The recommendation layer that added 1.4 points to search conversion

For a signed-in shopper, ASOS personalised search no longer opens with the same clothing rail every other visitor sees. The retailer says the first 200 rows can be reordered using what that customer viewed in the previous month and what they bought, bagged or saved during the previous four months. The query still retrieves the products. A recommendation system decides which stand first.

The test question was narrow: would filtering search through a customer’s existing taste signals make more visits end in an order?

In results published on 5 November 2026 for the 52 weeks to 30 August, the British fashion retailer gave its answer. ASOS reports that personalised search lifted conversion by 140 basis points, equivalent to about £31 million in annualised revenue. The public methods stop short of the detail needed to check it.

How ASOS personalised search reorders results

ASOS sells its own clothing and products from hundreds of partner brands. That breadth is part of the proposition and part of the search problem. A shopper can arrive with a precise phrase, a loose occasion or only a style in mind, and keyword retrieval can match the words while placing the least useful option first.

The retailer had already built a recommendation layer from behavioural signals. Its public ranking explanation for partners says logged-in search results consider product views from the past month, plus purchases, bag additions and saved items from the past four months. Logged-out shoppers instead see country-level signals such as bestsellers, newness, strategic importance, size availability and stock position.

The FY26 results describe the intervention as bringing that recommendation technology into search. Search still retrieves a candidate set, and the recommendation layer reranks it for the individual shopper.

This AI search personalisation layer decides which products get the scarce positions at the top of a high-intent journey. It also needs a data pipeline that joins identity, recent behaviour, availability and commercial rules fast enough to serve a live page, because a delay or stale stock signal can erase the benefit before the shopper sees it.

The treatment changed the order of results, not the query

The most copyable part of this recommendation engine experiment is its boundary. ASOS did not replace the shopper’s query with a chat session or generate a new catalogue. Search handled the expressed need, and recommendation signals changed the order of the answer.

That separation makes the comparison easy to design. One group keeps the incumbent ranking and the other gets the same search engine with personalisation applied. Assignment should persist per shopper, since someone moving between rankings contaminates both conversion paths. The primary metric can stay close to the decision, such as conversion or revenue per eligible searcher, with sitewide revenue per visitor showing how much the result matters to the whole business.

ASOS says it tests ranking changes on selected customer groups and updates its published ranking logic when a change becomes permanent. Its annual results also say AI support more than doubled the number of experiments the company ran during the year. The proportion with a positive outcome improved almost fivefold, while the share of conclusive experiments rose by more than 15 percentage points.

Those figures point to a wider experimentation system rather than one isolated launch, which is the high tempo testing that Hacking Growth by Sean Ellis and Morgan Brown puts at the centre of a growth team. They do not show how the personalised search comparison itself was assigned, measured or stopped, and the results give no allocation table, test dates or statistical threshold.

A 1.4 point lift became a £31 million estimate

A 140 basis point conversion lift means an absolute movement of 1.4 percentage points. That is not the same as a 140 percent increase. If baseline conversion were 5 percent, for example, a 1.4 point increase would take it to 6.4 percent, a 28 percent relative lift. ASOS does not publish the baseline, so the relative change cannot be calculated from the announcement.

The company translates the reported lift into about £31 million of annualised revenue. The word that matters is annualised: the figure projects a measured rate across a longer period or larger eligible audience. It does not say £31 million of additional revenue appeared in the reported accounts.

An outside reader also cannot tell whether the estimate uses retail sales, gross merchandise value or another revenue definition, nor whether it adjusts for returns, discounts or partner commissions. ASOS reported group gross merchandise value of £2.349 billion for the year and adjusted revenue of £2.240 billion, two different denominators because flexible fulfilment sales are increasingly commission based. The estimate needs the same definition before it can be set against a finance line.

The absence of public sample size matters too. A 1.4 point difference can be compelling or noisy depending on eligible traffic and variance. ASOS gives no confidence interval, device split, geography or exposure, and does not say whether only signed-in customers were counted, although its public description limits behavioural reranking to logged-in users.

The public ranking rules show the operating model

Personalisation does not act alone. Inventory, country demand, newness and full-price priorities still shape rankings, especially for logged-out shoppers, and recent behaviour is bounded by time windows, so an old purchase cannot decide every future result.

Merchandisers can still demote discounted products outside sale periods, raise strategically important ones and protect size availability. The recommendation system supplies a customer-specific order inside that commercial frame and does not own assortment or pricing.

The next test should publish the denominator

ASOS says its next layer will combine search, outfit building, a customer wardrobe and an AI stylist, so ecommerce search conversion becomes one input into a larger taste model.

The follow-up should keep a stable holdout while those features arrive. A useful readout gives eligible searchers and sessions, the treatment split, baseline conversion, absolute and relative lift, revenue per searcher, statistical uncertainty and the definition behind annualised revenue, split by new and returning customers, logged-in and logged-out visitors, market, device and query type. That is what turns a search revenue lift into a forecast.

Guardrails belong beside the sales result. Product diversity can fall when a recommender keeps showing the same style, and latency can worsen while zero-result rate improves. Returns and gross margin matter too, because more orders are not automatically more profitable orders.

ASOS took a recommendation asset it already ran, put it on a high-intent surface and reported a business metric rather than carousel clicks. The public record supports the mechanism and the 140 basis point claim. It does not yet support the precision of the £31 million figure.

Experiment detail	What ASOS reports
Commercial problem	Make high-intent search results more relevant without replacing the shopper’s query.
Intervention	Recommendation engine reranks the first 200 search rows for logged-in shoppers.
Signals used	Views from the past month; purchases, bag additions and saved items from the past four months.
Reported result	140 basis point conversion lift, about £31 million annualised.
Programme context	Experiment volume more than doubled, positive outcomes almost fivefold, conclusive tests up more than 15 percentage points.
Not published	Market, dates, sample, allocation, baseline conversion, confidence interval and revenue definition.

Sources and disclosure

• ASOS FY26 results for the 52 weeks to 30 August 2026
• ASOS ranking rules for partners

ASOS published the personalised search result in its own annual results and has not released the test design, sample, allocation or confidence interval. The £31 million figure is the company’s annualised estimate.

---

FAQ

Compare more measured experiments in the Signal Index, browse the Growth Signal archive, or use the benchmark tools in the IC Lab.