GOOGLE SEARCH CONSOLE FOR MARKETERS 28

{PODCAST} In-Ear Insights: Why Consumer Recommendation Engines Fail

In this episode of In-Ear Insights, Chris and special guest John Wall discuss the state of consumer recommendation engines. Why are recommendations so narrow and ineffective many times? What could we do to improve them beyond what we get now? Listen in as we discuss limitations of computational power, algorithm choice, and more.

[podcastsponsor]

Watch the video here:

{PODCAST} In-Ear Insights: Why Consumer Recommendation Engines Fail

Can’t see anything? Watch it on YouTube here.

Listen to the audio here:

Download the MP3 audio here.

Machine-Generated Transcript

What follows is an AI-generated transcript. The transcript may contain errors and is not a substitute for listening to the episode.

Christopher Penn 0:02
In this week’s In-Ear Insights we have a special guest John Wall, cohost marketing over coffee Katie Robbert is on vacation enjoying some well earned time off so this we’re gonna talk about consumer AI and recommendation engines we had alzolay right in saying like, why is consumer AI why recommendation engines so bad.

He was talking about an experience where he looked at some brightly colored clothing on on Instagram and now he gets all these like trashy pieces of clothing as ads that have no relevance to anything he wants to buy.

And then when he was on Netflix, you watch one show.

And now every recommendation is all the same style of show and he’s like, I don’t want to watch it.

Other things.

So why is this? So? And there’s a bunch of reasons for this.

But john, what’s been your experience with recommendation engines? Are you finding that when you’re using anything from Amazon to Netflix, whatever that recommendations are on target, or are they off target? How do you How is your experience been?

John Wall 2:17
Yeah, it’s obviously it’s first generation early stage stuff.

I mean, once in a while, I do get wowed, you know, I’ll buy, say, a docking station.

And there’ll be a recommendation for a couple cables that normally go with that.

And like, that’s fantastic.

That’s right on the money.

But at the other end of the spectrum, I’ve got my key tester here from Ollie Express, shipped straight from China.

Very exciting stuff, right? On the nerd friend.

Yet for some reason, I get an email every week from Ollie Express about hair extensions.

I don’t really understand why I would want hair extensions or female clothing to that’s always in there.

But the weird thing is I just get that and I don’t get nerd stuff.

So yeah.

I think it’s just such an early stage that there’s a lot of garbage out there.

Christopher Penn 3:03
Yeah.

So there’s a couple of things that are happening behind the scenes on these.

One is, there are so many different algorithms when it comes to recommendation engines.

And with big companies like Amazon and Netflix and stuff.

One of the primary considerations is how computationally expensive an algorithm is, do recommendation, I’m doing something like clustering, which is, you know, nearest neighbor, hey, people who bought this also bought this is cheap and easy, but it may not be accurate.

And then the second consideration is what data you have to work with, which is a much bigger limitation.

There’s a whole bunch of information that could improve recommendations, but consumers don’t necessarily want to give artificial intelligence, that level of detail, like you can specify, you know, here’s my age, my gender, my sexual orientation, you know, my my other likes, and you could calibrate on that.

But some of the things are protected classes that, yeah, you shouldn’t be using an algorithm or at least If you’re going to use them you have to demonstrate clearly for auditors why your software is not biased.

So the bigger challenge then is how do you What’s that that balance? And it seems like if you’re getting hair extension that it’s not quite there so what’s the fix for it? What do you when you think about recommendation engines? Are you being asked to give feedback on their recommendations at all?

John Wall 4:30
Yeah, that’s another fantastic point is that I have seen there is at least some work going into gathering more data about that.

Like with Amazon, it’s great because you’ll get all your your history and you can actually go in and click and say okay, no, don’t show me these anymore.

Which is the classic scenario of you know, you’re going to buy a gift for somebody, you don’t want 50 more of those gifts.

You know, just because I bought my daughter a lol, you know, one of these plastic surprise toys I don’t want to see lol dolls.

For the other 51 weeks of the year, so being able to flush stuff out is very important.

And then the other one is you can also use that as a win to you the other option right there, show me more.

So if you were shopping for something you want to come back, you can just grill through and get more.

