A large online catalogue solves one problem while creating another. It makes many products accessible, but leaves visitors with the work of finding what suits them. Amazon’s early recommender systems addressed that second problem. The important commercial question was not simply how much a retailer could offer. It was which small selection should appear before a particular person next.
There is a specific milestone associated with 1998. In their technical retrospective, Brent Smith and Greg Linden dated Amazon’s introduction of item-based collaborative filtering to that year. The well-known explanation by Greg Linden, Brent Smith and Jeremy York appeared in 2003. Deployment and publication are connected events, but they are not the same date (Smith and Linden, 2017; Linden, Smith and York, 2003).
This is a more precise history than saying that Amazon invented artificial intelligence or personalized recommendations in 1998. Earlier projects had already investigated how people’s experiences could help others select information. Amazon’s contribution lay in a consequential application to a large retail operation. It connected calculation and interface design with a commercial task: helping customers discover relevant products.
What 1998 meant, and what already existed
Amazon’s report for the 1997 financial year already referred to instant recommendations, collaborative filtering and other personalization features. It also described a gift centre introduced in November 1997 with editorial and personalized suggestions. Recommendations therefore did not begin at Amazon in 1998, although the report does not reveal every technical detail of the earlier arrangements (Amazon.com, 1998).
The distinction matters because personalization is a broad category. Greeting a returning visitor by name, changing displayed content and calculating a product selection are different activities. They do not necessarily depend on the same procedure. An editor’s selection is also different from a relationship inferred from behavioural data.
A careful timeline consequently treats 1998 as a milestone for a particular approach. It should neither absorb all earlier activity nor make that activity disappear. Company reports, technical publications and later recollections illuminate different parts of the development. Their dates cannot be substituted for one another without changing the claim.
Commercial improvement does not require inventing every visible feature again. An existing function can become faster, more suitable or more practical at scale. The historical advance may lie in organizing a familiar task more effectively. That is one reason the exact technical contribution matters more than a generic label of innovation.
The background included electronic documents, not just shopping
David Goldberg, David Nichols, Brian Oki and Douglas Terry published their Tapestry paper in 1992. Developed at Xerox’s Palo Alto Research Center, the system belongs to the early history of collaborative information filtering. Its setting involved electronic documents and human judgements rather than an Amazon shopping basket (Goldberg et al., 1992).
The origin is revealing. Recommendation does not have to begin with a transaction. It addresses the relationship between an abundance of alternatives and limited human attention. Other people’s experience can help an individual choose without examining every possibility. Retailing placed that selection problem within a commercial setting.
The distinction from straightforward text retrieval is also useful. A search can ask which documents contain particular terms. A recommendation can ask which material may fit earlier judgements or interests. The tasks can support one another without relying on identical signals.
The history should therefore include communication and information research alongside the famous commercial application. A successful retail system did not emerge from an intellectual vacuum. Marketing history becomes more explanatory when it identifies the earlier problems and research communities involved, rather than treating the best-known business as the origin of every underlying idea.
GroupLens and Ringo made preference relationships computational
The 1994 GroupLens paper described predicted preferences for Netnews articles based on readers’ ratings. Paul Resnick and his co-authors presented an open architecture in which rating servers gathered and distributed judgements. Earlier agreement was used to help anticipate later interests. This was a concrete application outside online retail (Resnick et al., 1994).
In 1995, Upendra Shardanand and Pattie Maes presented Ringo, a system recommending music artists and albums. Their paper discussed use by more than 2,000 people and comparisons among several algorithms. These different domains establish a research history before Amazon’s item-based milestone (Shardanand and Maes, 1995).
“Collaborative” did not mean that participants had to meet and compose a recommendation together. Their contributions could be combined within a procedure. The cooperation was organized through recorded judgements, and people did not need personal relationships with one another for the system to operate.
That produced a different kind of mediation. A suggestion could depend on people the recipient had never met. Rules were needed to determine which similarities mattered and how evidence should be combined. The software did not create human preferences from nothing. It changed how traces of those preferences could be aggregated and passed on.
Word of mouth became a reference point, not an exact duplicate
Paul Resnick and Hal Varian’s 1997 introduction situated recommender systems alongside the ordinary practice of consulting others when personal experience is insufficient. They favoured the broader term “recommender system” because this assistance need not always involve the same collaborative-filtering process (Resnick and Varian, 1997).
The analogy explains some of the attraction. Nobody can read every book or try every available product. Guidance can reduce the effort of choosing. Computational mediation nevertheless changes the basis of confidence. A friend’s advice comes through a recognizable relationship; an algorithmic suggestion may draw on many relationships the recipient cannot inspect.
