Intro

PressPlay by Phiture automates the creation, deployment and analysis of A/B tests on the Google Play Store. It uses generative AI to create visual assets in huge volumes, leveraging Phiture’s decade-long expertise in ASO. These assets, in the form of icons, feature graphics and screenshots, are then tested on the actual Google Play Store, powering tens of simultaneous experiments, with close to none human intervention.

This is enabling gaming companies like Wildlife Studios and Lockwood Publishing to achieve millions of additional projected app installs, in just a few months.

However, one of the hardest problems in AI is reliable consistency. It’s one thing if you’re prototyping creative ideas in AI, but to actually publish an AI-generated image requires a level of consistency that’s hard to achieve with prompting alone. The image isn’t going to be approved if it doesn’t conform to the style used across a company’s brand assets.

Phiture came to me with this problem – having done remarkable work to automate the testing of new brand assets with their PressPlay tool, the new bottleneck was creating new assets to test. AI-generated assets weren’t cutting it, because it was an onerous trial-and-error approach to match a client’s style, and even then they were always noticeably off. If they could automate the generation of assets, PressPlay could automatically cycle through many iterations of tests using AI to generate new ideas, driving growth for their clients many times faster than would be otherwise possible.

Thankfully I had written in my prompt engineering book (published June 2024 through O’Reilly) about creating consistent characters with AI, including fine-tuning a model based on a specific style or object. In this post I’ll walk through how I approached it, so you can see what’s possible with AI.

Fine-tuning with Dreambooth

I’ll start with the hardest technique, because it’s what gets the best results. Dreambooth is a training technique you can use on open-source models like Stable Diffusion XL, which actually changes the parameters in the model – it ‘learns’ about the concept you teach it. That concept is typically either a style or subject, of which you need 5-30 images to upload for the model to train on. There is another popular technique for fine-tuning called LoRA, but I use Dreambooth because it gets better results on faces.

Here’s an example from my book, where I trained Stable Diffusion 1.5 to generate images of myself:

There are plenty of options online in terms of tutorials you can follow (I like this one), but I would recommend some technical ability if you’re going to try this method. You probably won’t have to write any code, but you will have to know how to run it. The good news is that you don’t need the latest M3 Macbook or gaming PC with access to a GPU (graphics processor used by AI models), as Google’s Colab Notebooks offer access to GPUs for free or very cheap (~$15 a month), depending on which model you’re using.

Note: you can also try out Replicate, which I’ve had great experiences with before. However, as of June 2024 their customer service team admitted to me that their training methods were somewhat out-of-date relative to the results you can get within Colab, and they will be revisiting this feature.

Here’s the training data we uploaded on behalf of the client: ~30 cut out images with no background, of a specific character in the game Avakin Life in different poses and positions:

The training only took around 15 minutes to complete, and at the end we had a new set of model weights we could prompt with the word “UKJ” (a made up word, that the new model associates with our images):

Compare this to what we got when we prompted the normal Stable Diffusion XL model that didn’t know about our character and brand style:

When Phiture used fine-tuned models to automatically deploy icon tests created by PressPlay’s generative AI, Avakin Live saw remarkable results. Ouni Kwon, Senior Marketing Director at Lockwood Publishing said: “Phiture’s PressPlay has been instrumental in scaling our experimentation efforts and deepening our understanding of how to effectively communicate our brand to players. We are seeing successes where we didn’t expect to and have been able to apply the learnings across other areas in our marketing while staying true to our brand identity.”

As you can see, fine-tuning makes an astonishing difference in the quality of the assets, and really nails the client’s unique style. AI-generated assets aren’t replacing human designers any time soon, and if you have good quality assets to use you should. However, AI being able to fill the gap in scenarios where you need assets over and above what would be possible for your team to produce, is a powerful edge. You can always take the winning variants from AI generated asset testing and then make human versions of them later, to achieve stronger performance in less time.

Deep-dive: Running Stable Diffusion XL In Google Colab

If you don’t code, you can stop here, because we’re going to go through how this technique works. If you have a developer in your team, you could send this post to them and they should be able to replicate my results in half an hour.

