Summary
The shelf at the bottom of an App Store product page is two different things stacked together. The top of it is advertising inventory that Apple sells. Everything below is generated by a recommendation model that decides which apps are related to the one you are looking at.
Apple documents the advertising half in detail. It documents the organic half nowhere at all. A granted Apple patent describes it, and the mechanism is not what most ASO guides assume: relationships are not derived from category, and they are not derived from keywords alone.
Methods used
| Element | Method | Where it is stated |
|---|---|---|
| Top slot | Paid placement, shown to users who scroll to the bottom of a product page, built from your app name, icon and subtitle, targeted by app category | Documented |
| Organic slots | Item to item relationship model, queried with the current app as the seed | Patent |
| Relationship types | Overlapping, meaning items serving a similar primary purpose, and complementary, meaning items that extend rather than replace | Patent |
| Explicit signal | Co-download metrics across user populations | Patent |
| Implicit signal | Embedding similarity derived from content tags, descriptions and topic modeling | Patent |
| Training | Online meta-learning from user interactions, including observed non-engagement | Patent |
| A second, separate similarity graph | For search results, similarity built from overlap in the functional queries that led to downloads of each app | Patent |
What was not previously known
Already documented by Apple: that ads appear at the top of the You Might Also Like list, that they are shown to users who have scrolled to the bottom of a product page, that they are assembled from assets already on your listing, that they can be targeted across app categories or refined to specific ones, and that the placement is not currently available in mainland China.
Not documented anywhere: how the organic part is built. There is no Apple help page for it and no substantive industry explanation. What follows comes from the patents.
- Apple distinguishes two kinds of related app. Overlapping items serve a similar primary purpose, so they are substitutes. Complementary items extend the experience without replacing it. These are different relationships with different commercial meaning, and the system models them separately.
- Your description and tags feed the model. The implicit half of the relationship signal is embedding similarity derived from content tags, descriptions and topic modeling. Metadata influences which apps you are grouped with, even where it does not influence keyword ranking.
- Being shown and ignored is a negative signal. The model trains online from user interactions, and the patent specifies that it learns from observed non-engagement as well as from engagement. Appearing on a shelf and getting no taps is training data arguing against you.
- The claim names the surface explicitly. It covers receiving a request for a seed application’s landing page and presenting candidate applications identified through detected relationships. This is not a general recommender being repurposed. It is written for the product page.
- Apple has at least two different definitions of “similar app”. The product page model works on co-downloads and embeddings. A separate patent builds a different similarity graph for search results, using overlap in the functional search queries that led to downloads of each app. Same phrase, different mathematics, different surfaces.
- Category is not in the mechanism. Neither patent derives similarity from the App Store category you selected. That assumption is one of the most common in ASO writing and neither document supports it.
How the shelf is filled
The top of the list is for sale
Start with the part nobody disputes, because it changes how you read everything else. Apple’s advertising help states that product page ads appear “at the top of the You Might Also Like list to users who have scrolled to the bottom of product pages.”
Three details matter for anyone studying the shelf:
- The ad is at the top of the list, not beside it. The first thing in the section is the slot Apple sold.
- It is triggered by scroll depth. The user has already read your page and kept going, which is the moment they are most open to an alternative.
- Targeting is by app category, which means the advertiser chooses the neighbourhood. The organic apps below it did not choose anything.
If you are studying which apps appear under yours, separate the paid slot from the organic ones before drawing conclusions. They are produced by completely different systems.
The organic slots come from a relationship model
Below the ad, the apps are produced by an item to item relationship model. Your app is the seed, and the system returns candidates it has already determined are related to it.
The relationships are precomputed rather than decided in the moment, and the model runs continuously and refines itself. So the shelf under your listing is a readout of a stored relationship graph. It changes when the graph changes, not when you change your page.
Two signals build the graph
The explicit signal is co-downloads. Which apps get installed by the same people, measured across user populations. This is behavioral and you cannot write your way into it.
The implicit signal is embedding similarity, derived from content tags, descriptions and topic modeling. This is textual, and it is the half you control. Two apps that describe the same task with the same vocabulary land near each other whether or not anyone has installed both.