So that and then the other one, which has surprises me is the ad networks not having enough knobs and switches on those, you know, when I’m getting ads on Facebook, I should be able to say, Hey, I already bought that, you know, I mean, just shut that down.

And that would give them some knowledge about what else they could show me because now they know I’ve converted, which the other way otherwise wouldn’t get.

And then yeah, a lot of times you don’t have the option to just say no, stop showing me this entirely.

You know, I don’t need to see the, the hair weaves anymore ever.

Like that’s just a complete waste of everybody’s time.

But so yeah, at least there’s more data collection going on at that.

And I don’t know how about as far as architecturally, because it seems interesting that you could have, you could have the data file the customer, or you could tie it to the product.

I mean, it seems like it would make more sense to anchor things to the product because then it’s not PII and, you know, there’s kind of permanent links you could do.

Christopher Penn 6:04
So there is there is that.

But going back to something you just said one of the, the big challenges that you’re going to get with recommendation engines is that the network’s especially for ad recommendation engines don’t have an incentive to not show ads.

If you think about it, there’s no incentive for me to as an ad tech, not keep spending your money, even if the customers already bought the thing.

We’ve all had that experience.

You go on Amazon, you buy this thing, and suddenly you’re getting retargeted for like, I just bought that.

But there’s no communication from the merchant back to the ad network to say, Yeah, don’t show this to this person anymore.

It’s not relevant.

And, you know, the Facebook and Google ads and and Tiktok and whoever else gets paid on the impression, when you look at the reporting behind the scenes, it’s all the call ecpc and stuff like that.

It’s all CPM ads.

It’s still they’re still earning their money on.

I got to show this thing as well.

number times to get paid.

And so there’s it’s a point of confusion for you why advertisers are not saying, Hey, I don’t want to pay you anymore for this.

Once they’ve converted, I put the tracking code on the forums.

So if you keep showing it, I’m not going to pay you for those impressions anyway, have you ever run into an ad network that’s actually agreed to that?

John Wall 7:24
Well, there are Yeah, there’s a couple networks where you can do cost per conversion, you know, they usually charge a 10 X or some insane, you know, uplift on that.

But it could be worth it on the product.

Yeah, I think it’s an interesting point.

It’s just in the fact that it’s so much additional work to lay the pipe for all that data, that nobody’s willing to go the extra steps to make that happen.

But you would think that in the long run, somebody would finally come out and say, Hey, you know, we guarantee you’re going to do 20 or 30% better because we’re filtering out all this garbage that you’ll never have to pay for.

But I think that’s just it’s a matter of time thing.

And there’s so much more On that space between the vendors and the tools, and the outlets, and it’s just kind of this highest High Seas pirate mass.

Christopher Penn 8:12
Yeah, so the other thing is that, on the recommendations, recommendations typically only follow paths that people have followed.

It’s very similar to attribution, in that it looks at the the things that clustered and clumped together, but they’re not behaviorally or psychographically driven.

In a lot of cases.

They’re just simple associations.

And so what tends to happen is that you get stuff that is categorically similar.

So if you buy a microphone, you’re going to see at best you’re going to see like cables that people bought with that microphone.

There is no psychographic data that says, If you bought this microphone, maybe you need a camera to go with or maybe you need a green screen or maybe you need some sound baffling or Hey, maybe you need a book on how to record good audio, because the machinery does not understand that.

The classic example of this is in the Superman Target stores have gotten really good at point of sale.

If you go to some of the smartest stores that do this, and you go to like the feminine hygiene product aisle there, the smart stores also have a either an end cap or a mid.

I’ll pop up for chocolate bars.

Like, psychologically it makes total sense.

Biologically, it makes total sense.

But those are not items that are going to be clustered together in a standard clustering analysis.

And so a recommendation engine will miss that will go, Oh, these are not related.

I’m not going to show one of the other Where’s humans go? Well, duh, you know.

And I think that speaks to the fact that a lot of the folks who are doing this martec stuff, you don’t have subject matter expertise don’t have domain expertise and things like psychology and anthropology.

You’ve done a ton of stuff in martec in over the StackOverflow podcasts and things like that.

