HTTP protocol is “stateless”, in other words “an HTTP server is not required to remember any information from one request by a client to the next.” He then raises the question of how the server might be able to “remember” things that it’s required to remember, e.g. a user’s login credentials, which brings us to cookies.

“[A cookie is] a small piece of information passed between programs. When a server sends a web page to a browser, it can include additional chunks of text (up to about 4,000 bytes each) that the browser can store; each chunk is called a cookie. When a browser makes a subsequent visit to the same server, it sends the cookies back. So in effect, the server uses memory on the client to remember something about the client’s previous visit.”

In addition to remembering your login credentials, cookies are also used for advertising, as follows:

a. When you visit a website “A”, the server sends cookies from advertising platforms (e.g. Google, Facebook) to be stored on your device.

b. After this, when you visit a website “B”, the browser sends to the server not only the cookie information associated with website “B” (the first-party cookie) but also the information from the visit to website “A” for use by Google and Facebook (third-party cookies). The way this works is that when you request a URL via your browser, the server sends JavaScript code that asks for and retrieves information from all these cookies.

Note the distinction between first-party and third-party cookies. Note also how the browser needs to cooperate with the server to enable the functioning of cookies. We’ll revisit these points in a later section.

2. IDFA: IDFA stands for (Apple’s) Identifier for Advertising. It’s best explained through an illustration. 

a. Suppose you browse an e-commerce app “A”, which has the Facebook SDK. The Facebook SDK captures key events in the app, e.g. add-to-basket, purchase, and ties it to you via the IDFA.

b. When you go to the Facebook app, Facebook now knows the events which you’ve triggered in app “A”, since your IDFA stays the same across apps on the same device.

Facebook can now show you relevant ads based on your event history.

A couple of things to note here: 

a. A user has the same IDFA across apps.

b. The user has an option to reset the IDFA at any time: however, my guess is that this option is rarely used (the default is for IDFA to sustain). This allows Facebook or Google to use IDFA as a reliable identifier.

Advertisers do 2 things:

1. They reach people who are amenable to persuasion, in a cost-efficient way. This is the discipline of Media planning.

2. They craft a persuasive message. This is the discipline of Communication.

For Media planning, advertisers in India use a tool called the NCCS grid. Indeed, throwaway references to “NCCS AB segments” and the like are a calling card of the seasoned and wannabe advertiser alike. I’ll spare you its gridly alphanumeric details here. We’ll talk about a much more interesting story: the story of its genesis.

Let’s set the scene. Here’s an advertiser working in the e-commerce industry, who wants to advertise on TV:

1. She identifies the target segment as follows: all people in India who have shopped online at least once in the last 12 months.

2. The advertiser tells her TV planner: “Show my ad to all people in India who have shopped online at least once in the last 12 months.”

3. The TV planner now shrugs and says: “I don’t know how to identify this target segment.” 

Nor would a TV planner know how to identify “people who are likely to buy high-end phones” or “businessmen looking for reliable supplies of some widget.” 

Rather, a TV planner presents a menu of standard segments for the advertiser to pick. This is the NCCS grid. 

Using this, the advertiser maps her target group (“all people in India who have shopped online in the last 12 months”) to the best approximation on the NCCS grid (say, NCCS A, B). And that’s the input she gives to her media planner.

In the digital world, the targeting options are more direct. You can simply specify, for instance:

a. People who are likely to be “in-market” for travel;

b. Algorithm-generated Lookalike audiences which are similar to your existing customers;

c. People who are likely to have “affinity” for e-commerce purchases;

d. You can even leave the targeting almost entirely unspecified, for machine learning algorithms to play by ear and figure things out.

Returning to the world of TV: the question is, what’s the best way to arrive at these standard segments? A little digression: a couple of weeks back I was listening to a podcast where the guest, Shishir Mehrotra, described the concept of an “eigenquestion”:

“For a simplistic definition, the eigenquestion is the question where, if answered, it likely answers the subsequent questions as well.”

With this framework in mind, what’s the best way to arrive at these standard segments? Turns out, there are a couple of eigenquestions:

1. What method will lead to as much Discrimination between segments as possible? 

2. What method will lead to as much Homogeneity within segments as possible?

(on reflection: these are the questions you’d ask to arrive at any type of segmentation).

This isn’t rocket science. It simply expresses the intuition that different segments ought to be different from each other and similar within themselves.

Now, India had something called Socioeconomic Classification (SEC) for several decades to solve exactly this. However, times change, requirements change, and there was an attempt to figure out a better way to do this.

In the early 2010s, Market Research Society of India (MRSI) and the Media Research Users Council (MRUC) came up with the New Consumer Classification System (NCCS), which the Indian Readership Survey (IRS; a survey that tracks newspaper readership and product consumption) adopted in 2014 and the Broadcast Audience Research Council (BARC; an organization that tracks TV viewership) in 2015.