This raises a marketing question: why should somebody regard the output as useful? Its statistical basis might be reasonable without being obvious to the visitor. The interface influences how the suggestion is understood, not merely whether it is visible.
An unfamiliar recommendation accompanied by an intelligible connection may be interpreted differently from the same item presented without context. Translating calculation into something a person can assess is consequently part of the system’s achievement. It is not simply decorative packaging placed around a completed mathematical result.
Different approaches answered different questions
Marko Balabanović and Yoav Shoham’s 1997 Fab paper combined content-based and collaborative recommendation. Its existence shows that early research was not confined to one approach. Information about the material itself and evidence from its users could be brought together in different ways (Balabanović and Shoham, 1997).
A simple illustration separates the possibilities. A content-based method might connect two books because they discuss a similar subject. A collaborative method might connect them through patterns of selection or appreciation. The relationships need not coincide. Two items can look different in vocabulary or category while appealing to overlapping interests.
A bestseller list answers another question. It says something about general popularity rather than necessarily identifying what is appropriate for a particular person. An editorial selection can follow an explicitly chosen theme. None of these arrangements is inherently useless; each has different informational requirements and potential applications.
Historical explanation should therefore go beyond calling a system intelligent. What information enters it? Which relationship is calculated? How does that relationship shape what is shown? These questions identify the actual change more clearly than the label alone.
The item-based approach shifted the comparison towards products
Badrul Sarwar, George Karypis, Joseph Konstan and John Riedl examined item-based techniques in 2001, including similarity measures and ways of generating recommendations. They addressed quality and computational performance together. The work illustrates a broader technical investigation of relationships between items rather than only relationships between people (Sarwar et al., 2001).
The basic difference can be explained without a formula. A user-based approach asks which other people resemble a particular person. An item-based approach first identifies suitable relationships among products. Individual suggestions can then be assembled from the items associated with that person’s interests or activity.
An explicitly invented example is a reader who has bought an introductory photography book. Relationships associated with that item might help identify another book about composition. The connection need not result from identical words in the titles. It could arise from observed patterns of interest. This illustrates a principle rather than reconstructing an Amazon dataset.
The shift matters commercially because part of the work can be prepared before a live visit. Calculation, freshness and quality become design variables. A theoretically interesting suggestion is less useful to a functioning shop if it arrives only after an impractical delay. The retailer needs a procedure that operates within the visitor’s activity.
Scale also meant being ready at the right moment
Smith and Linden’s 2017 account explained the separation between prepared item relationships and rapid selection during use. It connected the approach with scale, responsiveness and intelligible links between suggestions and prior interests. These are retrospective statements by involved specialists, not an independent audit of every earlier feature (Smith and Linden, 2017).
The architecture reveals an easily overlooked condition of marketing usefulness. The result must arrive within the time available for a decision. A person browsing a store cannot be assumed to wait indefinitely for a marginally better list. Timely assistance performs a different role from an elaborate analysis completed after the visit.
Scale has several dimensions. More customers, more products and more current interactions do not necessarily increase the same part of the workload. Improving one bottleneck does not make all processing free. The achievement concerns a better allocation of work, rather than the disappearance of computational costs.
Claims about an efficient online stage should therefore not be read as claims that preparation, storage and maintenance require no resources. A recommender is an operational arrangement with several stages. The short visible list is an output of that arrangement, not a complete picture of its cost or complexity.
A patent supplies another dated record
US Patent 6,266,649, “Collaborative recommendations using item-to-item similarity mappings,” names Gregory Linden, Jennifer Jacobi and Eric Benson as inventors. It was filed on 18 September 1998 and granted on 24 July 2001. The record provides an additional dated source for technical work on the approach (Linden, Jacobi and Benson, 2001).
The names also show why a history should not be reduced to the authorship of one later article. Different records can make different contributions visible. Publishing a paper does not establish that only its authors participated in every earlier stage of development.
A patent does not, however, determine the exact date of first operational deployment or the commercial effect of every described feature. It documents a procedure and a patenting event. Claims about implementation, public explanation and business outcomes still require evidence appropriate to each question.
This is historical source criticism rather than an assessment of present-day patent rights. It permits a specific milestone to be retained without turning it into a universal claim of invention. Such distinctions can be lost when a complex development is compressed into one line on a timeline.
Recommendation became a marketing task inside the store
J. Ben Schafer, Joseph Konstan and John Riedl’s 1999 study examined recommender systems across several e-commerce sites. It related recommendations to finding products, additional selling opportunities and customer relationships, while distinguishing interfaces, technologies and inputs. Recommendations were already being analysed as business tools rather than only unusual computing experiments (Schafer, Konstan and Riedl, 1999).
