Amazon Filed a Patent in 1998 That Described Exactly How 'Customers Who Bought This Also Bought' Works
In Patent Archaeology #2 (Nikola Tesla, US381968), a motor patent from 1888 turned out to contain the design logic of modern EV powertrains. This time I'm digging somewhere closer to daily life.
The recommendation box that appears on every e-commerce page — "Customers who bought this also bought" — has a design document. It's a 1998 Amazon patent.
Conclusion First
Patent number: US6266649B1 Title: "Collaborative recommendations using item-to-item similarity mappings" Filed: September 18, 1998 Issued: July 24, 2001 Inventors: Gregory D. Linden, Jennifer A. Jacobi, Eric A. Benson Original Assignee: Amazon.com Inc Current Assignee: Amazon Technologies Inc Status on Google Patents: Expired - Lifetime (Anticipated expiration: 2018-09-18) Claims: 49 Litigation family: exists
Note: Legal status reflects Google Patents display only. For actual freedom-to-use analysis, verify at USPTO Patent Center.
Separating strong analogy from technical identity.
This patent is not modern AI recommendation or embedding-based search. Item-to-item collaborative filtering and dense vector nearest-neighbor search operate on fundamentally different computational principles. But "precompute similarity, serve by lookup" is a design orientation that recurs in modern large-scale recommendation infrastructure.
The core idea is simple: stop comparing every user to every other user every time a recommendation is needed. Instead, precompute which items are similar to which other items, store those tables, and at recommendation time, just look up and combine.
1. How I Found It
The patent appeared in the AI & ML Patent candidate database under "highest modern relevance score / confirmed primary source." Amazon's design for EC recommendation is documented at the primary source level on Google Patents. Abstract and mechanism description confirmed; Claim 1 verbatim text not retrieved (JavaScript rendering limitation on the Claims page).
[STEP 1] Query: "Amazon collaborative filtering patent 1998 item-to-item USPTO"
[STEP 2] Identified US6266649B1 via Google Patents
[STEP 3] Retrieved abstract and mechanism description; Claim 1 full text unconfirmed
[STEP 4] Selected: design is documented at primary source level;
clear connection to modern recommendation system design problems
2. What the Patent Describes
From the abstract on Google Patents:
A recommendation service generates individual recommendations using a customer's purchase and browsing history. It uses a pre-generated table to map items to lists of similar items, retrieves and combines those lists from a customer's known interests, and produces recommendations. The table is updated periodically by an offline process.
Three technical decisions drive this design:
1. Precompute item-to-item similarity tables offline Similarity is derived from co-occurrence in user behavior: which items appear together in purchases, views, or lists across the customer base.
2. Keep the online computation cheap At recommendation time, the system looks up a table and combines lists. It does not recompute similarity on the fly. Latency stays low regardless of user base size.
3. Compare items to items, not users to users Classic "user-user" collaborative filtering compares every pair of users — computation that scales with user count squared. Item counts tend to be more stable. Item-to-item comparison scales better.
3. Then vs. Now
| US6266649B1 (1998) | Modern recommendation systems | Assessment |
|---|---|---|
| Offline item-to-item similarity table | Batch-built item embedding index | Similar (shared problem orientation; implementation differs) |
| Co-occurrence from purchase history | Collaborative filtering trained on behavior logs | Similar |
| Lookup + list combination for real-time serving | Low-latency ANN (approximate nearest neighbor) serving | Analogy (speed motivation overlaps; mechanisms differ) |
| Item-to-item mapping | Dense vector embedding space similarity | Does not map well (design is fundamentally different) |
On rows 3 and 4. The aspiration to return recommendations cheaply is shared. But the 1998 patent builds sparse co-occurrence tables from behavioral counts. Modern embedding-based systems build dense vector representations through neural training. The computational substrate is different. "They solved the same speed problem" is accurate; "they're the same system" is not.
4. Why This Patent Is Not Often Cited
Linden, Mooney, and Good published an explanatory paper in IEEE Internet Computing in 2003 (often cited as "Linden et al. 2003"), making the item-to-item collaborative filtering concept widely known to engineers without reading the patent. Once the concept entered the literature, the patent itself became less referenced — even though the patent predates and grounds the paper.
This is a hypothesis. Internal records of how the paper was written relative to the patent have not been confirmed.
5. Why This Is Worth Reading
Recommendation in 2026 involves multi-tower retrieval, real-time feature computation, and large language model re-ranking. The 1998 patent is none of that. What it documents is the shift from "recalculate everything online" to "precompute what you can, serve from a table." That shift is a design principle that recurs, not a technique that was inherited.
Reading the patent gives you the design reasoning that the 2003 paper assumes you already know.
6. Pitfalls
Pitfall 1: Amazon did not invent recommendation
GroupLens (1994) had already implemented user-user collaborative filtering. Amazon's contribution was industrializing item-to-item recommendation for large-scale e-commerce. Not the same as inventing recommendation.
Pitfall 2: Item-to-item tables are not embedding vectors
Dense embedding similarity (used in two-tower models, DPR, and similar systems) has a different computational basis from co-occurrence tables. The direction — "find items similar to the user's interest" — is shared. The mechanics are separate.
Pitfall 3: Google Patents "Expired" is not a legal opinion
A litigation family exists for this patent. For any commercial use decision, USPTO Patent Center verification is required.
Strictly Speaking
Confirmed from primary source Google Patents: US6266649B1 / Filed 1998-09-18 / Issued 2001-07-24 / Three inventors (Linden, Jacobi, Benson) / Assignee Amazon Technologies Inc / Expired-Lifetime (anticipated 2018-09-18) / 49 claims / litigation family exists / Abstract and mechanism description confirmed
Author's interpretation "Important precursor to modern recommendation infrastructure design" is an interpretive claim. Technical inheritance from this patent to modern embedding-based systems has not been confirmed at the primary source level.
Analogy and metaphor Row 3 of the table (real-time serving speed) is an analogy. Row 4 (item mapping vs. embedding space) was assessed as "does not map well" — the designs are architecturally incompatible.
Unconfirmed Claim 1 verbatim text / Technical comparison with 2003 IEEE paper / Litigation family outcomes / Relationship to GroupLens and other 1994 prior art / Comparison with subsequent embedding-based approaches
Where the comparison breaks down The item-to-item table is a sparse co-occurrence structure. Modern embedding search is dense vector retrieval. The connection between them is at the problem orientation level — "find similar items efficiently" — not at the implementation level. No primary source evidence of technical inheritance has been confirmed.
Sources:
- Primary patent: US6266649B1 on Google Patents
- AI & ML Patent #2 (full note): LeCun CNN weight-sharing US5067164A (1989)
- Patent Archaeology #1: IBM ZISC US5717832 (1995)
- Patent Archaeology #2: Nikola Tesla US381968 (1888)