Creative Planning Meeting – Ultra Offices

In your regular creative planning meeting, a question has come up. The team wants to know not just what creatives are working, but why? To do that you have to do an analysis of what’s in each creative so you can identify patterns of what’s working. 

To see what attributes of our creative are working we have to tag them with what’s in the image
But I don’t think that’s realistic given we’ve tested hundreds of ads
I have an idea though
Why don’t we use Google Vision to tag the images for us?
AI is getting pretty good, and we wouldn’t have to do all this manually and can get insights right away
When we launch new creative we just need to run the script again to tag it
It could even be set up as a fully automated report
Do you want to check it out and see how it works?

Once thought futuristic and unlikely to happen in our lifetimes, the ability for computers to “see” has progressed exponentially in recent years. Machine learning algorithms are trained on millions of images and their labels, so that they can learn to recognize objects, faces, and landmarks. The technology can be used to automatically tag images with metadata, so that they can be easily searched and indexed. There are a range of applications from security and surveillance to image and video search, and decomposition analysis: breaking marketing creative down into its component parts to analyse what’s working.
Companies like Google and OpenAI have democratised these techniques through open source libraries and commercial APIs bringing the cost of experimenting with them down considerably. Now anyone can tag any image they like in a few seconds without owning a supercomputer or knowing anything about machine learning. Marketers who want to go a level deeper with creative analysis, can use computer vision to automatically tag each image used in an ad, and then pivot performance metrics to see what labels are correlated with your best and worst performing ads. Then these ideas can be used as new hypotheses to test in your creative testing plan, to validate if they work.