The marketing task was partly relocated within the offer. Not every commercial influence had to originate in an external advertisement. Choosing which products appeared in the shop could itself shape which possibilities a visitor considered. Selection and ordering became consequential decisions.
This did not abolish assortment management. A retailer still had to determine availability and provide useful product descriptions. Recommendation added another layer by selecting a smaller visible set from the larger stock of possibilities. That set did not have to be identical for everyone.
The relationship between product and communication also changed. An effective selection aid could simultaneously be a service and a route to further sales. Customer usefulness and the retailer’s commercial interest could align, but alignment was not guaranteed merely because the procedure was personalized. It depended on what the system actually helped the person do.
The 2003 account described selection, not generative AI
Linden, Smith and York’s 2003 paper discussed recommendations on web pages, in shopping baskets and in email campaigns. It reported higher click-through and conversion than untargeted content, but did not provide a publicly reproducible, isolated return-on-investment result for 1998 (Linden, Smith and York, 2003).
The system should therefore not be equated with present-day text generators. Selecting and ranking existing products does not require generating new product descriptions. “AI in marketing” refers here to a field of data-based inference and selection. The phrase should not imply capabilities that the historical application did not demonstrate.
The data description needs similar care. The possibility of using different signals does not prove that every early implementation processed every search, click and purchase in the same way. A historical account should identify documented inputs rather than making a blanket claim of complete behavioural knowledge.
The narrower explanation remains consequential. The innovation did not require knowing everything about every person. It involved using particular observed relationships to create a useful choice. Its effectiveness depended on the quality of that relationship, rather than on the breadth of the language used to describe artificial intelligence.
Observed behaviour was evidence, not a complete preference profile
A purchase might be intended for the buyer or selected as a gift. Viewing an item can indicate enthusiasm, comparison or curiosity. A missing rating need not express dislike. These examples expose an interpretive problem: the recorded event is not identical to the interest behind it.
More data does not necessarily eliminate every such ambiguity. People change interests and may purchase on behalf of several household members. A large dataset remains a collection of observable actions. The system must decide which actions to emphasize and how strongly.
Personalization can obscure this limitation. An interface feels individual because it responds to a recognizable history. That does not mean the system comprehensively understands the person. A limited predictive procedure can be useful without being a full account of human taste.
The experience also depends on how mistakes are encountered. An unsuitable suggestion may be accepted as an occasional error or interpreted as evidence of general unreliability. Technical output and perceived quality are separate, connected dimensions. Marketing outcomes depend partly on that human interpretation.
New items created a problem that calculation alone could not remove
Andrew Schein, Alexandrin Popescul, Lyle Ungar and David Pennock’s 2002 paper examined cold-start recommendations for items nobody in the community had yet rated. It combined content information with collaborative evidence. The study addressed a limit that a purely historical-rating approach could not solve simply by running faster (Schein et al., 2002).
A new book can fit somebody’s interests while lacking a recorded pattern connecting it with other titles. Commercial availability and recommendation evidence can therefore develop at different times. The gap is also a merchandising problem: entering the catalogue does not automatically make an item easy to recommend through every method.
New visitors present a related but distinct challenge. Without a recognizable history, there is less evidence for an individual selection. General starting suggestions, explicitly stated interests or current context are possible information sources. Whether a particular early implementation used them requires its own evidence; these possibilities do not establish additional Amazon features in 1998.
The dependency among assortment, audience and data becomes visible here. More products expand choice while initially lacking the relationships a procedure needs. More visitors may enrich evidence over time while still needing help on arrival. Personalization was not one switch that produced the same quality in every situation after it had been turned on.
Explanation was already a research question in 2000
Herlocker, Konstan and Riedl’s 2000 research investigated how and why collaborative-filtering recommendations should be explained. Their experiments provided evidence that explanations could improve acceptance. The study concerned the presentation of recommendations, not a measured Amazon revenue effect (Herlocker, Konstan and Riedl, 2000).
The issue matters commercially. An unfamiliar item may look inappropriate until a connection with an earlier selection becomes understandable. Explanation need not alter the computed order to change how the result is assessed. It supplies additional information for interpretation.
Greater acceptance is not automatically a better decision, however. A persuasive explanation can also make a weak suggestion attractive. Evaluating the interface therefore requires more than asking whether it encourages clicking. It also matters whether people can understand the basis of the recommendation. This is an analytical implication, not an additional reported result of the study.
The early research shows why computation and judgement cannot be separated completely. A calculated output still has to be translated into a usable interaction. The history of recommender systems includes that translation as well as the development of algorithms.
