ITMTB · Enterprise search
Your search says “no results.” Your database says otherwise.
When people can't find a product, a report or a document you already have, they assume you don't have it. We replace the search behind your existing search box, prove it on your own content before it goes live, and spend the first month tuning it to what your users actually type.
For one publisher, searches that found nothing fell from 53% to 5% after launch.
The problem
Most failed searches aren't missing content. They're missing the match.
Most search boxes match the exact words stored in your records. People don't search that way. They misspell, abbreviate, use their own word for a thing, or type more than the title says. Each mismatch ends the same way: an empty page, or the wrong results at the top.
On one catalogue we worked on, more than half of all searches returned nothing. In a sample of 500 of those failures, 470 were for a report that already existed.
Why searches fail
- •A specialist term "corrected" into a common word by a general-purpose dictionary.
- •A longer, more specific search returning nothing, because every word had to match.
- •Short terms such as AI, 5G or EV silently left out of the index.
- •Filler words such as "latest" or "report" treated as mandatory.
What it costs you
- •Customers conclude you don't have it, and buy elsewhere.
- •Employees ask a colleague, or redo work that already exists.
- •AI assistants built on the same search answer confidently from the wrong material.
- •The business keeps writing, pricing and hosting content that nobody can reach.
Nobody reports an empty result as a bug, so the business rarely sees the cost.
What fixing it did for one customer
From our live deployment for The Business Research Company, a market research publisher with 25,700 reports.
How we fix it
Search tuned to your content and your users, not a generic dictionary.
Every fix starts from real failed searches in your own logs, and nothing goes live until it scores better on your own content.
Understands how people actually search
Typos are corrected using your own vocabulary, so a specialist term is never "fixed" into a common word. Longer, more specific searches narrow the results instead of returning nothing. Abbreviations, hyphenated names and short terms like AI or 5G are found, not dropped.
Proven on your content before it goes live
We build a test set from your own items, each searched for the ways people get it wrong: misspelt, a word missing, words reordered, abbreviated. Every change is scored against it. If any kind of search gets worse, the change does not ship, even if the overall score went up. After launch, a console shows every search, how good its results were, how fast it was and what people clicked.
Safe to put in front of revenue
It runs as a separate service with its own database access, so it cannot touch your orders or enquiries. If it ever fails or runs slow, your current search answers instead. Switching back is a single setting.
One search for your people and your AI assistants
The same engine is available to AI assistants and agents through MCP (the Model Context Protocol), so an assistant gets the answer your search box would give. Optional meaning-based matching catches what keyword matching misses, with a safeguard so that nonsense still returns nothing.
What you get
The engine, plus everything needed to run it well.
Search engine
Keyword matching tuned to your content: your vocabulary, the filler words your users add, graceful handling of long searches, and a boost for exact titles, subjects, places and years.
Search console
Every search, its results and whether they were good, weak or empty. Speed and clicks for each place search appears on your site, a query explorer and CSV export.
Business rules
Pin, hide or redirect results for specific searches yourself, with a preview. No developer, no release.
Access for AI assistants
The same search available to AI assistants and agents through the Model Context Protocol (MCP), with access scoped per project.
Operations
The index rebuilds nightly, refreshes every fifteen minutes, checks itself against your source data and raises an alert if updates stop. Releases are versioned and approved.
Meaning-based matching
Answers searches that keyword matching cannot, and suggests related items. Runs on a small local model: no external AI service, no per-search cost. Off by default, and switched on only when it proves itself on your data.
How an engagement runs
Prove it first. Run it where you want. Tune it for a month, included.
Evaluate on your content
Before you commit to anything, we test your current search against a sample of your items and your real failed searches. You see where it breaks, kind by kind, and how much a tuned engine would recover. The decision rests on numbers, not a demo.
Deploy on your servers or ours
Run it on your own infrastructure beside your existing application, or let us host it. Either way it keeps its own index, and your current search takes over automatically if anything goes wrong. Your search box and results page stay exactly as they are.
A month of fine-tuning, included
The first month after launch is when search learns your users. It is part of the engagement, not a change request.
- •Someone reads the console every day and reviews every empty, weak or rescued search.
- •Each real failed search becomes a permanent test before it is fixed, so it cannot break again.
- •Ranking adjusted to what your users actually type: long searches, filler words, places, years, compound terms.
- •Synonyms and vocabulary drawn from your content and your search logs, reviewed with your team.
- •Business rules set up with your team for searches where commercial priority should decide the order.
- •Speed targets and timeouts tuned to your traffic, with error counts confirmed at zero.
- •Meaning-based matching tested on your data, and switched on only if nonsense searches still return nothing.
- •Your team trained on the console: what people want that you don't have, and what they click.
Measure every month
How much of your content people actually reach, what they ask for that you genuinely don't have, what they click, and how search leads to the enquiries, orders or downloads that matter to your business. Every engine update is tested before it ships.
Success story
The Business Research Company
A global market research publisher with 25,700 reports was telling more than half of the people searching its site that the report they wanted did not exist. In almost every case we checked, it did. Three weeks after design sign-off, our engine was live on their site, with the old search kept as a safety net throughout.
- 53% → 5%of searches returning nothing
- 98.8%of known items found in the top five, across 20 kinds of misspelling, reordering and abbreviation
- 1 in 10searches rescued by typo correction that uses the catalogue's own vocabulary
- 0fallbacks to the old search, degraded responses or lost events, all monitored
Ready to Transform Your Business?
Join industry leaders already scaling with our custom software solutions. Let’s build the tools your business needs to grow faster and stay ahead.
FAQ
Frequently asked questions
Reading on enterprise search
The method, the measures and the deployment, in full.
Everything on this page is written up with its evidence. Start with the success story.
The Business Research Company: zero-result searches from 53% to 5%
What was wrong, what was built, how it was measured and what changed for a 25,700-report catalogue.
Read
How to measure the success of an enterprise search tool
Twelve measures across offline tests, live telemetry and business outcome, and five ways the numbers can mislead.
Read
Why enterprise search is important
Six reasons search decides more than it is given credit for, and a short check for your own catalogue.
Read
A realistic dataset for comparing enterprise search solutions
The 20 query classes, the nonsense guardrail, real failed searches, and how to score vendors fairly.
Read
Improving enterprise search for people and AI agents
The retrieval failures found in a real catalogue's logs, and why fixing retrieval came before any interface.
ReadFind out what your search is missing.
Send us a sample of your content and a week of search logs. We'll show you where your current search fails, kind by kind, and how much a tuned engine would recover.