# Where personalization comes from

> Personalization's history is a repeated contest over who holds the model of the person: the user, the community, or the firm. It runs from 1990s collaborative filtering through one-to-one marketing, the accuracy era, the martech explosion, the CDP era, the privacy turn, and today's real-time and generative systems.

_Last updated 2026-10-30. From The Personalization Guide, https://personalization.guide_

Personalization did not begin with the web, and it did not begin as a technology. It began as a
question about who gets to hold the model of a person. Its sixty-year history is best read as a
repeated contest over who owns that model: the user, the community, or the firm. In every era the
technical capacity to infer things about people grew faster than the social license to do so, and
each correction arrived late and left the question open.

This is not an argument that personalization is bad. The evidence shows it often helps people find
things, and some of the most famous charges against it, such as strong filter-bubble effects, are
weaker in the research than in popular memory. The argument is narrower: personalization's
legitimacy has always depended on something its builders kept treating as optional, which is the
person's knowledge of, and say over, the model built about them.

This page tells the story. For the critique of personalization and for privacy and regulation read
as ethics, see [personalization and its critics](/ethics). For how the methods themselves work, see
[how personalization works](/library/techniques).

## Who holds the model of the person

The history can be read as four shifts in who owns the model, and on what terms.

**The community model (late 1980s to mid 1990s).** The first collaborative filtering systems were
tools people used to help each other filter information overload. The model of a person was built
mostly from ratings that person chose to give, and the person knew she was being modeled because she
was the one doing the rating.

**The firm's relationship model (1993 to 2005).** One-to-one marketing reframed the customer as an
asset whose value a firm should grow, and Amazon made that commercially real at scale. The model now
belonged to the merchant, and its purpose was to sell.

**The surveillance-scale behavioral model (around 2005 to 2018).** Ad tech, social feeds, and later
customer data platforms turned implicit behavioral exhaust, such as clicks, dwell time, location, and
cross-site browsing, into the main input. The person rarely saw the model and often did not know it
existed.

**The contested and generative model (2018 to today).** Privacy law and platform policy pushed back,
while large language models made it possible to generate the personalized artifact rather than merely
select it. Who holds the model is an open question again, now with higher stakes.

Across all four, one pattern recurs. Each generation of builders defined personalization by the value
it gave the person: relevance, convenience, a store built for you. Each generation of critics defined
it by what it took: attention, autonomy, privacy, price. Both were partly right.

## The waves

### Wave 0: segmentation, direct mail, and information filtering (1950s to 1980s)

Three older traditions supplied the concepts. Market segmentation, set out in Wendell Smith's 1956
article on product differentiation, established the strategic idea of dividing a market into more
homogeneous groups and tailoring to each. Every later personalization claim is in some sense the
limit case of segmentation: shrink the segments far enough and you reach the segment of one. Direct
and database marketing, running through the 1970s and 1980s, introduced the practice of keeping
records on individual customers, scoring them, and addressing them by name. Email and CRM
personalization descend from this lineage more directly than from computer science. In parallel,
information filtering in computer science, from H. P. Luhn's selective dissemination of information
to Elaine Rich's 1979 work on user modeling, supplied the idea of a standing profile matched against
a flow of new documents. Mass customization, named by Stan Davis in 1987 and developed by Joseph
Pine in 1993, was the industrial counterpart: serve individual wants at mass-production cost.

### Wave 1: collaborative filtering and the community model (1992 to 1997)

The phrase "collaborative filtering" was introduced in 1992 by a team at Xerox PARC describing
Tapestry, an experimental system for filtering email and news with human judgments in the loop. In
1994 the GroupLens project at the University of Minnesota automated the idea, predicting how much a
Usenet reader would like an article from the ratings of readers whose past ratings correlated with
theirs. Its five stated design goals included privacy, which is worth noting for this history:
privacy was a first-wave design goal, not a later imposition. At the MIT Media Lab, Ringo recommended
music from the same principle and was spun out in 1995, becoming Firefly Network. Firefly built a
profile product, Firefly Passport, that let people control which sites could use their data, and its
people contributed to an early web privacy standard. Microsoft acquired Firefly in 1998 and, by one
contemporary account, was more aggressive about the privacy product than about the recommendation
engine. The first commercial personalization company, in other words, ended up as an identity and
privacy product. In 1997 a special section of the *Communications of the ACM* helped fix "recommender
systems" as the umbrella term. What this wave assumed, an explicit rating, a cooperative community,
and a user who knew she was being modeled, would not survive commercialization.

### Wave 2: one-to-one marketing and the firm's model (1993 to 2005)

The commercial canon was set by Don Peppers and Martha Rogers in *The One to One Future* (1993),
which argued that the basis for marketing should be share of customer rather than market share, and
that firms should identify, differentiate, interact with, and customize for each customer. The book
is widely credited with launching the CRM movement. Two details are usually forgotten: its media
examples were fax, voice mail, and interactive television rather than the web, and a later edition
included a chapter on making money by protecting privacy rather than threatening it. The founding
commercial text named the privacy tension in 1993 and framed privacy as a competitive asset.