When you talk to vendors, how many of them have any kind of anthropology or ethnography or any kind of expertise like that in their technology teams.

John Wall 10:00
Well, that ends up being, you know, a primary determinant of who’s going to win or who’s going to lose, you know, because you what you find is that the winners are the people that have that domain knowledge.

So it’s usually somebody on the team that is like, Hey, this is the problem that I face every day.

So I’m going to build this to fix it.

And on the other end are the people who are like, hey, we’ve built this widget, we just have to find somebody who has this problem.

And you know, those people never win, they never are able to find their part of the world of people who want to buy.

So yeah, that kind of knowledge right there is what separates the, you know, unlocks the demand.

If you can’t get that knowledge of what people are looking for and why they’re shopping.

You’re just never going to crack the code on it.

And it’s, yeah, and it’s the same thing, what we always talk about with marketing automation tools, in that you have to build these from the ground up, and it’s very easy to go through the three week sales process and buy the tool, but it’s four years of work to build email sequences and figure out the dependencies of the product.

And map them all out and a B test them.

You know, that’s all the hard work which, unfortunately, most marketers are just way too lazy.

They’re not up for that task whatsoever, you know, because I see I think about with Amazon like every book, if you buy any series for any kind of book, like you should just automatically be entered in a series of like, okay, you read this fourth Stephen King book 60 days ago, we should be pounding you with the next Stephen King book, you know, around the clock.

And but it’s, it’s just a lot of work to build that kind of web and to figure out how that all works.

And yeah, I don’t know if the average cubicle going person has the, the guts to do that extra work,

Christopher Penn 11:40
or may not have the ability to do it.

I mean, when you look at algorithms, around recommendation engines, they go from very simple, like, you know, k nearest neighbor clustering or K means clustering, all the way up to Hey, you need a deep neural network that has you know, 400 layers and it’s constantly needing to be retrained with active learning models in order for it to To improve, but if the improvement really scales, again, challenge with that is Amazon can’t do that.

Even Amazon can’t do something like that, because that is so computationally expensive for the size of data they have, they have to do sampling that to do subsets.

And you know, obviously, anytime you subset data, there is always the risk of, you’re going to get less than the full picture.

When you’re a small company, like you know, the size of Trust Insights is we don’t have to sample.

But then we also don’t necessarily have enough data to get a realistic sample versus the marketplace overall.

And so one of the challenges there is how do we say how do we use our data and then extrapolate and figure out what is the bigger marketplace look like of all the people that frankly have never heard of us, because we’re just too small to build that.

And so for recommendations like engines like that, you always have to look alike modeling.

If you think about look alikes, in advertising, look alikes, and recommendation engines are very much the same thing.

As long as you have enough data to to form a point of recommendation, you can then go and start hunting down other stuff.

I was building a tool for a writing group I’m in.

There’s like 1000 different authors and 1000, you know, 10,000 pieces of writing.

And all you only need to fingerprint an author’s one or two good sized chunks of text.

And then using machine learning algorithms, you can figure Okay, here are the other pieces of text that are stylistically similar to this one.

So if you’d like this story, he’ll probably like these these additional stories because they’re very similar in language and tone and emotion and emotional arc and all that stuff.

Whereas when you’re doing like product stuff, there’s actually a lot less data to work with.

Because you don’t we don’t know why the consumer bought this thing.

Like, you go onto Amazon and you buy a lavalier mic.

If that’s your only purchase.

There’s no context around why you made that purchase.

So Amazon’s gonna be like, okay, we’re just going to guess about the other things like here’s a wrench and test it out.

See, like, Did In, you know, for the thousand people who saw this, this, we bought this microphone and we, you know, showed 50 of them a wrench, how many of those 50 bought the wrench to go with it? You know, zero is like, okay, that didn’t work.

And that’s where, again, a lot of marketers are not willing to not willing or not able to put in the extensive amounts of testing needed to see like, is there are there associations we don’t know about? There’s a concept called perturbation testing, which is wait like, 5% of time in any kind of test, you just throw in a random thing, like, you know, here’s here’s a salad.

You know, you’re buying a microphone, here’s a salad.

Does that work? No.

But sometimes that’s the only way that you can find those things that are are a little bit out of the box.