First, you need to check you have an Nvidia GPU, which you can accomplish by typing !nividia-smi and pressing Option + Enter on a Mac, CTRL + Enter on a PC. If you don’t have a GPU then you can assign one by clicking the dropdown in the top right corner, and choosing ‘Change runtime type’, before selecting one of the GPUs (I used the L4 one).

Next you need to install the diffusers library by HuggingFace, which is what helps you run Stable Diffusion: %pip install --upgrade diffusers[torch] -q

If you want to run the base version of Stable Diffusion XL, you can run it with the following code. Stable Diffusion is open-source, so it’s useful to be able to run it locally like this, and all you have to do is run on the code. Just change the prompt to be whatever it is you are trying to create:

from diffusers import DiffusionPipeline, AutoencoderKL
import torch
from PIL import Image


def image_grid(imgs, rows, cols, resize=256):
   assert len(imgs) == rows * cols


   if resize is not None:
       imgs = [img.resize((resize, resize)) for img in imgs]


   w, h = imgs[0].size
   grid_w, grid_h = cols * w, rows * h
   grid = Image.new("RGB", size=(grid_w, grid_h))


   for i, img in enumerate(imgs):
       x = i % cols * w
       y = i // cols * h
       grid.paste(img, box=(x, y))


   return grid

vae = AutoencoderKL.from_pretrained(
    "madebyollin/sdxl-vae-fp16-fix",
    torch_dtype=torch.float16
)
pipe = DiffusionPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    vae=vae,
    torch_dtype=torch.float16,
    variant="fp16",
    use_safetensors=True,
)
pipe.to("cuda");

prompt = "sims3 avatar of a woman riding a horse"


image = pipe(prompt=prompt, seed=42, num_inference_steps=50, num_images_per_prompt = 3)
image_grid(image.images, 1, 3)

We found the closest prompt to the client’s style we could find was evoking the name of sims3 and specifying it was an avatar. Here’s the resulting set of images:

Here’s what the code does, step-by-step:

1. Imports and Setup:
   • from diffusers import DiffusionPipeline, AutoencoderKL: Imports the DiffusionPipeline and AutoencoderKL classes from the diffusers library. These are used for setting up and using the diffusion model.
   • import torch: Imports the PyTorch library, which is used for tensor operations and as the backend for the model computations.
   • from PIL import Image: Imports the Image class from the Python Imaging Library (PIL), used here to handle image operations like resizing and merging into a grid.
2. Function Definition image_grid(imgs, rows, cols, resize=256):
   • This function takes a list of image objects (imgs), the desired number of rows and cols for the grid, and an optional resize parameter to adjust the size of each image.
   • assert len(imgs) == rows * cols: Ensures that the number of images matches the expected number of grid slots.
   • Resizing images (if resize is not None): Resizes each image to the specified square dimensions.
   • Grid dimensions are calculated based on the number of columns (cols), rows (rows), and dimensions of the first image.
   • A new blank image grid is created with the computed width and height.
   • Images are placed into this grid in row-major order using a loop that calculates the position for each image.
   • The function returns the final image grid.
3. Model and Pipeline Setup:
   • vae = AutoencoderKL.from_pretrained(...): Loads a pre-trained autoencoder model. This model is specialized for handling images in a reduced precision (float16) for efficiency.
   • pipe = DiffusionPipeline.from_pretrained(...): Sets up the diffusion model pipeline with the previously loaded autoencoder, specifying model settings such as precision and safety features to manage GPU memory efficiently.
4. Moving the Pipeline to GPU:
   • pipe.to("cuda"): This command moves the pipeline’s computations to a GPU (if available), allowing for faster processing.
5. Generating Images:
   • prompt = "sims3 avatar of a woman riding a horse": Defines the text prompt based on which the images will be generated.
   • image = pipe(prompt=prompt, num_inference_steps=50, num_images_per_prompt=3): Generates images using the defined pipeline and settings. The settings indicate that 50 inference steps should be used, and three images should be generated from the prompt.
6. Displaying the Generated Images in a Grid:
   • image_grid(image.images, 1, 3): Calls the image_grid function to organize the three generated images into a grid with one rows and three columns.