General popularity was different from distinctive relevance
Larry Hardesty’s 2019 Amazon Science retrospective discussed why frequent joint purchasing does not automatically demonstrate a particularly strong relationship. Very popular items appear in many histories. It also described later corrections concerning unusually active purchasers. The account reveals continuing work on the meaning of statistical association (Hardesty, 2019).
A hypothetical bestseller illustrates the point. If it appears in many orders, it will also occur alongside numerous specialist titles. That alone does not make it distinctive for every specialist audience. General frequency and an unusually strong connection are different statements.
The choice of measure affects what becomes visible. A procedure can follow its rule correctly while repeatedly foregrounding familiar products. Its suitability must therefore be judged against the intended task, rather than inferred from the mere fact that it operates mathematically.
For marketing history, this is more than a technical detail. Recommendation distributes attention within an assortment. How it handles popularity can help determine whether it supports discovery or reinforces existing visibility. The computational rule participates in the organization of demand.
Better prediction was not automatically greater business value
Jonathan Herlocker and his co-authors’ 2004 review distinguished tasks, datasets, metrics and user evaluations. It showed that recommender systems had been assessed in many, often difficult-to-compare ways. A single accuracy score could not express every relevant property (Herlocker et al., 2004).
A retrospective test might ask whether known preferences can be predicted. An operating retailer must also ask whether the displayed suggestions help at that moment. An accurately predicted item may already be familiar. A less predictable suggestion might provide useful discovery. The value of the property depends on the task.
Dietmar Jannach and Michael Jugovac’s 2019 review addressed business value directly. It identified continuing questions about quantifying real-world effects and about translating algorithmic improvements into commercial outcomes. The distinction between an improved offline test and demonstrated business value remains essential (Jannach and Jugovac, 2019).
For the 1998 milestone, deployment can be documented, as can later technical significance. A specific assertion about additional profit still needs a separate demonstration. Neither the company’s prominence nor an attractive algorithm substitutes for that evidence.
Click-through, conversion and return on investment asked different questions
A click-through rate relates interactions to impressions. A conversion rate relates defined outcomes to a specified starting population. Financial return also depends on costs, margins and outcomes that actually occurred because of the intervention. Without those distinctions, a precise percentage can support an imprecise historical narrative.
Consider an explicitly invented comparison. Under comparable conditions, one variant produces four purchases per hundred visits and another five. That is initially one additional purchase in the group, not proof of an equivalent gain in profit. Basket value, returns, displacement of other buying and operating costs still matter. These are illustrative numbers, not Amazon results.
A purchase after a recommendation click is not necessarily caused wholly by the recommendation. A customer may already have intended to acquire the product or might have found it through another route. Estimating the additional effect requires a suitable comparison. The observed sequence of events is insufficient by itself.
This is why a blanket statement that recommendations proved higher marketing ROI in 1998 goes beyond the evidence unless an appropriate study is identified. A careful claim remains useful: a powerful form of personalized selection developed, and its different effects require different kinds of measurement.
The historical issue is not whether the technology could be commercially useful. It is whether a particular outcome, magnitude and date have actually been established. Maintaining that distinction avoids turning a credible innovation into an unsupported success statistic.
Personal discovery did not guarantee greater market-wide diversity
Daniel Fleder and Kartik Hosanagar’s 2009 study used analytical modelling and simulation to investigate sales diversity. It showed how individual discovery and the distribution of aggregate demand could diverge. Under particular conditions, recommendations could broaden personal consumption while concentrating overall sales (Fleder and Hosanagar, 2009).
The distinction is intuitive. Many people can each discover something new to them while discovering the same already popular product. Their personal experience has expanded, but demand across the market has not necessarily become more evenly distributed. Individual novelty and collective diversity operate at different levels.
The finding does not prove that Amazon’s 1998 system displaced every niche. It also prevents the opposite automatic conclusion that recommendations must always promote the entire breadth of a catalogue. Actual effects depend on the procedure, audience and initial distribution.
For international marketing history, this leaves an empirical question. A method may operate differently across assortments and patterns of use. A prominent American application does not supply a complete description of every subsequent deployment around the world.
What remains of Amazon’s milestone
The year 1998 is a useful reference point for Amazon’s introduction of item-based collaborative filtering, not for the beginning of all personalization. The background includes Tapestry, GroupLens, Ringo and hybrid approaches. The company report records earlier Amazon features, the patent documents 1998 technical work and the 2003 article explains a later publicly described implementation.
For marketing, the significant change was that selection within a shop could become an ongoing computational service. A large catalogue did not have to be presented identically to everyone. Recorded evidence could help identify possibilities that a visitor might otherwise overlook, making presentation part of the offer itself.