And even then, you need a lot of traffic to do it.

John Wall 14:44
Yeah, right.

But that does seem to be the holy grail for somebody like an Amazon that it’s just impossible to know every single combination of every purchase and what goes with other things.

And hopefully, you would, at least on a regular basis be unearthing stuff that that you can then Just continue to clone.

The other

Christopher Penn 15:02
thing is that Amazon in particular does this but even we do this.

The recommendations are sometimes a little self serving.

like Amazon says, Hey, we’re going to, you know, say Sennheiser comes to Amazon says, Hey, we want to be number one in the recommendation engine for anything microphone related and they could put their thumb on the scales, okay, Sennheiser is gonna get shown 16% more of the time, when you go onto the Trust Insights website.

If you look at the end of any of our blog posts, there’s you might also enjoy reading, those are not contextually based on the content of the article you read.

Those are based on what we call most valuable page assessment where you look at the pages that are most likely to lead to conversions.

And we put 10 of those into a randomized and so the five are shown at any given time, because the premise is if you’re reading the article, and you see one of those other five that you like, no, a lot of our contents stylistically similar, but we know those are more likely to get you to convert We’re going to try and boost our conversion numbers that way, even though it’s not a truest form of recommendation, it’s because it’s not Data Wise related to the post you’re reading.

So there’s that other aspect of recommendation engine students, you may be showing things that are in the interest of the company and not necessarily the interest of you as the user.

John Wall 16:18
Yeah, right.

So look alike detector.

Unknown Speaker 16:21
Exactly, exactly.

Christopher Penn 16:24
But I would say for anybody who’s trying to figure out how to do this stuff, that’s an easy win, right? That’s an easy sell to the boss to say, like, Hey, we want to try doing a content recommendation engine on the blog, and we’re going to feature content that has that helps convert I can’t imagine there are 30 people like no no no, we don’t want to do that.

More conversion Well, let’s but that’s a good place to as a stepping stone to get started with a recommendation engine that can be static.

It doesn’t have to be you know, some real time crazy Matthew, anything you can start out with batches on the stuff.

Have you seen this before? done anywhere with podcasts.

John Wall 17:03
Yeah, you know Apple is the leader of that one.

It’s kind of like okay, hey, you like this podcast, here’s five other podcasts that are similar to that.

And so they’re just doing you know raw numbers as far as like okay the people that download this have downloaded that stuff.

Yeah, I can’t say that I’ve seen anything very smart or intelligent where the recommendation wasn’t just like for other marketing podcasts, you know, it would be interesting if somebody was like, you know, hey, this is uh, this is number one over in true crime or something like that.

So ya know, and then you also it’s as bad as consumer goods and that you really just know that they made the download but you actually don’t even know if they even listened to it.

Right? I mean, it could have been that they download it listen to 30 seconds and said oh my god, this is terrible.

I never want to hear this again.

So I get you know, you could at least try and limit it because they do you know, if you don’t download the last three episodes, I think you get you know, disable that stops auto downloading So you do kind of know what’s hot and what’s live.

But yeah, it’s only the most basic, you know, kind of Oh, you like this, check out that.

Christopher Penn 18:08
Yeah, the one to watch is going to be Google the new Google podcasts thing because Google has shown again and again, that they are willing to invest in like the text mining and the transcription of rich media content to figure out what it is people like when you read how YouTube’s recommendation engine works.

You know, they actually published the paper, the academic paper on it in 2016.

And it’s like, okay, they look at your search history, they look at the query, then they look at the video, the title, the description, the tags and the automated transcript, they can detect the words in what you’re saying, and match that with the behaviors and other things to recommend videos.

I can’t imagine they wouldn’t repurpose that existing code for podcasts.

And so having good clean audio and having, you know, audio that’s on topic with what you want to be known for and be found for is going to automatically just the base level like here’s what you need to do.

Well with Google podcasts now the question is whether Google can actually get cast on that platform.

Yeah.

And

John Wall 19:07
it’s, you know, that’s a brings forth another thing as far as these ad networks, you know, I would love I would gladly throw down 1000 bucks a month to drive traffic to marketing over coffee, you know, to have ads run next to Entrepreneur on Fire, or some of these other casts that we know that when people like this, they like that, but really everywhere I’ve only seen is like, you know, no, unless your Ford Motor Company and you’re gonna drop 150 grand, like, we’re not interested in your thousand bucks a month models.