Running Dreambooth Fine-Tuning In Google Colab

Now we have established a baseline, we can run the Dreambooth training process to fine-tune the model on our data. We’re going to use autotrain, which is another package by HuggingFace:

%pip install -U autotrain-advanced

Make sure you restart the Colab Notebook now, to free up memory from running the base version of Stable Diffusion XL, and to apply the autotrain library. The final thing you have to do before running training is upload the images to Colab. On the left-most menu, click the file icon, right click then click ‘New folder’, and call it image_training. Upload the images you want to train on to this folder by right clicking the ellipses on the right hand side of the folder, and clicking upload.

Note: You want to upload your image files to the folder without opening it, or you’ll run into an error – it will add an ipython-checkpoint file which annoyingly causes an error with the training script. If you accidentally open the folder then delete it, and try again with a different folder name.

Finally, you can run training, which can be done with the following code:

!autotrain dreambooth \
--train \
--model "stabilityai/stable-diffusion-xl-base-1.0" \
--project-name "DreamboothSDXL" \
--image-path "image_training" \
--prompt "photo of UKJ" \
--seed 42 \
--resolution 1024 \
--batch-size 1 \
--num-steps 500 \
--lr 1e-4 \
--mixed-precision fp16

Here’s what each parameter means:

• !autotrain dreambooth: This initiates the autotrain tool specifically for the Dreambooth task, which is a method of personalizing generative models (like image generators) to specific styles or subjects.
• --train: This flag specifies that the operation is a training run, as opposed to other possible operations like evaluation or inference.
• --model "stabilityai/stable-diffusion-xl-base-1.0": This specifies the model to be used for training. In this case, it is using the stable-diffusion-xl-base-1.0 model provided by Stability AI, which is likely a large-scale version of the popular Stable Diffusion image generation model.
• --project-name "DreamboothSDXL": This sets the name of the project under which the training run will be saved. It helps in organizing and retrieving the training run data later.
• --image-path "image_training": This option points to the directory where the training images are stored. These images are what the model will use to learn from during the training process.
• --prompt "photo of UKJ": This specifies the text prompt associated with the training images. The model will learn to associate this prompt with the visual content of the images in the specified directory.
• --seed 42: This sets a specific seed number for the random number generator used in the training process. This is important for reproducibility, ensuring that training runs are consistent if they are started with the same seed.
• --resolution 1024: This determines the resolution of the images that the model will generate during training. Here, it is set to 1024 pixels, which implies a resolution of 1024×1024 pixels.
• --batch-size 1: This sets the batch size to 1, meaning the model processes one image at a time during the training. This is often used for more memory-intensive models or when GPU memory is limited.
• --num-steps 500: This specifies the number of training iterations or steps the model will run through. More steps generally lead to better learning but increase training time and resource consumption.
• --lr 1e-4: This sets the learning rate to 0.0001. The learning rate controls how much the model weights are adjusted during training in response to the error rate. Smaller learning rates can lead to more stable but slower training.
• --mixed-precision fp16: This flag enables mixed precision training using the fp16 (16-bit floating-point) format, which can speed up the training process by reducing the computational demand on the GPU.

The training should take about ten to fifteen minutes, and when it’s done you could go into the files and download the pytorch_lora_weights.safetensors file, to use in AUTOMATIC1111’s Stable Diffusion Web UI, ComfyUI, or most other popular Stable Diffusion user interfaces. However, you also just run the new model weights directly in Colab, like we did with the base model.

