ASO Patents

Publication number	US20080091549A1
Official title	User supplied and refined tags
Applicant	Apple Computer, Inc., now Apple Inc.
Inventors	Kenneth N. Chang, David A. Koski, Pedraum R. Pardehpoosh
Priority and filing date	September 29, 2006
Published	April 17, 2008
Granted as	US8275666B2, September 25, 2012
Status	Expired, fee related
Source	patents.google.com/patent/US20080091549A1

What This Patent Covers and Why It Matters for ASO

Feature it describes: a community tagging layer on an online store, where customers supply the tags, other customers vote on which tags are useful, and the store uses the result as searchable metadata.

Scope note. This is a 2006 filing and its worked example tags a MacBook Pro and an iMac, so the online store in question is Apple’s retail store, not the App Store, which did not exist yet. Treat it as background on how Apple thinks about tag metadata rather than as evidence about app ranking.

It earns its place here for one reason. Tags are a named signal in Apple’s modern app profile filings, appearing in the content metadata category of US20240086412A1 alongside curation information and developer descriptions. This patent is the oldest document in Apple’s portfolio that explains what a tag is for and, more importantly, how its value gets measured.

The mechanism worth taking away is the weighting rule. A tag’s usefulness score is not just a popularity vote. The patent describes adjusting the score dynamically when users add tagged items to a shopping cart, raising the usefulness of tags that correlate with purchase.

That is the same pattern that runs through every Apple search filing on this site, twenty years apart: the words that describe an item are validated by whether they lead to a transaction. In 2006 the loop closed at the shopping cart. In 2012 it closed at the download distribution. In 2024 it closed at a query statistics vector holding conversion rates. The mechanism keeps getting more sophisticated and the principle does not change.

Practical carry over for app metadata work is limited but real:

• The patent asks users how they use an item, or how they expect others would use it. Function first vocabulary, again, four years before the topic model filing made the same choice.
• Tag suggestions propagate between similar items. If products share a tag, the other tags on one are suggested for the other. Category adjacency shapes vocabulary, which is why your competitors’ terminology tends to become the category’s terminology.
• Superlatives are filtered out as stop words, along with obscene language and proper names. “Best” was being stripped from Apple store metadata in 2006.
• Thesaurus expansion is applied to the supplied keywords, so a single tag pulls in synonyms and related terms automatically.

Patent Summary

“Managing online content via user supplied and/or refined tagging is disclosed. In some embodiments, a selection of an item available from an online store is received, and one or more tags to be associated with the item are received from a member of an online community comprising users of the online store. In some embodiments, an indication of a degree of usefulness of a tag is received from a member of an online community, and an attribute of the tag is updated to reflect the indication.”

US20080091549A1, abstract

The system has two halves. In the first, community members submit tags for store items through a web interface, prompted to describe how the item is used. In the second, community members rate the usefulness of existing tags, and those ratings update an attribute on the tag itself. Between the two sits a refinement layer: automated filtering of prohibited and low value words, thesaurus based expansion into synonyms, and suggestion of tags drawn from related items.

Tags then function as searchable metadata. A user entering keywords retrieves items whose tags match, and the tags themselves are displayed in an order and manner determined by their usefulness scores, with the most useful shown in bold.

Detailed Explanation

How tags get collected

Tags come from an interface presented to the user over the network. When an item needs tagging, the user is asked how they use it, or how they would expect others to use it, and submits through a control labeled done or submit.

The framing of that prompt is the design decision that matters. The system does not ask what the item is. It asks what it is for. A store that collects “what is this for” metadata from its customers is building a functional index, which is precisely the gap that the 2010 topic model patent would later identify as the core failure of app stores: they let you search by name when you needed to search by function.

The three refinement layers

Automated processing. Keywords are identified from the submission and filtered. Stop words and phrases are removed, and the patent names the categories explicitly: superlatives, obscene or otherwise prohibited language, and proper names. The remaining keyword set is then expanded using thesaurus resources to include synonyms and related terms. Text analysis, linguistic analysis, spell checking and thesaurus checking are all named.

Community feedback. Members vote on whether a tag is useful, and the vote updates an attribute of the tag. This is the mechanism named in the abstract’s second embodiment, and it is what makes the tag set self correcting without editorial staff reviewing every entry.

Tag suggestion from related items. The system examines the tags on similar items. Where products share a tag, the other tags on one become suggestions for the other. The patent describes this in terms of first degree relationships, meaning a common tag, and higher order relationships built from chains of those.

The worked example

A MacBook Pro carries the tags computer, portable and PowerBook. An iMac carries computer, desktop and Mac.

A user selecting the portable tag is offered three suggestions, each arriving by a different route. Laptop is suggested because it means the same thing in context, which is the thesaurus and synonym route. Mac is suggested because of a shared root and similar items, which is the related items route. And iLife is suggested because it is frequently mentioned in the associated materials, which is a co-occurrence route.

Three different justifications for three suggestions in one interface is a compact illustration of how much machinery sits behind a tag list, and the co-occurrence route in particular is the direct ancestor of the indicator terms in US9280595B2.

How tags affect search and display

Tags are searchable metadata. A user enters keywords into a search box and retrieves items whose tags match one or more of them.

Usefulness then governs presentation. The patent describes the usefulness attribute determining the order and manner in which tags are displayed, with the most useful tags shown in bold. So the same data serves two purposes: retrieval, and a visible signal to the next customer about which descriptions the community trusts.

The dynamic weighting, and why it is the important part

Figure 9 covers dynamic usefulness weighting and Figure 10 covers mining transaction data. The described behavior is that when users add tagged items to a shopping cart, the usefulness scores of the tags that correlated with those purchases increase.

This turns a subjective label into a measured one. A tag can be popular with voters and still lose ground if it does not accompany purchases, and an unglamorous tag can gain weight because the people who arrived through it bought something.

Two things follow for anyone reading Apple’s later filings. First, community input is treated as a hypothesis to be validated by behavior, never as ground truth. Second, the validation signal is commercial rather than engagement based, and that choice recurs consistently: purchases in 2006, downloads and download distributions in 2012, conversion rates in the 2024 priority filings.

Spam and abuse handling

Three defenses are described. Prohibited language is filtered during tag suggestion. Content owners are questioned as part of the process. And users can request removal of a tag, with deletion enabled when many community members have indicated that the tag is not useful.

That third one is notable for what it implies about the failure mode Apple expected. The remedy is not detecting malicious tagging automatically. It is making the community’s negative signal strong enough to remove a tag. Given what happened to open tagging systems over the following decade, it is a light defense, and the feature never appeared at scale in Apple’s stores.

What is claimed

Claim 1 is two steps long: receiving a selection of an item available from an online store, and receiving from a member of an online community of that store’s users one or more tags to be associated with the item.

That is close to the broadest possible statement of community tagging, and it granted in 2012 on a 2006 priority date, a six year prosecution that a claim this broad would explain. None of the refinement, weighting or transaction mining appears in the independent claim, which means the interesting engineering is disclosed rather than protected.

What became of it

Apple never shipped customer tagging as described here. The App Store has no user supplied tag layer, and the tag signals that do appear in Apple’s recent app profile filings are internal and editorial rather than crowd sourced. The patent expired for unpaid maintenance fees.

Google Play went further in this direction than Apple did, and the contrast is instructive. Apple’s answer to the function versus name search problem turned out to be deriving vocabulary automatically from descriptions, external coverage and behavior, rather than asking users to supply it. This patent is the road not taken, and it is worth reading precisely because the underlying question it tried to answer never went away.