ASO Patents

Patent number	US20240281490A1 (pending application)
Official title	Ranking of content based on implied relationships
Assignee	Apple Inc.
Inventors	Xiaoyuan Goodman Gu, Yuan Yen Tai, Moliang Zhou, Dayvid Victor Rodrigues de Oliveira, Brian Knott, Jin Cao, Jia Huang
Priority date	February 20, 2023, via provisional application 63/486,011
Filed	August 30, 2023
Published	August 22, 2024
Status	Pending. Not granted, so it confers nothing yet
Family	EP4418144A1 (European application 24155945.9), US20250258883A1 (continuation, published August 14, 2025)
Source	patents.google.com/patent/US20240281490A1

What This Patent Covers and Why It Matters for ASO

Feature it describes: how the App Store orders results for a keyword when the competing apps carry no explicit link to one another.

The patent is unusually direct about where it applies. It states that “the search input could be keywords received in an App Store user interface” and that the problem is “ranking search results for Apps in an App Store based on a user-provided keyword”. The worked examples are real app names: INSTAGRAM, SNAPCHAT, FACEBOOK, TIKTOK, WHATSAPP, MESSENGER, YOUTUBE and TWITTER. This is not a generic search document that happens to mention apps.

Every ranking model this site has covered so far works on one axis: how well an app matches the query, or how similar it is to something the user already likes. This application adds a second axis that has nothing to do with similarity. It calls it influence, and it means one app pulling a user toward another app.

Three things are named as the sources of that influence, and two of them should stop an ASO practitioner in their tracks.

• Download sequences. If a large number of users download one app and then download a second one shortly after, an edge is drawn from the first to the second. The strength of that edge decays as the gap between the two downloads grows, so a second download ten minutes later counts for far more than one a week later.
• Campaign keyword targeting. The document says content publishers “can indicate that their content is relevant to specified keyword targets that can reference other content”, and that this creates an edge in the same graph. In plain terms, the keywords advertisers choose to bid on are treated as a declaration of relationship between apps, and that declaration feeds the ranking graph.
• Review text. When the reviews on one app mention a second app by name, an influence edge runs from the first to the second.

The second one is the finding. Search Ads targeting data and organic search ranking are usually discussed as separate systems, and Apple has always described them that way publicly. This application describes advertiser keyword targeting as an input to the graph that orders search results. It is an application and not a granted claim, so it is evidence of what a team designed rather than proof of what ships, but it is Apple’s own document saying it.

The third one matters too. Reviews have always been discussed in ASO as a conversion factor and a ratings signal. Here the actual words inside reviews are read for the names of other apps and turned into graph structure.

The final piece is that the balance between similarity and influence is not fixed. It moves per query. The document explains it with two examples. On the keyword INSTAGRAM, “the influence relationship between apps connected to the App INSTAGRAM would be important in ranking search results”. On the keyword biology, “the influence relationship between apps in the search results is comparatively less important”. Brand queries lean on influence. Topical queries lean on similarity.

And the motive is stated outright. Without this, the patent says, “highly popular Apps such as SNAPCHAT, INSTAGRAM, FACEBOOK, TIKTOK, WHATSAPP, MESSENGER, YOUTUBE, AND TWITTER are ranked highly in the search results” even when they are not topic dependent. This is Apple writing down that raw popularity was swamping topical queries, and describing the fix.

What this changes for practitioners: your keyword ranking is not only a function of your metadata and your own performance. It is also a function of which apps users open your download session alongside, which advertisers point their targeting at you, and which apps your reviewers talk about.

Patent Summary

The system builds an entity graph in which each node is a content item, in the primary example an app. Two kinds of weight can sit on an edge between two nodes.

The first is the entity similarity weight, written esw(A, B). It is “calculated via cosine-similarity (or similar methods) between the embeddings of two entities A and B”. This is the conventional axis: two apps that describe themselves in similar ways sit close together.

The second is the entity transition weight, written etw(A to B), which the document defines as the weight “which implicitly indicates how likely the user transition from entity A to B”. It is directional. It is derived from aggregated user activity sequences with a decay function applied, so the longer the interval between the two interactions, the weaker the resulting weight. Campaign keyword targeting data and review text that references other content feed the same directional weight.

The two are combined into a single edge weight by a tunable hyperparameter p:

w(A→B, p) = esw(A, B)^p + etw(A→B)^(1-p)

The document describes p as a value that “adjusts the power of entity similarity weight and entity transition weight under different contexts”. A smaller p pushes the graph toward influence. A larger p pushes it toward similarity. The context that sets p is the search input itself, which is what produces the different treatment of INSTAGRAM and biology.

Ranking then runs on that weighted graph. The document states that “the entity graph service 106 can use the contextualized Page Rank to calculate the contextualized ranking score”. The graph is first filtered to the candidate set that matches the search criteria, and the PageRank style score is computed over that filtered, context weighted graph rather than over the whole catalog.

The independent claim is broader than the machinery. It covers identifying result content items that match a search criterion, and ranking those items using both a similarity relationship and an influence relationship between the items. The specification is where the download sequences, the campaign keyword targeting, the review mining, the decay function and the contextualized PageRank live.

One signal worth recording. The family has a European counterpart and a continuation published in August 2025, two and a half years after the priority date. Apple is still spending money on this line.