from diffusers import DiffusionPipeline, AutoencoderKL
from PIL import Image
import torch


def image_grid(imgs, rows, cols, resize=256):
   assert len(imgs) == rows * cols


   if resize is not None:
       imgs = [img.resize((resize, resize)) for img in imgs]


   w, h = imgs[0].size
   grid_w, grid_h = cols * w, rows * h
   grid = Image.new("RGB", size=(grid_w, grid_h))


   for i, img in enumerate(imgs):
       x = i % cols * w
       y = i // cols * h
       grid.paste(img, box=(x, y))


   return grid


vae = AutoencoderKL.from_pretrained(
   "madebyollin/sdxl-vae-fp16-fix",
   torch_dtype=torch.float16
)
pipe = DiffusionPipeline.from_pretrained(
   "stabilityai/stable-diffusion-xl-base-1.0",
   vae=vae,
   torch_dtype=torch.float16,
   variant="fp16",
   use_safetensors=True,
)
pipe.to("cuda");
pipe.load_lora_weights("DreamboothSDXL", weight_name="pytorch_lora_weights.safetensors")


prompt = "photo of UKJ riding a horse"


image = pipe(prompt=prompt, seed=42, num_inference_steps=50, num_images_per_prompt = 3)
image_grid(image.images, 1, 3)

Here are the resulting images:

Most of this is the same as before, but there are two changes:

• vae = AutoeratorKL.from_pretrained(...): Loads a specific pre-trained version of an autoencoder, set to operate in float16 for efficient GPU computation.
• pipe.load_lora_weights("DreamboothSDXL", weight_name="pytorch_lora_weights.safetensors"): This line loads additional specialized weights into the model, to use our fine-tuned model weights.

As we have demonstrated, employing techniques like Dreambooth for fine-tuning AI-generated brand assets can significantly elevate their relevance and consistency with existing brand aesthetics. For companies like Phiture, building advanced AI tools like PressPlay can offer a real edge in improving performance through creative testing. Embracing these technologies not only enhances efficiency but also opens new possibilities for creative innovation in digital marketing.

Apple, holding significant market share, sets industry standards with its operating system updates. At WWDC (Worldwide Developers Conference), Apple annually announces significant changes impacting both businesses and consumers.

We watched the keynote for you and summarized the key updates from WWDC 2024 for app marketers.
We’re updating this article continuously as new updates are announced throughout the week.

iOS 18 introduces new customization options for the home screen and control center

iOS 18 and iPadOS 18 will bring updates like more flexible app arrangement, and customizable app icons, with dark backgrounds or tints. Icons on the home screen can now also be enlarged. Users can lock specific apps with a passcode or Face ID and hide apps from the home screen, which will then appear in a “Hidden” folder in the app library. Additionally, app names can optionally be hidden, leaving only the app icon on the home screen. Apple has already updated documentation around app icon variations for iOS and iPadOS 18 here.

The control center receives significant updates. A new Controls API now enables third-party apps to create controls that can then be placed into the control center or lock screens by users.

What the new customization options in iOS and iPad OS 18 mean for app marketers

Having recognizable elements in your app icon becomes more important than ever. Once tinted app icons get adopted by users, your brand colors likely won’t play a big role on all of your users’ home screens anymore. Make sure your users can still easily identify your app by including unique brand elements or text on your icon.

Especially apps in the utilities categories should take advantage of the new Controls API, which will let you add certain actions from your app in more places throughout iOS — like to the control center, or onto the lock screen. This is not only beneficial for your app usage and feature discoverability, it should ultimately be a UX improvement for your app users.

MacOS Sequoia’s Continuity feature brings iPhone even closer to the Mac

Together with iOS 18, updates to the Continuity feature on Mac include iPhone mirroring — which quite literally brings your iPhone screen to your Mac, now effectively allowing users to use all of their iPhone apps seamlessly on their computer. This includes receiving notifications from iPhone while you’re on your Mac. Apple also made it easier to share files between devices using Continuity, which now allows you to just drag files from your Mac “into” an iPhone app you’re currently mirroring.

How macOS Sequoia’s update to Continuity benefit app marketers

These updates mean you can be close to your users in more situations than ever before — even if you don’t offer a dedicated Mac app. Users who use the new Continuity features will be able to receive your app’s push notifications directly on their computers, and can use your app in more situations.

Apple Intelligence integrates AI deep into Apple’s devices — including third-party apps

Apple’s newly announced take on AI, Apple Intelligence, introduces advanced AI features across its entire ecosystem, enhancing OS-level features with large language models (LLMs), and offering sophisticated writing tools for generating, rewriting, and proofreading text. These features are accessible system-wide, including in third-party apps via the App Intents API.