NCCS asks its respondents the following questions:

1. What’s the education of the Chief Wage Earner?

2. What’s the number of items used in the household, which in my mind fall into the following categories:

a. Roti: Agricultural land (owned), LPG stove, refrigerator

b. Kapda: washing machine

c. Makaan: electricity connection, ceiling fan, AC

d. Travel: 2-wheeler, 4-wheeler

e. Technology & Entertainment: Colour TV, Laptop/computer

Based on this you can arrive at the NCCS grid, available at the link below:

https://www.afaqs.com/news/media/42980_a-dummys-guide-to-nccs

If you own 9 items or more in the above list, and your educational qualifications are “graduate and above” you fall into the NCCS “A1” classification, which forms <1% of the population of India. At the other end is NCCS “E3”, which is the segment that owns none of the durables in the list and is illiterate. 

That, in short, is the story of the NCCS grid.

Notes

1. https://mruc.net/uploads/posts/7709d00f2bfd8ec61574c05948641b7d.pdf

2. https://coda.io/@shishir/eigenquestions-the-art-of-framing-problems/eigenquestions-3

“When you get to be an expert at something, you’re hoping to achieve a deeper understanding, not just have a large set of recipes to go execute.” - Jim Keller, microprocessor engineer, on the podcast “Artificial Intelligence with Lex Fridman”

If you’re managing an online product (app or website), it’s an oddity if you aren’t running experiments. The beacons shine bright: Amazon’s Add-to-cart button is a comically well-refined marvel of experimentation. Facebook isn’t far behind either.  

Experiments are useful. But like many things in life, too much of a good thing isn’t always a good thing. Part of the problem is that experiments are used as a way not to validate ideas, but as a substitute for philosophy, taste, judgment, heart. Sure, an experiment can tell you if a blue variant of a button results in higher conversion rates than a green one, but it can’t tell you which one makes the product in its entirety more pleasing to the customer. When you take an experiments-based approach to its extreme, you’re making not a cohesive product but a loose archipelago of features, mumbling to the user the truth that it’s been designed by no one at all. 

Experiments have a way to ensnare you into the trivial, unless you’re vehemently vigilant. The story of Ulysses and the sirens comes to mind. I was watching this talk by Alan Kay, a pioneer in the computing world, where he says, “Better and Perfect are the enemies of what’s actually needed.” Sure, maybe a blue button works better than green, but is that what’s actually needed? 

The problems don’t end there. Even when you run an experiment for the right reasons, the waters are fraught. There are three parts to an experiment. Sampling, Running the experiment, and (if the experiment is successful), Expanding it to the wider world. These are three swirling rapids, each to be regarded with trepidation.

1. Sampling: Condensing the view of Walter Shewhart, a statistician, Dr. Deming said: “There [is] no such thing as a random sample. There is, however, such a thing as a sample selected by a random operation." It follows, for instance, that you need to pay the utmost attention to the operation by which you select buckets A and B for an A/B test. 

Say you choose “customers with a yahoo email ID” as bucket A, and “customers with a gmail ID” as bucket B. It’s hard to imagine these buckets being identical, because customers with a yahoo email ID are likely to be early entrants into the Internet world: and quite likely more affluent on average than the group with a gmail ID. One of the ways to check this is by comparing the behaviour of customers in these two buckets even before you start the experiment.

A few weeks back, I listened to this podcast episode on how mice used in lab testing are different (they have longer telomeres) from wild mice, resulting in potentially flawed results when drugs are tested on them. 

In a nutshell: sampling isn’t simple.

2. Running the experiment: Unlike in a lab, it’s hard to maintain “everything else constant” in the context of a company. For example, let’s say you’re trying to figure out if advertising is effective or not. One way you could do this is: pick two cities that have showed similar revenue behaviour in the past, turn on advertising in one city (say, Pune) but not in the other (Chennai) and see the effects on revenue. However, the competitive ground beneath Pune may well be shifting differently compared to Chennai.  It’s hard even to ensure that other variables e.g. pricing, distribution, stay constant during the experiment. To avoid these pitfalls, you have to be diligent and view results with the necessary dose of skepticism.

3. Expanding it to the wider world: Even if you have a successful result from an experiment, there’s no guarantee it will work when you expand it to the wider world. Think of the following scenarios:

a. Suppose you run an experimental training programme in a retail store. It works well. Now you have to expand this to a hundred stores across. Are you sure you’ll be able to achieve the same level of training in all these stores as you did in the experiment? If the training isn’t as effective as in the experiment, are you sure it won’t worsen the experience for customers (for example, because the training confused the sales associates in the store)? 

b. Suppose you make a tweak on your app, which improves page load time in an experimental study. Are you sure this performance improvement will hold up when you expand it to a wider audience? For instance, it’s possible that the tweak you made requires a new API call, which works well under experimental load but breaks down under high load.

To be clear, it’s worth running experiments. But it’s also worth being careful what experiments you pick and how you run them and use the results.

References

1. https://web.ma.utexas.edu/users/mks/statmistakes/randomsample.html