Autocomplete and suggestions

For a buyer who knows the article number, the dropdown under the search box is the entire shop. They type four characters, click a row, and are on the product page. Everything else — the results list, the facets — is what happens when autocomplete failed.

What the dropdown should contain

Four sections, in this order. Most B2B shops need the first two and benefit from all four.

All of these are fetched in a single request, not four. That matters: the dropdown has to redraw on every keystroke, so the budget is well under 100 milliseconds end to end.

Article-number prefix matching

This is the part that decides whether buyers use your shop.

A buyer typing 4711 must see 4711-A and 4711-B by the fourth character. A buyer pasting KAT-4711-A-2024 out of their own system must see 4711-A. A buyer typing 6205 must see the SKF bearing even though your own number for it is something else entirely.

What that requires:

• Prefix matching enabled on identifier fields, so a partial code matches from the front.
• Mid-string matching enabled on the same fields, so a code embedded in a longer string is found.
• Typo tolerance off on those fields. In a dropdown, one allowed typo means 4711 also suggests 4712 — and the buyer clicks without reading.
• An exact whole-string match pinned to row one, visually distinct from the fuzzy suggestions below it.

Recent and popular, before a key is pressed

An empty search box is an opportunity. Two lists cost nothing and are used constantly:

• Recent searches for this buyer, from their own session and account.
• Popular searches for this channel — the queries other buyers in this storefront make most.

Popular searches are computed from your own query log. Filter out zero-result queries before you show them; suggesting a search that returns nothing is worse than suggesting nothing.

Per-customer suggestions from order history

The highest-value autocomplete in B2B is not clever text matching. It is "you ordered this before".

For a signed-in buyer, promote articles from their own organization's order history above general matches, and label them. A purchaser re-ordering the same twelve articles every month wants row one to be the thing they ordered in March, not the newest equivalent.

Three variants worth building, in order of effort:

All three respect the customer's assortment automatically, because the suggestion query carries the same channel and organization filter as the main search. A buyer never sees a suggestion for an article they cannot order.

Steps

1. Confirm identifier fields are prefix- and mid-string-matchable in App Studio › Services › Search.
2. Configure the dropdown sections and their row counts. Start with products, the pinned exact match, and categories.
3. Set the query fields for suggestions to identifiers and name only. Including description fills the dropdown with articles whose relevance the buyer cannot see.
4. Enable recent searches, and popular searches with zero-result queries excluded.
5. Enable previously-ordered promotion for signed-in buyers.
6. Test on a phone. Innendienst and fitters use the shop on phones far more than your analytics suggest, and a dropdown that covers the whole screen with ten rows is unusable there.

What to check

Take ten real article numbers off recent orders and, for each one:

1. Type the first four characters. The article must appear.
2. Type the whole number. It must be row one, exact-matched.
3. Type the number with a wrong last character. It must not silently offer the neighbouring article as if it were a match.

Then sign in as a real customer and confirm that an article they have never been entitled to buy does not appear in any suggestion.

When it does not work

Next

• Merchandising and boosting
• Search analytics — where popular queries and click-through come from.