Siri is now more personal and context-aware, capable of natural speech commands, in-app actions, and advanced queries, and can be invoked by simply double-tapping the home bar and typing a prompt or query.
New tools like Genmoji, Image Playground, and Image Wand enable custom emoji creation, on-device image generation, and photo cleanup or editing with the help of AI.

AI enhancements to iOS, iPadOS, and macOS also include improved semantic search for images and videos and direct integration of ChatGPT-4o with access to premium features if users connect their premium ChatGPT account.

Considering Apple’s focus on privacy, it’s no wonder processing for most features occurs on-device, while more demanding queries get sent to a server securely through a new architecture called Private Cloud Compute.

These updates will be available in beta with iOS 18, iPadOS 18, and macOS Sequoia. Learn more about individual Apple Intelligence features here.

What Apple Intelligence means for App Marketers

Users might soon interact with your app in a whole different way — through Siri acting as an AI agent, fulfilling requests and collecting information on their behalf. Explore the App Intents API and make sure to implement the API in time before the iOS 18 release in September to be ready from the get-go.

AppAttributionKit expected to replace SKAdNetwork, supports alternative app stores

While the AdAttributionKit framework has been introduced with iOS 17.4 already, it is now expected to fully replace SKAdNetwork. According to Apple, AdAttributionKit supports re-engagement, click-through attribution (including support for custom creative), JWS formatted impressions and postbacks, and more.
A new developer mode for AppAttributionKit has also been announced, which makes debugging easier by shortening conversion windows and removing random time delays.

In contrast to SKAdNetwork, AdAttributionKit also works with alternative app stores.

Custom Product Pages (CPPs) to support deeplinks

A session description on the Apple Developer website indicated that Custom Product Pages will soon support deeplinks. Whether this applies to apps that aren’t yet installed remains to be seen — we’ll update the article as new information becomes available.

Nominate your apps for featuring on the App Store

App Store Connect very soon will let you submit your apps for possible featuring on the App Store. A session description on Apple’s Developer website gave us an early hint, and now, members of our ASO Stack Slack Community have reported that the featuring nomination feature on App Store Connect is already rolling out to their accounts. Check your consoles and let us know if it’s already available for you too!

Featuring notifications and marketing assets via the App Store Connect mobile app

The App Store Connect mobile app also gets a welcome new feature: You can now choose to receive push notifications on your iPhone when your app gets featured on the App Store somewhere in the world. You can then even go one step further and generate and export marketing assets for important achievements or updates — right from App Store Connect.

Win-back offers for subscriptions

Developers of subscription apps can now encourage churned subscribers to resubscribe with new win-back offers that appear in the App Store or inside the app itself. On macOS, developers can even offer discounted or free subscriptions for a period of time, this feature won’t be available on iOS yet.

Hear about these updates and more — Live on stage!

The ASO Conference is the ultimate knowledge-sharing event for mobile growth leaders. Today, we’re incredibly excited to announce the first wave of speakers!

Join us to hear from industry leaders sharing growth lessons that you can apply to your app or game.

→ Get your tickets here: https://asoconference.com

In May 2023, Google started testing a new branded ad slot on the Play Store, which has since become a permanent fixture. This slot requires app publishers to bid on their own brand names to secure the top ad placement, a move that has significantly reshaped app marketing strategies and impacted organic download metrics.

What is the new branded ad slot on Google Play about?

Now, when you search for a brand on the Play Store, the top result can be a paid ad, followed by organic results below. This means publishers must financially compete for visibility on their own brand names, making it harder to rely on organic search as a free resource. If they don’t bid, competitors can easily take the top ad spot, potentially diverting traffic that would have naturally gone to them.

Impact on organic downloads

Through working with our clients, Phiture noticed some clear patterns. Despite the total download numbers having remained steady, the proportion of organic downloads has been declining. This shift suggests a growing emphasis on paid downloads, which complements the traditional focus on organic App Store Optimization (ASO) strategies, highlighting the need for a balanced approach to enhance overall app visibility on the Google Play Store.