The significance does not require an unsupported revenue percentage. Data could influence what appeared next, rather than only describing past activity. A computational procedure stood between human interests and a commercial interface. Its relevance, usefulness and financial contribution remained questions requiring concrete evidence.
References
Amazon.com (1998) Form 10-K for the year ended December 31, 1997, especially ‘Recommendations and Personalization’. https://s2.q4cdn.com/299287126/files/doc_financials/annual/123197_10k.pdf.
Balabanović, M. and Shoham, Y. (1997) ‘Fab: content-based, collaborative recommendation’, Communications of the ACM, 40(3), pp. 66–72. https://doi.org/10.1145/245108.245124.
Fleder, D. and Hosanagar, K. (2009) ‘Blockbuster Culture’s Next Rise or Fall: The Impact of Recommender Systems on Sales Diversity’, Management Science, 55(5), pp. 697–712. https://doi.org/10.1287/mnsc.1080.0974.
Goldberg, D., Nichols, D., Oki, B.M. and Terry, D. (1992) ‘Using collaborative filtering to weave an information tapestry’, Communications of the ACM, 35(12), pp. 61–70. https://doi.org/10.1145/138859.138867.
Hardesty, L. (2019) ‘The history of Amazon’s recommendation algorithm’, Amazon Science, 22 November. https://www.amazon.science/the-history-of-amazons-recommendation-algorithm.
Herlocker, J.L., Konstan, J.A. and Riedl, J. (2000) ‘Explaining collaborative filtering recommendations’, Proceedings of CSCW ’00, pp. 241–250. https://doi.org/10.1145/358916.358995.
Herlocker, J.L., Konstan, J.A., Terveen, L.G. and Riedl, J.T. (2004) ‘Evaluating collaborative filtering recommender systems’, ACM Transactions on Information Systems, 22(1), pp. 5–53. https://doi.org/10.1145/963770.963772.
Jannach, D. and Jugovac, M. (2019) ‘Measuring the Business Value of Recommender Systems’, ACM Transactions on Management Information Systems, 10(4), Article 16, pp. 1–23. https://doi.org/10.1145/3370082. Author version: https://arxiv.org/abs/1908.08328.
Linden, G.D., Jacobi, J.A. and Benson, E.A. (2001) Collaborative recommendations using item-to-item similarity mappings, US Patent 6,266,649, filed 18 September 1998, granted 24 July 2001. https://patents.google.com/patent/US6266649B1/en.
Linden, G., Smith, B. and York, J. (2003) ‘Amazon.com recommendations: item-to-item collaborative filtering’, IEEE Internet Computing, 7(1), pp. 76–80. https://doi.org/10.1109/MIC.2003.1167344. Full text: https://www.cs.umd.edu/~samir/498/Amazon-Recommendations.pdf.
Resnick, P., Iacovou, N., Suchak, M., Bergstrom, P. and Riedl, J. (1994) ‘GroupLens: an open architecture for collaborative filtering of netnews’, Proceedings of CSCW ’94, pp. 175–186. https://doi.org/10.1145/192844.192905.
Resnick, P. and Varian, H.R. (1997) ‘Recommender systems’, Communications of the ACM, 40(3), pp. 56–58. https://doi.org/10.1145/245108.245121.
Sarwar, B., Karypis, G., Konstan, J. and Riedl, J. (2001) ‘Item-based collaborative filtering recommendation algorithms’, Proceedings of WWW ’01, pp. 285–295. https://doi.org/10.1145/371920.372071.
Schafer, J.B., Konstan, J. and Riedl, J. (1999) ‘Recommender systems in e-commerce’, Proceedings of EC ’99, pp. 158–166. https://doi.org/10.1145/336992.337035.
Schein, A.I., Popescul, A., Ungar, L.H. and Pennock, D.M. (2002) ‘Methods and metrics for cold-start recommendations’, Proceedings of SIGIR ’02, pp. 253–260. https://doi.org/10.1145/564376.564421.
Shardanand, U. and Maes, P. (1995) ‘Social information filtering: algorithms for automating “word of mouth”’, Proceedings of CHI ’95, pp. 210–217. https://doi.org/10.1145/223904.223931.
Smith, B. and Linden, G. (2017) ‘Two Decades of Recommender Systems at Amazon.com’, IEEE Internet Computing, 21(3), pp. 12–18. https://doi.org/10.1109/MIC.2017.72. Full text: https://cdn.amazon.science/76/9e/7eac89c14a838746e91dde0a5e9f/two-decades-of-recommender-systems-at-amazon.pdf.
