Case study / Ecommerce experimentation
When search returns nothing, fix the search
How an international B2B ecommerce business addressed failed product searches through evidence-led experimentation, improving the reported order rate by 14%.
Client anonymised · Individual experiment within a wider international optimisation programme
01 / The problem
Customers knew what they wanted. Search wasn't helping them find it.
On-site search is often one of the clearest signals of customer intent. Someone has told you what they are looking for, and the business has an opportunity to help them find it.
Global search analysis identified a substantial proportion of searches returning no results, with a high exit rate from those pages. This established the scale of the problem across the business, before individual-market experiments were run.
That is a significant point of friction. Customers could be searching for a product the business stocked, but not using exactly the same terminology, part number or phrasing as the search catalogue expected.
The question wasn't how to make the empty-results page look better. It was whether we could improve the search logic so that fewer customers reached a dead end in the first place.
02 / The intervention
Make the search more flexible, not just the message more persuasive
The experiment introduced multiple ways to match a customer's original search to relevant products. Instead of accepting the full query as an all-or-nothing match, the experience could look for other plausible routes to a result.
01 / Break up phrases
Search the component words
A phrase such as “rugged hammer” could be searched as “rugged” and “hammer” separately, rather than only as the complete phrase.
02 / Reduce the query
Try shorter versions
The experiment tested reducing the original query in parts to identify potentially relevant results that an exact full-string search could miss.
03 / Extend matching
Look beyond the standard catalogue
A separately created extended-range product database offered another opportunity to find and present relevant products.
04 / Offer alternatives
Give customers a useful next action
Where the alternative logic could identify relevant items, it surfaced products rather than leaving the customer with an unhelpful dead end.
These elements were combined in the experiment. The outcome cannot be attributed to any single component without further testing.
03 / The experiment
An uplift in orders, with different performance by market
We evaluated the approach in the UK first and then in the USA. Both were part of the international experimentation programme, but the results were not identical. The UK delivered stronger results than the USA.
The archived experiment summary reports a 14% relative uplift in order rate, with statistical significance described as close to 100%. This is the reported result from the archived presentation, not a newly calculated aggregate of the two country-level tests.
The change in exit behaviour was much less pronounced. Improving product discovery did not necessarily mean every behavioural measure would move in the same direction.
What mattered was the commercial outcome. Search quality was evaluated in terms of orders, rather than assuming a lower exit rate alone would demonstrate success.
The initial diagnostic described global search behaviour. The later tests measured individual markets, so those exit-rate populations should not be treated as interchangeable. Market-level order rates, sample sizes and traffic volumes are intentionally withheld.
04 / What happened next
The result informed product development, not just a test report
The learnings were passed to the agile development squad, with particular attention to reductive search and breaking searches into component terms.
The team also planned further experimentation around stripping prefixes and suffixes from queries, alongside potential UX and UI refinements.
What the evidence supported
Making search more tolerant of imperfect queries was associated with a higher reported order rate in the A/B test.
What remained to investigate
Which matching methods created the most value, how different search intents behaved, and what further refinements would be worth implementing.
05 / The learning
Don't optimise a failure state if you can address the reason it exists
It would have been easy to treat the zero-results page as a design problem: improve the copy, add a more prominent CTA or encourage customers to browse a category.
Those might be useful ideas. But the underlying question was why so many customer searches were failing to return useful products.
By looking at the search behaviour and changing the matching logic, the experiment addressed the problem closer to its source. That is a more meaningful form of optimisation than simply making the dead end more attractive.
The lesson: start with the customer problem, investigate the mechanism behind it and choose the intervention that is most likely to improve the actual business outcome.
Evidence and interpretation
What the original report establishes
This anonymised retrospective case study is grounded in the archived internal test presentation. Intervention details, the reported relative uplift and next steps are drawn from that source. Exact traffic volumes, baseline conversion rates and market-level performance figures are withheld because they are commercially sensitive.
The original presentation reports a 14% relative uplift and describes statistical significance as close to 100%. Full statistical outputs, sample-size planning and separate market-level inference are not included in the material reviewed, so this is a reported historical result rather than a fresh statistical verification.
The report does not establish a standalone revenue figure for this experiment or a confirmed full international rollout. Its outcome should not be added separately to the wider programme's attributed commercial value.