Moreover, conversion rates for organic traffic have dropped since the new ad slot was introduced, suggesting that the quality and engagement of organic visits might be weakening.

Impact on download decisions

Despite these challenges, the impact on actual download decisions seems minimal for users with high intent. For example, someone searching specifically for “N26” is unlikely to download “Revolut” instead, even if “Revolut” appears in an ad.

Interestingly, Apple’s approach in the App Store offers more creative freedom. Their ads can show up to three screenshots of the competitor’s app, which might capture user attention more effectively than Google’s format that only displays the app icon and metadata.

How to test the impact of Google Play’s new branded ad slot for your app or game

To truly understand the impact of this ad slot for your specific app or game, Phiture recommends testing by turning off all Google App Campaigns (GAC) to analyze the cannibalization effect and the true value of organic downloads when they aren’t overshadowed by paid ads. We also suggest strategically removing branded keywords in specific markets to mitigate unwanted ad placements, although Google might counter this by serving the ads in different markets.

This change in the Play Store’s ad strategy highlights the need for app publishers to continually adapt their marketing strategies. Balancing visibility through paid placements with nurturing organic growth has become more complex, requiring a dynamic approach to App Store Optimization and marketing budget allocation. As the digital landscape evolves, staying ahead of these changes and adjusting strategies accordingly will be crucial for maintaining a competitive edge and ensuring sustainable growth in the app stores.

First prepared by the Phiture team for clients and partners, we’re now publishing our take on Apple’s response to the EU Digital Markets Act (DMA) publicly.

Update 03/29/2024: The European Commission has now launched investigations into Apple, Google, Meta, and Amazon for potential non-compliance with the Digital Markets Act (DMA), with possible heavy fines for violations. As a result of these investigations, we expect to see further changes to Apple’s policies and will keep this briefing up to date to reflect those in the future. Read more here: The Verge

For app publishers who choose to opt in to enable alternative distribution and payments, Apple’s recent proposed changes in response to the EU Commission’s Digital Markets Act (DMA) mark a significant shift in its app distribution and payment processing policies.  

The DMA aims to promote competition by imposing restrictions on “gatekeeper” companies (including but not limited to Apple and Google) that operate at a certain scale. 

Regarding app store usage, the DMA stipulates two important requirements for gatekeepers: one, allowing alternative in-app and out-of-app payments, and two, enabling the installation and use of third-party software applications or application stores. 

Google’s compliance with the DMA focuses more on alternative payments. This is because Android is a more open platform that has supported sideloading and alternative app stores on many iterations of its devices since it was first released. Google introduced user choice billing in 2022, which allows developers to offer alternative payment methods, but still charges a reduced platform fee on those transactions. 

Apple’s compliance with the DMA is more complex because it must support alternative payment methods as well as alternative app stores. 

Apple has introduced new business terms in the EU that give developers a choice. They can either abide by existing App Store policies or opt into new terms that allow them to utilize alternative app stores and billing methods under different commercial conditions. Under these new terms, developers will pay reduced commission fees for payments made through their apps distributed from the App Store. They will also pay a payment processing fee if they use Apple’s native in-app payments process but no fee if they use alternative providers or accept payments via an external website link. Additionally, all developers who opt into these new terms will be obligated to pay a Core Technology Fee (CTF) for every first annual install of their app above 1 million generated from either the App Store or an alternative app marketplace.

There is no deadline to enter the new business terms, and developers are encouraged to take their time to decide on what is best for them, as once they sign the addendum to change the agreement, the new terms will apply to all apps in the account, and developers can’t go back to the previous business terms.

Summary of Apple’s changes in response to the DMA

1. New (opt-in) business terms, which unlock alternative app distribution
   These include a reduced commission on App Store transactions, an optional payment processing fee for using Apple’s system, and a highly contentious Core Technology Fee for apps distributed from the App Store or alternative marketplaces. This fee is €0.50 for each first annual install over a 1 million threshold on iOS in a 12-month period and includes first-time installs, re-installs, and app updates by iOS users with accounts from the 27 EU countries.
2. Apple introduces a new set of business terms for app distribution and monetization, offering developers two options:
   Stick with the existing App Store policies, with no changes, or adopt the new terms to access alternative app stores and billing methods under revised commercial conditions.
   