The web of the late 1990s then used "personalization" in two senses that later merged: user-driven
customization, such as choosing modules on a portal, and system-driven inference sold by rules
engines and collaborative filtering vendors. That first personalization-software market collapsed in
the 2000 to 2001 bust. The lesson some drew, which recurs here, was that personalization software was
easy to sell and hard to make pay. The era's lasting technical contribution was Amazon's item-to-item
collaborative filtering, patented in 1998 and described publicly in 2003, which recommended neighbors
of what a shopper was viewing rather than finding similar customers. It is, ironically, one of the
least person-centric forms of personalization: it models relationships between products, and the
person enters only as a bag of recent items. Much of what users felt as "Amazon knows me" was really
"people who bought this also bought that." That gap between the felt intimacy of personalization and
its actual mechanics recurs throughout the history, and it explains both vendor overclaiming and some
misplaced fears.

### Wave 3: the Netflix Prize and the accuracy era (2006 to 2012)

In October 2006 Netflix offered one million dollars to the first team to improve the accuracy of its
rating predictor by ten percent, and released a hundred million ratings to work with. The prize was
won in September 2009 by a team that reached the target. It popularized matrix factorization and
large ensembles and made recommender systems a mainstream machine-learning subject. The more
instructive history is what happened next. Netflix later explained that it had adopted some of the
simpler winning techniques but that the final ensembles were not worth the engineering effort to
deploy, and the business had shifted to streaming, where implicit signals mattered more than star
ratings. Two lessons from the prize run through the rest of this story. A proxy metric drifts from
value: optimizing prediction accuracy was not the same as helping people choose what to watch.
And anonymization is fragile: in 2008 researchers re-identified prize records by matching them against
public ratings elsewhere, and a planned second prize was canceled after a privacy complaint.
Personalization's own benchmark became a founding case study in privacy research.

### Wave 4: behavioral advertising, feeds, and the martech explosion (around 2005 to 2017)

In this period personalization moved from the merchant's store into the whole web and the social
feed. Three shifts matter historically. The input changed from explicit ratings to implicit
behavioral exhaust, collected without a moment of user choice, through cross-site cookies, mobile
advertising identifiers, location, and social signals. The surface changed from the catalog to the
feed: Facebook's News Feed and its later algorithmic ranking, YouTube's recommendations, and ranked
timelines made personalization the default way people met news and culture, which is what gave the
filter-bubble critique its force. And personalization became a category of software purchase: the
annual marketing-technology landscape grew from about 150 tools in 2011 to more than fifteen thousand
by 2025. The cautionary tales of this wave are about inference rather than collection: a retailer
building a pregnancy-prediction score from purchases, researchers showing that social media activity
could predict private traits, and the Cambridge Analytica episode. Whether such targeting worked is
often contested; the data access usually was not. Price steering also arrived, with reporting in 2012
that one travel site showed different users different hotels.

### Wave 5: the CDP era and the single customer view (2013 to 2020)

The customer data platform emerged as a category in the mid 2010s, promising to unify first-party
data into persistent profiles that marketing tools could act on. It mattered for two reasons: it was
the enterprise's answer to its dependence on ad-tech intermediaries and third-party data, and it
re-centered the firm-owned profile that one-to-one marketing had imagined in 1993, now with the means
to act on it in near real time. This is also the era of the most-quoted statistics, and they should
be read with care. The figures that say most consumers want personalization, often given as eighty or
ninety-one percent, are stated-preference surveys from firms that sell personalization services. The
figures that say it lifts revenue ten to fifteen percent are consultancy estimates. Against them sat a
2019 analyst forecast that most marketers who had invested in personalization would abandon the effort
by 2025 for lack of return. The same industry produced both the demand statistic and the abandonment
forecast within two years.

### Wave 6: the privacy turn (2016 to 2025)

Law and platform policy pushed back on surveillance-scale personalization. The European Union's
General Data Protection Regulation, applicable from May 2018, made profiling a regulated activity that
needs a lawful basis and gave people rights to object. California's Consumer Privacy Act, effective
from 2020, introduced an American opt-out model that other states followed. Apple's tracking
protections, and in 2021 its App Tracking Transparency prompt, changed the economics of cross-context
tracking, and most people, asked plainly, declined. Google announced in 2020 that it would remove
third-party cookies from Chrome, then delayed repeatedly, abandoned the plan in 2024, dropped even a
standalone choice prompt in April 2025, and retired most of its replacement APIs in October 2025. The
direction of travel was toward first-party, consented, and contextual personalization. The reversal
at Google shows the turn was neither linear nor complete, and that privacy outcomes for billions of
people were shaped by one firm's position as browser, ad platform, and search engine. The deeper
account of these turning points, read as ethics rather than compliance, is on the
[critics](/ethics) page.

