Find the right item without the part number

B2B customers often order many complex items. Without a part number it can be hard to find exactly the right one, especially when the customer describes the need in their own words. If the search doesn't find the right item, the order can go to a competitor. This article shows how AI product search solves that, what it brings to the business, how quality is measured and what it costs to run.

An example: an installer needs to seal an LPG fitting and types “yellow thread tape LPG”. The product is called “PTFE thread tape for gas lines”, and the wrong tape on a gas line can leak. The AI product search puts thread tape for gas at the top, although neither PTFE nor gas line was typed.

The business case in brief

The business value. The goal is for more searches to end in an order, fewer cases to reach customer service and more customers to come back. A search with no hits can be a lost order.
Low risk. The search answers with items from your own catalogue. AI writes no text in the answer, so it can't invent items or prices.
Fits your setup. Built in Azure, so it can sit close to your existing data. Secure, with a cost cap and measurable quality.
From demo to production. The demo shows the core. Production adds, among other things, your own synonyms, control for marketing and sales, statistics and alerts.
Investment. Running costs are about 1,000–3,000 SEK a month. The big investment is the work on your product data, and it can pay off beyond search too.
Improves with data. Customers' real searches and buyers' knowledge of replacement items make the search better over time. With AI, that work goes faster than before.
More business opportunities. The same search can give a customer service chatbot real items to suggest, or match order lines from email to the right part numbers. It can also be part of a strategy for agentic shopping.

The demo is in Swedish. The recording shows a search for screws that won't rust in outdoor wood, and a V-belt by size.

Try the demo

How good product search creates value

A search with no hits can be a lost order. People who use the search box are often in a buying process. If the search doesn't find the item, the customer may buy it elsewhere.
Friction costs customers. More than half of B2B buyers say they are likely to switch suppliers if buying is not smooth. 54% of them cite poor digital experiences.¹
Easier buying can mean larger orders. Buyers who get helpful information that eases the purchase are three times as likely to choose the bigger option.³
Lower cost per order. A customer who finds the right item alone doesn't need to call or email customer service and sales.
Returning customers. A buyer who quickly finds the right item has a reason to come back. In B2B, returning customers often make up a large share of the business.
An edge over competitors. In a review of 23 B2B sites, 56% only returned results when the search term exactly matched the product text.² That is an edge to take.

More figures and what they show are in What does research say about B2B e-commerce?

Hybrid search

The technique is called hybrid search. A standard keyword search and a search on meaning (semantic search) run at the same time, and the results are merged. Keyword search is best for part numbers. Meaning search is best when the customer describes a need.

Plain words. The search understands what the customer means, not just the words that are typed. “Screws that won't rust” finds stainless and coated screws.
Part numbers. Part of a number is enough, and hyphens or spaces don't matter. “6203-2RS” finds the whole family of 6203 bearings.
Sizes and specifications. “3L280 V-belt 3/8 x 28” finds the right V-belt, with sizes converted to millimetres.
Variants. One product in several sizes shows as one row with selectable variants, not twenty near-identical hits.

The catalogue is based on an open research dataset with product data from Amazon, translated into Swedish. In a real project it would be your own products, prices and stock levels.

Applications

Other systems can use the same search through an API. That is where the value grows, because the same investment pays off in more parts of the business.

Customer service chatbot. A chatbot that invents products is a risk. If the bot gets its suggestions from the search instead, it answers with real items, part numbers and prices. Customer service can get relief and the answers can be checked.
Orders from email and PDF. Free-text order lines are matched to the right part numbers before they go into the ERP. That can save manual entry and reduce wrong deliveries.
Customer service and sales. An internal tool where staff quickly find the right item, even when the customer describes it in their own words on the phone. The goal is shorter calls and the right item the first time.
Spare parts and accessories. Find a replacement for a discontinued item, or parts that fit a given product. It gives the customer a reason to stay with you instead of searching at a competitor.
Configurators. Suggest components that meet the customer's requirements, with rules that guarantee the combination is valid. That can mean larger orders and fewer returns.
Agentic shopping. More purchases may come to be made by AI agents that search and order on the customer's behalf. A search with an API that answers with real items, prices and part numbers gives the agent something to work with.
Product data. The same AI step that reads product texts also finds missing or conflicting attributes in the PIM. Better data gives better search, filters and product pages.

From demo to production

The demo shows the core of the search. Production adds tools for the people who own the online store, for example:

Synonyms. The e-commerce manager adds their own synonyms and trade terms, without a developer.
Control for marketing and sales. Lift campaign items, lower discontinued ones and steer what shows for important searches.
Statistics. See what customers search for, what returns no hits and which searches lead to a purchase.
Monitoring and alerts. Alerts when response times rise or when the share of searches with no hits suddenly grows.

The demo also has limitations:

The demo has no customer prices, stock levels or contracts. In a real project they come from the ERP.
The demo catalogue is machine-translated research data. Some texts are clumsy and some products don't belong in an industrial catalogue.
The interface is in Swedish. English queries work but give weaker results.

Expected cost in production

Running costs are about 1,000–3,000 SEK a month, and more with guaranteed uptime. A simple way to judge it is to compare that cost with the margin on the orders the search saves.

The big investment is the work on your product data. Much of the time in a project goes into structuring, enriching and correcting it. That investment can also pay off in the PIM, on product pages and in other channels.