   

   

3. Alternative app distribution and payments
   Developers can now distribute their iOS apps outside the App Store and use alternative payment processors within their EU apps across Apple’s operating systems.
   This could lead to an increase in reach, but costs associated with the new business terms are significant.

   

4. Enhanced user protections
   Apple will introduce new safeguards to minimize risks associated with sideloading and alternative payment processing. These include notarization for iOS apps, authorization for marketplace developers, and disclosures on alternative payments.
   These safeguards are good news for both developers and consumers.
5. EU-specific adjustments
   Apple will allow EU iPhone users to select a default web browser and contactless payments app, potentially bypassing Apple Pay.
   This could be interesting for e-commerce apps and others that wish to drive off-platform purchases while still distributing their app through the App Store.

Phiture’s take

• Apple has worked hard to make both side-loading/3rd party stores and alternative payment methods commercially unattractive to app developers. As Eric Seufert coined it: Apple is playing “Heads I win, Tails you lose” in what amounts to an attempt at malicious compliance.
• Only very high average revenue per unit apps currently look like potential winners from accepting the new terms.
• It’s still very early days, and Apple’s current response is just a proposal
  • Regulators will review the proposal, along with those from other gatekeepers, starting March 7, 2024. 
  • Apple may be required to modify its offering based on this review.
  • Aside from Epic, who have a well-documented ‘beef’ with Apple, we haven’t seen many developers showing enthusiasm about the chance to take Apple up on their new terms.  
  • The proposed new terms are not available to opt into yet: this is all still somewhat theoretical as of today (Jan 31, 2024).
• It’s not a one-size-fits-all situation: Factors such as reach, development effort, revenue model, user base size, and other fees/charges incurred all need to be considered when deciding which model is best for your app.
• Phiture expects other gatekeepers such as Microsoft – and possibly Amazon – to mull creating their own iOS app stores, which could significantly boost the reach of apps in the EU that opt into Apple’s new business terms.
• Some niche-audience apps may be able to operate very effective businesses while staying under one million EU installs, but the App Store Small Business Program already offers a reduced 15% fee, which only drops to 13% under the new terms. However, this ties developers into the €0.50 Core Technology Fee once they hit one million EU installs: this fee will be a killer for most existing app business models.
• Under the current rules, app updates (including updates for users who are inactive but still have the app installed) will be counted toward the ‘installs’ total. This implies that developers will have to pay the CTF annually even for inactive users. The CTF makes the economics very tough for most businesses.

Industry reaction

• Consumer concerns: There’s a mix of opinions among consumers. Some appreciate the EU’s stance on consumer rights while others express concerns about potential security risks and reduced user experience due to multiple app sources.
• Regulatory compliance questions: There are uncertainties about whether Apple’s restructuring will comply with the DMA as the act requires fair, reasonable, and non-discriminatory conditions for access to app stores. The introduction of the CTF and Apple’s new business terms are central to this debate.

Further Reading

• Apple’s official announcement and business terms changes link
• Mobile-centric analysis from renowned industry commentator Eric Seufert link
• Mobile-centric analysis of Apple’s DMA response from the leading mobile measurement platform, Appsflyer link
• Explanation of the DMA by CookieBot (UserCentrics sub-brand) – less mobile-centric link 
• Reuters’ coverage of Apple’s response to the DMA and industry reactions link
• 9to5Mac’s discussion on the new labeling of apps to inform and protect EU users link
• MacRumors’ report on industry opinions and the App Store split link
• TechCrunch’s analysis of Apple’s new ‘core tech’ fee and its implications link
• MacRumours report on Apple’s recent non-compliance with EU rulings (related to Epic Games) requiring Apple to open up distribution to non-Apple store purchases link
• Epic has declared they will launch an app store on iOS, mostly to distribute Fortnite, despite decrying Apple’s “Hot Garbage” EU business rules link
• Developers call out Apple’s ‘malicious compliance’ link