SERVICE
AI Recommendation & Search Engines
Ranking and semantic search that turns browsing into buying.
semantic retrieval at scale
Reviewed by Mohammed Affaan Khan, GenAI & Agentic AI Engineer · Updated July 2026
How much does recommendations cost? See what drives the price — and get a quote scoped to your budget.→What recommendations actually means
A recommendation engine decides what a person sees first. That makes it a ranking problem, not a matching problem — the question is never "what is related to this" but "what should occupy the position that gets almost all of the attention".
The judgement that separates a useful system from a harmful one is the objective. Optimise for clicks and you will get clicks, along with returns, unsubscribes and a catalogue that collapses onto whatever was already popular. The objective has to be the outcome the business is actually paid for.
What you get
- Conversion-weighted ranking models
- Natural-language semantic search
- Real-time vector retrieval (<50ms)
- Full MLOps: training → registry → serving
How we build it
Objective definition
What the system is optimising — margin, retention, completed orders net of returns. Everything downstream is determined by this and it is a commercial decision, not a technical one.
Signal collection
Views, carts, purchases, returns, dwell and search terms, joined to a stable identity. Weak identity resolution is the most common reason these systems underperform.
Candidate generation
Narrowing a large catalogue to a few hundred plausible items quickly, using collaborative and content-based signals.
Ranking
Scoring candidates against the objective, with business rules — margin, stock, contractual placement — applied explicitly rather than smuggled into the model.
Cold start
Sensible behaviour for new users and new items, which is most of the catalogue on any given day and where naive systems visibly fail.
Online evaluation
A/B testing against the real objective. Offline metrics are directionally useful and routinely disagree with what happens in production.
The stack
RETRIEVAL
Vector search and approximate nearest neighbour for semantic similarity, sub-50ms at production volume.
RANKING
Gradient-boosted models or neural rankers, chosen for the data volume rather than for novelty.
FEATURES
A feature store keeping training and serving consistent — skew between the two is a silent and expensive bug.
SERVING
Low-latency APIs with fallbacks, because a slow recommendation is worse than a generic one.
EXPERIMENTATION
A/B framework with correct assignment, guardrail metrics and honest readout.
MONITORING
Coverage, diversity, popularity bias and per-segment performance.
What it connects to
- Product catalogues and PIM systems
- Ecommerce platforms and storefronts
- Site search
- CDP and customer data platforms
- Email and marketing automation
- Analytics and experimentation tooling
Where teams use it
Ecommerce
Product discovery ranked on completed orders net of returns rather than on clicks.
Media
Content ranking optimised for sustained engagement rather than the next session only.
Travel
Property and itinerary ranking where availability and price change constantly.
Marketplaces
Two-sided ranking balancing buyer relevance with seller fairness.
B2B catalogues
Reorder prediction and cross-sell on long, technical catalogues.
Search
Semantic ranking that handles the exact product codes pure vector search misses.
How a build runs
Objective and audit
1–2 weeks
The metric agreed, signal quality assessed, identity resolution checked.
Baseline
2 weeks
A simple, honest baseline — often popularity — to measure everything else against.
Build
4–8 weeks
Retrieval and ranking serving live traffic behind an experiment.
Test
2–4 weeks
A/B results against the real objective, with guardrail metrics watched.
Iterate
ongoing
Segments, cold start and business rules refined on evidence.
When this is the wrong answer
A model that lifts click-through while raising returns has made the business worse. If the objective is wrong, better modelling accelerates the damage.
Below roughly a few thousand items and a few thousand active users, simple heuristics compete well with a learned system and cost far less to run.
Personalisation requires a unified customer record. Joining that data is usually the real project, and the model is the easy part.
Popularity bias is the default failure. Without deliberate diversity handling the system narrows the catalogue and the long tail stops selling.
Proof
Booking-optimized discovery engine with sub-50ms search
Multi-phase ranking system with GRU event encoding and natural-language venue search served from a real-time vector store.
Frequently asked questions
What makes a recommendation engine convert?
Training on your real funnel: we weight models toward booking and verified-attendance events, not just clicks, so ranking optimizes revenue.
Can users search in plain English?
Yes — semantic search handles queries like “date night in Glasgow” via embeddings, returning relevant results in under 50 milliseconds.
Do you handle the ML infrastructure?
End to end: offline training data assembly, model training and registry, embedding stores and quantized production inference.
How much data do we need?
Enough interaction history to learn from — typically a few thousand active users and a few thousand items before a learned system reliably beats good heuristics. Below that, popularity and rules compete well and cost far less to operate. We will tell you which side of that line you are on.
What should it optimise for?
The outcome you are paid for, which is rarely clicks. Revenue per session net of returns, margin, or retention. This is a commercial decision rather than a technical one, and getting it wrong means better modelling accelerates the damage.
What about new products with no history?
Cold start is designed in rather than patched later, using content signals and attributes until behavioural data accumulates. New items are most of the catalogue on any given day, and a system that cannot rank them quietly suppresses new stock.
Will it just recommend our bestsellers?
Not if diversity is handled deliberately. Popularity bias is the default failure of these systems — left alone they narrow the catalogue and the long tail stops selling. We monitor coverage as a first-class metric alongside conversion.
How do we know it improved anything?
A/B testing against the real objective, with guardrail metrics watched. Offline ranking metrics are directionally useful and routinely disagree with live behaviour, so a system without an experiment framework cannot tell you whether it helped.
Can it work alongside our merchandising rules?
Yes, and those rules stay explicit rather than being smuggled into the model. Margin, stock position and contractual placement are applied as visible business logic, so a merchandiser can see why something ranked where it did.
How do we know whether it worked?
We agree the metric before the build and baseline it before anything changes, so the comparison is possible afterwards. Without a baseline every result can be described as a success, which is why so many AI projects are.
Can our own team maintain it after handover?
That is the intended end state. We use standard, widely-known technology rather than anything clever, document decisions as they are made rather than at the end, and run handover sessions with your engineers. If a system can only be maintained by us, we have built it wrong.
Who owns the code and the models?
You do, from the first commit. Work happens in your repository under your licence and the contract assigns IP outright. We keep no rights, hold no keys you cannot rotate, and build nothing proprietary that makes leaving expensive.
Recommendations by industry
What this is built on
The engineering disciplines behind recommendations, each with its own scope and constraints.
Recommendations near you
North America
United Kingdom
Europe
Asia-Pacific
Latin America
Before you choose anyone
Written to be useful whether or not you hire us — including the parts that argue against hiring an agency at all.
How to choose an AI development company
ReadAI agency vs in-house team
ReadCustom AI vs off-the-shelf
ReadOffshore vs local AI development
ReadAI Voice Agents: The Complete Guide for Businesses (2026)
ReadRAG vs Fine-Tuning: Which Does Your Business Need?
ReadHow Much Does AI Development Cost in 2026?
ReadBuild recommendations with Midalaxy
Tell us what you are building. We reply within one business day.