Image Recognition

Hey, I'm gonna show you how to use AI to tell what's performing inside of your creative images. So we're gonna use Google's vision api and to do that, we just need to run pip and store Google Cloud Vision.
Still gonna install the library that we need. And then we're gonna just import these libraries here. So when you're running these cells, it's literally just click into the cell and shift enter, or command enter if you're on a Mac or you can click play here.
So what we're gonna do is we're gonna pull the campaign stats, the campaign results in first. So these are the, this is the performance data that you get from your campaign.
So here we have like a specific ID for each creative that we ran, that each one of these is an image and we've got the reach impressions, click spend, conversions whatever else you need.
So before we can do that, we need to upload it. And you should be able to see that. We can see the campaign results CSV here.
So what we need to do, if you're in the notebook, it might you know, just have something like this. You just click on the folder and then click upload.
And then you can kind of choose the file to upload it there. You also wanna do that for the images.
So you're gonna get a link to all images in this folder. I've got all images here look through on my computer.
So you can see these are just the different ads that we ran you know, for our campaign. And we wanna know, okay, is it, you know, pictures of women or men?
Or is it kind of like, you know, what style of images worked really well for our campaigns? Or was it these kind of, you know, clay emoji style images that worked?
And you know, obviously there's like 20 ads here. Yeah, you could go through and kind of manually tag them but it might be interesting to see what the AI comes up with.
So once you've uploaded, you can just run this and get the campaign results. And then what we're gonna do is we're gonna use Google Vision to tag each image and then join that to the campaign results.
So what this is doing is just running through and once you've uploaded all the images, so you know, in this case you click upload and then find all the images on your computer and then upload 'em that way.
This is gonna make sure that you have all those images. So it's just gonna run through look in this folder in the Google CoLab notebook, and then it's just gonna kind of find everything that's a PNG in that folder.
So don't worry too much about how this works. It's just gonna display them with the name here. So just gonna let that run.
Here we go. Okay, so you can see what all the different images look like. This is useful because sometimes you need to remind yourself like, What, what, how did we run for that campaign?
Cool. The other thing you need to do is get a service account json. So this is basically like a password but it's a file that you can save.
So if you go to this page here, there's a link it tells you how to get started. So if you haven't set up a Google Cloud console project before it's gonna walk you through how to do that.
You need to enable billing. So it doesn't work unless you don't have a credit card behind it. It's not really gonna cost you very much unless you do something hilariously wrong.
But you know, for me it's like a, I think a couple cents, and you usually get some credit that you can use like a few hundred dollars.
When you set up the Google Cloud project you can also the, the next thing you need to do is enable the api.
So literally, if you click on this, it will actually just kind of look at the project and then it's gonna confirm you know, within your project if you're happy to enable it.
So that's the next thing you do. And that's just kind of a fail safe to make sure you don't use the wrong APIs.
And then the next thing you do is get a service account key. So if you click go to create service accounts then that's gonna kind of give you the different projects and you're gonna choose that.
And then you're gonna go through this flow and just create and continue. Once you've got a service account then you basically just need to click into it.
And then and then you wanna go to Keys. And then you can download a key here. So you can go add a key, create a new key and I'm gonna delete mine after this.
So you know, don't worry <laugh>, you, you won't be able to access yeah, cause I'll, I'll have it deleted by the time I put the video live.
And I recommend you don't share yours with anyone. But once you have, once you have the key, so like, if I create a new key, it'll say like, create, and then it'll download some j to my folder.
And then you basically just then upload it here. I've, I've renamed it so you can click into a file and go rename, and I've renamed it.
So account J just to make the code run, but you can just set yours to whatever your thing is named here.
So if you run that fine and then and then you go in here. So this is the actual code that finds the labels.
So this is just a test run, which try it with one of the images. In this case it's gonna be, let's say image 10.
So or the, I guess, let's say 11th image. Cause it starts counting from zero. So it's gonna open the image, so it's gonna read the image, then it's gonna turn that into A Google vision image.
And then when it's in that format, then you can get the label detection running. And there's actually a bunch of other things you can do with this, but label detection is the most immediately useful.
And then once you get the response, so this calls the API for Google Vision and it takes a few seconds, you get the response and then you can get the labels as well as other things as well.
But labels is all we need. And then what we're gonna do here is just to test its work and we're just gonna display the image with the labels underneath.
So let's run that and see if it's working. Here we go. Yeah, it worked. So this was our image. And then we got the labels of hair joint head, shoulder, eye flash, photography, neck sleeve, dressed eye and we can check other images as well.
So lemme try another one. This is the, the first image. So zero. So this is a lemon. Picked up the blue and the sphere and still life.
And then you can just basically, you know, kind of try out all, all the different images if you like, kind of see what it comes up with.
It's not it's not always fully accurate. I think it's a good kind of directional understanding. Sometimes it doesn't get them right.
These images themselves are actually generated by AI as well. So sometimes it does get confused on these more than it does with the real the real images that you know, that you would use from your account.
Cool. So now that we know it's working, we can just run this to process all the images. So it's gonna go through every single image and do the same thing again except instead of just printing the labels, although it will print them it will also then add them to a list.
And then we're gonna use that list afterwards to join the data together. So I'm not gonna run that again because it's you know, it's already run here, takes a few minutes, but it depends on how many images you have, but you can see that, you know, here, it can tell that that's a peach, which is pretty cool.
It thinks that this is a toy or maybe some office supplies <laugh>. So you can see this doesn't quite get it.
But you know, these, I think these fashion photography ones seem to work a little bit better. So, you know, it's picking up quite a few things here.
So it's quite interesting. So you can see all the logos as it comes through. And then and then if you scroll down to the end now this is where we start to get the data into the right format.
So what this does is it just goes through every label. And then for every label it adds it with the creative.
So you know, if C 39 had you C 79 creative had the label of ball and solo photography and circle and sphere and electric blue.
So this just showing the first five roads. And then and then we're basically gonna then join that together with our other our other data, the metrics df, which was the, you know, impressionist, click spend, et cetera.
So that just does merge. And there we go. Now we can see cuz C 39 had the label of ball, we've given that, you know, the, the metrics from C 39 and it's, you know, it's given that for each one of the labels.
And then what we're gonna do is kind of join it all together so that you know, if multiple creators had you know, this circle label then they're gonna add up all of their metrics.
So that's what this does here. Just gonna run this. And there we go. So this is kind of the beginning of the table and what this has done is, is kind of grouped by the creative and then just counted how many creatives it was in and, and kind of back which creatives it was in as well.
So so you see the action figure was in C 44 and arm was tagged in C3 C 34 and C 25.
So that, we're gonna use that in a second. The only thing left to do really is to attach all the, all the metrics.
So again, this is just going through each column and then getting you know, just the metrics data, so from column to onwards and then grouping by and summing it and then kind of adding it into that column.
So every creative, so you know, this one action figure it was just in creative C 44. So this, you know, the reach impressions clicks will be just from C 44 but this one arm was in C three, C 34 and C 25.
So it has the combination, the sum of all those metrics. The other thing I did was just calculated the click through rate.
That's what I'm gonna look at for this. We're gonna have a lot of conversions. But here it's just the clicks divided by impressions times by a hundred.
And that's what I'm kind of using to tell which creatives were good or not. Cool. So one thing you'll notice is that like, if it's only in one creative, it doesn't tell as much.
So it's saying action figures, you ranked the same as art.  1.18%, click through eight cuz that was the click rate for this one ad.
So what's more interesting if it's in multiple creators, but it wouldn't be that interesting if it was in all the creators, cuz then that would just be the average, right?
So what we want is any you know, we, we don't want it in all creatives. So this is saying basically filter for any time where the creative count was you know, equal or, you know, it's basically it needs to be less than less than the length of creative names.
And then this three or more filter basically you know, needs to have at least three at least three in the creative count call.
Cool. So then the filter df I've also sorted it by values, so sorted by click three Descending, and then that shows the head here.
So you can run that and you can see, there we go. And now we can start to tell you know, what, what types of tags are showing up as performant.
And you could look at the head and, and see that you know, there's this one has like it's only in three creatives, but it's actually like quite a high click rate when the circle thing, I think this is showing up in the labels, like in the in, in the, those specific you know, the kind of the claymation emoji thing.
So that, that tends to show up there. So that's an interesting thing to look at. We also see neck, when we show the neck.
The click rate is pretty high. Flash photography is good sleeve forehead. Okay, So that's kind of telling us roughly, you know, what's working.
We could also look at what's not working. So is the tail. And we can see the very end. So these are the things, the lowest click rate.
So when the notices, there's a hairstyle that seems to perform worse. You know, when there's a white color in there that seems to perform worse then this electric blue color, so that's in some of the claymation ones is performing worse.
So you start to get a sense of what sorts of things are improving performance and what sorts of things are decreasing in performance.
And these are just correlations. So you do need to test these hypotheses, but this is a great way to find new ideas of what to test.
All right. Hopefully this was useful.


Why might you want to work with computer vision to tag your ad creatives?

Which creatives feature the label “Forehead”?

What’s the worst performing tag in terms of CTR?

Why does it make sense to remove tags that don’t appear in at least 3 creatives?

Why might you have to be careful about making decisions solely off this type of creative analysis?

You know which creatives are performing, but do you know why? For that you have to look at what’s IN the creative, to see what’s really driving performance. 

Vexpower Learning Certificate of Completion

Congratulations Carina Obster 💪