### Wave 7: real-time and generative personalization (2016 to 2026)

Two changes define the present. First, the emphasis shifted from overnight batch segmentation to
in-session decisioning, which re-weighted personalization toward context, meaning what a person is
doing now, and away from biography, meaning who the person is. Some practitioners argue this is also
more compatible with privacy. Second, since the public release of large language models in late 2022,
systems can generate the personalized artifact rather than only select from a fixed inventory: a
product description, an email written for one recipient, an interface assembled per visit, an answer
shaped by remembered context. This reopens the oldest question in the history. Where earlier eras
asked who owns the model of the person, generative personalization adds a model of persuasion, able
to produce endless variants of a message and learn which works on whom. Where personalization
composes the interface itself, it overlaps with generative UI; this guide covers what is selected or
generated for whom and why, and the sibling
[Generative UI Guide](https://generativeui.guide) covers how interfaces are composed at runtime.

## What the word has meant

"Personalization" has never had a stable definition. Its meaning drifted along with who was selling
it, and the drift is itself part of the history.

- **User customization.** In early web usage, to personalize often meant what the user did: choose
  categories, set preferences, arrange a portal. In this sense personalization gave control to the
  person.
- **Relationship strategy.** One-to-one marketing folded that idea into the firm's ability to treat
  different customers differently, shifting the agent from the user to the firm and the goal from
  convenience to customer value.
- **Inference-driven adaptation.** Academic work settled on a neutral definition: adapting a system's
  behavior to an individual from a model of that individual. It is neutral about who controls the
  model, which is exactly the point at issue.
- **A euphemism for tracking.** Ad platforms called cross-site behavioral targeting "personalized" or
  "interest-based"; critics and regulators called the same practice "tracking" or "profiling." Each
  side's word served its interest.
- **A legal object.** The law re-coined the concept in functional terms: "profiling" under the GDPR
  and "recommender system" under the Digital Services Act. These terms deliberately avoid the word
  personalization, because the law regulates the mechanism, not the marketing promise.
- **Generation.** With large language models, vendors now use "personalization" for content created
  per person, which raises the question of whether a uniquely generated message is personalized or is
  simply targeted persuasion.

Read together, each commercial definition emphasized benefit to the person and each critical or legal
definition emphasized the mechanism and the data. Neither is neutral. Every definition of
personalization is also an argument about it.

## The intellectual roots

Five disciplines fed the field, and their values still shape the debate.

- **Information retrieval and filtering** supplied the standing interest profile and the vocabulary of
  relevance. Its central value, relevance to an information need, is the ethical root of the most
  defensible form of personalization: helping someone find what they were looking for.
- **Recommender-systems research** grew from those early systems into its own community and, unusually,
  developed its own critical sub-literature on popularity bias, feedback loops, and filter bubbles.
- **Marketing science** supplied the economic logic: segmentation, customer lifetime value, and the
  targeting models behind next best offer and personalized pricing.
- **Human-computer interaction and adaptive interfaces** asked how a system should adapt its interface
  to a person, and produced the work on explanation and user control. This lineage is the most direct
  ancestor of generative UI.
- **Data mining and machine learning** supplied the methods for inferring from implicit data, and with
  them a stance that treats predictive accuracy on held-out data as the standard of success. The
  critical tradition's central complaint, that proxy objectives diverge from human goods, is largely a
  complaint about that stance.

## Who the story leaves out

The usual canon runs through the collaborative-filtering pioneers, one-to-one marketing, Amazon,
and the Netflix Prize. Several contributors are routinely left out. Elaine Rich and the adaptive-
hypermedia researchers established user modeling well before e-commerce. Firefly's privacy work
showed that early personalization companies treated user-controlled profiles as part of the product,
an idea later lost and now being revived. The direct-marketing and database-marketing practitioners,
whose methods are the real ancestors of email and CRM personalization, rarely appear in
computer-science histories. The measurement and audit researchers who built the empirical
counter-tradition are usually treated as critics rather than as part of personalization's own story.
And non-US regulators are often overlooked; China's 2022 rules on algorithmic recommendation, for
instance, required an option to switch off personalized recommendations before European law did.

## The through-line

In each era the capacity to infer about people outran the person's knowledge of, and say over, the
model being built, and each privacy turn was a late and partial correction rather than a settled
resolution. The first wave briefly took the opposite for granted: the person rated things herself,
and privacy was a design goal. The durable question the rest of this guide returns to is whether
personalization can be built so that its benefits are real, measurable, and visible to the person,
and so that the inferences behind it are ones the person would accept if they were shown them. That
is less a prediction than a standard, and the history suggests technical capability will keep
outrunning it unless it is designed in.