Classification Framework · April 2026

Recommendation Systems Taxonomy

This taxonomy governs how every system in the table is classified. Classification is determined by decision-layer ownership — which component directly performs retrieval and/or ranking that determines what items are shown. FM use that only creates embeddings, features, or signals upstream of the decision layer does not qualify a system for an FM category.

Unit of Analysis

One row = one recommendation system. A system may have multiple components (retrieval, ranking, re-ranking) described within the same row. Classification is applied at the system level, not the component level. One company or product may have multiple recommendation systems, each recorded as a separate row.

Foundation Model — Working Definition

FM criteria (all three must be satisfied)

1. Foundational role — trained at large scale and serving as a core model across multiple tasks, surfaces, or objectives; not merely a component within a larger pipeline of specialised models.

2. Decision-layer participation — directly performs retrieval and/or ranking decisions; its role is more than signal generation, feature enrichment, or embedding provision upstream of a separate decision-making model.

3. Built for purpose, not lightly adapted — designed and trained for recommendation (whether from scratch or via deep domain retraining from a general-purpose base); not a general-purpose LLM that was lightly fine-tuned or prompted.

FM basis note — within FM-classified rows, the table notes whether capability derives from semantic understanding (initialised from broad language/content pretraining; enables cross-domain transferability and generalisation to unseen items via text) or behavioural scale (trained on interaction/ID-based sequences; achieves strong domain-specific performance without cross-domain transferability or text-based generalisation). This distinction does not change the classification label — it is informational.

Note on SID-based models (e.g. GRID): using an LLM to construct the discrete ID vocabulary offline does not make the recommendation model itself semantic. The recommendation model trains on interaction sequences over those IDs — the LLM is a preprocessing tool, not part of the model that serves recommendations. Such systems are classified as behavioural.

Classification Categories

RS-Dominant

Retrieval and ranking decisions are owned entirely by purpose-built recommendation models (collaborative filtering, two-tower networks, GBDTs, multi-task neural nets, bandits, sequential models trained on interaction signals, etc.).

Any FM present is used only for signal generation, embedding provision, or post-hoc explanation — all upstream or downstream of the decision layer. No FM participates in making retrieval or ranking decisions.

FM-Dominant

A Foundation Model (as defined above) directly owns retrieval and/or ranking as the primary decision-maker. The FM is not a component feeding an RS — it is the system's core decision layer.

RS+FM Hybrid

A Foundation Model directly participates in retrieval or ranking decisions alongside RS components. The FM does not need to be the sole authority but must co-own part of the decision layer — not merely enrich it with signals.

FM-Wrapped RS

A general-purpose or domain FM provides a conversational or natural-language interface (chat, Q&A, itinerary generation, explanations). Retrieval and ranking decisions are made entirely by a traditional RS underneath. The FM translates user intent and surfaces pre-ranked results but does not independently rank items.

GP-FM Special Case

A general-purpose LLM (GPT, Claude, Gemini, Llama, Grok, or similar broadly-pretrained model) that has been lightly fine-tuned or prompted for a recommendation task and is confirmed to directly participate in the retrieval or ranking decision layer in production (including A/B tests).

Distinct from RS+FM Hybrid because the model was not designed or substantially retrained for recommendations — its general-purpose character is substantively preserved. Flags cases where a GP model is placed in the decision layer, which is architecturally notable.

Note — if a general-purpose model undergoes deep domain-specific retraining (continuous pretraining at scale on domain data, not just fine-tuning), it crosses into RS+FM Hybrid or FM-Dominant territory instead.

Deployment Status

Production

The system is fully deployed and serving the primary traffic on its target surface. No experimental gating — it is the live, default experience for all (or essentially all) users on that surface.

Partial Production

Live in production but intentionally scoped — for example: serving a subset of traffic (e.g. 10%), limited to specific surfaces, geographies, or user segments, or deployed alongside a legacy system with gradual rollout in progress. The system is real and serving users, but not yet the default for everyone on that surface.

A/B Testing

Actively being tested against a production baseline in a controlled online experiment. The system is receiving real user traffic and producing real recommendations, but only for the experiment arm — it has not been promoted to the default serving path. Results from A/B tests are the primary evidence of real-world impact cited in papers and blogs.

Distinction from Partial Production — A/B Testing means the system is in an experiment with a control group and has not yet been adopted. Partial Production means the rollout decision has been made and is underway, even if coverage is not yet 100%.

Research

Published results (paper or blog) based on offline experiments, simulations, or academic evaluations only. No confirmed online deployment or A/B test against real user traffic at the publishing company. Included when the system is sufficiently novel or influential to warrant tracking, with the expectation of future deployment.

Key Disambiguation Rules

When in doubt, apply these rules

→Large sequential model trained purely on interaction/ID signals — RS-Dominant regardless of size, transformer architecture, or generative framing, unless it satisfies all three FM criteria. Architectural sophistication alone does not qualify.

→Pipeline component (e.g. a user sequence encoder within a system of 1000+ models) — RS-Dominant. Fails criterion 1 (foundational role), regardless of architecture.

→Deeply retrained GP model (continuous pretraining at scale on domain data, not just fine-tuning) — FM-Dominant or RS+FM Hybrid, not GP-FM Special Case.

→Lightly fine-tuned GP-LLM in the decision layer — GP-FM Special Case, not RS+FM Hybrid.

→FM used only for embeddings/features upstream — RS-Dominant. Criterion 2 (decision-layer participation) not met regardless of FM quality.

Back to table

System entries

RS-Dominant

(5, 6)