Running costs per month:

Part	Smaller catalogue	Larger catalogue or high load
Search index	Basic, about 740 SEK	Standard S1, about 2,450 SEK
Hosting (containers, logs, registry)	about 150 SEK	about 300 SEK
AI per search	about 3 SEK	about 30 SEK
Total	about 900 SEK	about 2,800 SEK
With guaranteed uptime (two index replicas)	about 1,650 SEK	about 5,200 SEK

The search index is the largest part. With more research and testing, it may be possible to find cheaper alternatives, but they are not always in Azure.

One-off cost for a catalogue of 100,000 items:

Vectors. About 7 SEK. Redone only for new and changed items.
AI enrichment of the product texts. About 1,500 SEK, or about 750 SEK as a batch run. Redone only when an item changes.

How it is calculated:

Prices. Azure list prices in Sweden Central in October 2026, at 1 USD ≈ 10 SEK.
Searches. 10,000 searches a day for the smaller catalogue and 100,000 for the larger. A search is at most 200 characters and costs about 0.00001 SEK.
Items. The one-off cost assumes about 300 tokens per item for vectors and 800 for enrichment. That is the catalogue text, not the searches.
Hosting. The smaller figure is what the demo costs. The larger is an estimate.
Number of items. Reliability and load drive the price, not the number of items.

Built in Azure

If the company already runs its ERP, product data and sign-in in Azure, the search sits close to the data and under the same contract, security rules and billing. That can mean a shorter start and lower risk.

Search index. Azure AI Search, with both a keyword index and vectors.
AI models. Azure OpenAI in Sweden Central, in Azure's EU data zone.
Infrastructure as code. The infrastructure can be recreated automatically.
Deployment. Every change is tested automatically before it reaches customers.

Azure is not the only option. Elasticsearch, OpenSearch, Algolia and similar services solve the same problem. Some may have strengths Azure lacks, such as ready-made merchandising tools or lower cost at very large volumes. The choice depends on what the company already has and needs.

Security and rules

In a customer project the most important question is who may see which prices and which assortment. Search should only show what the customer has a contract for.

The AI Act. Product search is not among the high-risk areas in the EU AI Act.⁴ The answers are real items from the catalogue, not text written by AI.
Least privilege. The website only has a key that can read from the search index, not change it.
No AI keys. The service signs in to the AI models with a managed identity in Azure. There are no keys to leak.
Cost cap. When the day's cap on AI calls is reached, search continues with keyword search only. The site stays up and the cost can't run away.
Personal data. The demo needs no login. It states that searches are logged and stores no names. In a customer project, searches can be stored to improve the search. They are then stored without names or accounts, with a set retention period, and handled under GDPR.
Abuse protection. The demo caps searches per visitor and AI calls per day. In production the search can also sit behind a login, so prices and assortment are shown only to customers and abuse can be traced to an account.

How do you know the search delivers?

The search is measured against a test set, not by trying a few queries by hand. Every improvement can therefore be shown in numbers before it reaches customers. For part numbers, the right product is in the top five for every test query. Across 90 mixed test queries, a fully relevant product is in the top five for 87% of queries, against 63% with a standard keyword search.

Query type	nDCG@10
Part numbers	0.95
Realistic B2B queries	0.79
Generated Swedish queries	0.78
Swedish queries, overall	0.71

The nDCG@10 metric shows how high the right products rank among the first ten hits, on a scale from 0 to 1.

Test queries. The test set has hundreds of queries, and every hit has been graded as relevant, partly relevant or not relevant.

Two lessons:

Measure after every change. A change that looked right on one query lowered the overall score from 0.82 to 0.77. The test set is therefore rerun after every ranking change.
More AI is not always better. A language model was set to interpret every query before searching. It helped English queries but lowered the realistic Swedish ones from 0.79 to 0.69.

Operations and quality over time

An AI search is not done when it is deployed. It needs the same care as the rest of the online store. With AI, much of that work goes faster than before, and more decisions can be based on data instead of gut feeling.

When product data changes. New and changed items are read in, get new vectors and are updated in the index.
Customers' real searches. AI can go through thousands of real searches, group them and show where the search falls short. For the demo, 280 test queries were generated and graded with AI, work that otherwise takes many hours by hand.
Buyers' knowledge. Information on what replaces what, for example in orders or replacement lists, can become suggestions in the search. This requires that the information exists in the systems. A person reviews the suggestions before they are used, since a wrong replacement item can be costly.
Quality. The test set is rerun after every change. It should grow with real queries that gave poor results.
Follow-up. Searches with no hits can show demand you don't meet today. That leads to synonyms, better product texts or new items in the assortment.
New product types. New categories and trade terms are added to the glossary, so search and filters understand them.

Next step

Try the demo. Want to know how the search would work with your catalogue?

Book an intro call

McKinsey & Company, "Five fundamental truths: How B2B winners keep growing", September 2024. B2B Pulse Survey of nearly 4,000 decision makers. ↩
VML, "The State of B2B Site Search". Review of 23 B2B sites against 22 criteria. The page is undated, and VML sells e-commerce implementation. ↩
Gartner, press release, April 2019. Sample not stated. ↩
Regulation (EU) 2024/1689 of the European Parliament and of the Council on artificial intelligence, Annex III, which lists the high-risk areas. ↩