So there’s still there’s still growth and hopefully things will, you know, get more advanced on that front.

Because Yeah, I’d like to take advantage of that.

Christopher Penn 19:42
If you had a way to model the propensity of people to visit other sites, maybe through SEO data, would that preclude you from like, reaching out to like, you know, I don’t know somebody else’s marketing podcasts and say, Hey, do you want it you know, I’m willing to pay you you know, price The inventory space for this.

John Wall 20:02
Yeah, definitely.

I totally wouldn’t be against that.

But yeah, you know, podcasting is definitely a very tribal thing.

I have talked to certain partners about that kind of stuff.

And it’s like, oh, god, no, I would never share anything with you, you know, that kind of stuff.

So, but yeah, it seems like there is a opportunity for some network advantage.

You know, I don’t see why four or five podcasts with similar topics, or with some kind of complimentary things couldn’t all get together and, you know, get their marketing dollars to go farther for them.

So, but yeah, it’s just never come around.

It’s weird that it’s always been on the production front, you know, you get somebody like gimlet that says, Okay, we have a lot of popular people.

So we’re just going to get under the same roof and I’ll make shows, whereas I’ve never seen for existing shows, say, okay, we’re going to give up a little piece of the land.

It’s maybe too complicated to make happen.

Christopher Penn 20:56
Hmm.

That might be a good exercise.

Something like, you know, Saturday night data party, whatever, we can figure out what the CO locations are different podcasts by domain name or our, you know, you know, for example, every, what two months, there’s another top 10 podcasts in your business to listen to kind of thing.

And it’d be interesting to see which cast of characters repeatedly shows up around a given show and say, Okay, these are the ones that may not be, you know, like you said, they may be complimentary and stuff competitive.

Let’s use that as the basis for forming a consortium or advertising together, or even directly targeting searches for that domain or searches for that show name in your like your pay per click ads, because you know, everybody in the cousins listening to, you know, Michael ports podcast and, and it’s always co located with marketing over coffee, it might be worth paying just to advertise if there’s inventory on his site or something like that.

So there’s opportunities there to essentially build your own recommendation engine, because the technology is the same and use it to target your advertising.

That could be fun.

John Wall 22:00
Yeah, no, I would definitely be up for that.

Because I said, I think there’s potential there.

You know, I think there’s no reason why shows can help build each other’s audiences.

Is it nobody is ever like, Oh, I only want to watch one Western happen.

Yep.

Christopher Penn 22:17
So to wrap up, recommendations engine suck for a bunch of reasons, one, thin data or bad data to self interest by the recommender three computationally expensive algorithms.

And the takeaway here is, if you want to try recommendation engines, could give some real thought if you have the technical capacity to building your own.

If you don’t have the technical capacity, hey, reach out to us.

Go to a TrustInsights.ai dot A and drop us a line.

If you have questions about it.

Go over to our free slack group Trust insights.ai slash analytics for marketers over 1200 folks who are hanging out asking questions making discussing this stuff like this all the time, and who knows we might try and some of these experiments that we’re talking about because I think there’s a there there.

Using recommendation technology to help better targeting of advertising and maybe even content distribution.

Again if you got follow up questions about this, please leave them in the comments wherever it is you’re consuming this or over on our website Trust insights.ai.

I’ll talk to you soon.

Take care want help solving your company’s data analytics and digital marketing problems, visit Trust insights.ai today and let us know how we can help you


Need help with your marketing AI and analytics?

You might also enjoy:

Get unique data, analysis, and perspectives on analytics, insights, machine learning, marketing, and AI in the weekly Trust Insights newsletter, INBOX INSIGHTS. Subscribe now for free; new issues every Wednesday!

Click here to subscribe now »

Want to learn more about data, analytics, and insights? Subscribe to In-Ear Insights, the Trust Insights podcast, with new episodes every Wednesday.

Leave a Reply

Your email address will not be published. Required fields are marked *

Pin It on Pinterest

Share This