A composite model combines them. That combination is the practical opening for a new app: with no install base there is no co-download signal at all, so early placement on anyone’s shelf has to come from the text side.
Substitute or companion
The overlapping versus complementary distinction is the most strategically useful idea in the patent, because it changes what a given shelf means.
| Relationship | What it means | What a user taking it costs you |
|---|---|---|
| Overlapping | Serves a similar primary purpose. A substitute. | A lost install. This is a competitor sitting on your page. |
| Complementary | Extends the experience without replacing it. | Nothing. The user can have both. |
Look at your own shelf and sort each app into one of these two columns. A page surrounded by substitutes is a different competitive position from one surrounded by companions, and the same number of apps means opposite things in the two cases.
The model watches what happens next
Online meta-learning means the model updates from what users do with the recommendations it produced. The patent is specific that this includes observed non-engagement.
So placement is not a prize, it is a test. If you appear on a relevant app’s shelf and nobody taps, the system learns that the relationship it inferred was weaker than it thought. Impressions without taps are not free exposure in this design.
The practical consequence is that your icon and subtitle are doing real work, because that is all the shelf shows. A listing that cannot earn a tap from a genuinely relevant audience will gradually stop being shown to it.
How it works
Step 1: every item gets a representation
The system covers applications, digital media such as music and video, electronic games and e-books. Each item carries a text derived representation built from its tags, description and topic modeling, and a behavioral profile built from who downloads it.
Step 2: relationships are detected and typed
Pairs of items are assessed on both signals, and a relationship is recorded along with its type, overlapping or complementary. This is the stored graph the shelf reads from.
Step 3: someone opens a product page
A request arrives for a seed application’s landing page. The platform identifies candidate applications through the detected relationships and selects which ones to present in the interface.
Step 4: the ad is inserted above them
If the user scrolls far enough, a paid placement occupies the top of the list, drawn from a campaign that targeted the category.
Step 5: the outcome feeds back
Engagement and non-engagement with what was shown returns to the model, adjusting the weighting of signals over time.
Step 6: search has its own, different shelf
Worth keeping separate in your head. When someone searches an app by name, the results page also contains similar apps, but those are chosen by a different method: overlap in the functional search queries that led to downloads of each app, scored by how large a proportion of an app’s downloads came from those queries.
Two systems, two definitions of similar, two sets of apps. If the apps under your product page differ from the apps beside you in search results, that is expected behavior rather than an anomaly.
What this means for how you work
- Audit your own shelf monthly, separating the paid top slot from the organic apps, and label each organic entry as a substitute or a companion. That list is the clearest available picture of who the store thinks you compete with.
- Use vocabulary deliberately. The text half of the relationship graph is built from tags, descriptions and topic modeling. Describing your task the way the apps you want to sit beside describe theirs is a real lever.
- New apps compete on text first. Without installs there is no co-download signal, so metadata is the only route onto anyone’s shelf early on.
- Treat icon and subtitle as shelf assets. They are all that appears in the recommendation, and non-engagement is recorded.
- Do not assume category drives it. Nothing in the patents supports that.
- If you want the top slot, buy it. The organic positions cannot be purchased and the paid one cannot be earned.
Sources and caveats
- Apple Ads, Product Pages placement help. The paid slot at the top of the list, the scroll trigger, the creative assembly and the category targeting.
- US12314272B2, Online meta-learning for scalable item-to-item relationships. The relationship model, the overlapping and complementary types, the two signals and the online training. Granted May 2025, active.
- US10394838B2, App store searching. The separate similarity graph used in search results, built from functional query overlap. Granted 2019, active.
Caveats. A patent describes what a company designed and claimed, which is strong evidence and not documentation. Apple has never confirmed that either model powers the shelf you see, and the only part of this section Apple describes publicly is the advertising. Where a patent and the documentation disagree, the documentation describes the shipped product.
One more note for completeness. A frequently cited patent titled “Identifying similar applications”, which infers similarity from an app’s interface elements and usage, belongs to Microsoft rather than Apple. It is interesting and it is not evidence about the App Store.
Bir yanıt yazın