Intent parsed from the query
A query is read for what it constrains, not only for the words in it, so a search that names an attribute your catalogue records is used as a constraint rather than as a bag of tokens.
Search
Search that ranks on what the shopper is deciding.
It runs beside the search you already have rather than replacing it, scores every result on four factors you can read, and always holds a share of traffic back so you can see what it is actually doing.
Where you are
Keyword search is not the reason your shoppers cannot find things.
One query, four factors
- Matches the wordsyour search engine, already fineScored
- Fits the implied budgetfrom what this session browsedScored
- Separates the near-identicalneeds the deciding attributeNo field
- Explains its own positiona reason per resultScored
1/4of the four factors needs a field the catalogue does not record yet. The ranking says so rather than guessing.
Most search tools match the words in the box against the words in your catalogue, and they do it well. That is why replacing one with another rarely moves anything: the matching was never the weak part.
The weak part is what happens after the match. A query for a laptop returns ninety products that are all genuinely laptops, and the ordering has to decide which of the ninety this particular person should see first. Sorted by relevance they are near identical. Sorted by popularity you show everyone the same thing.
There is also a floor under all of it. If your catalogue never recorded screen size, no ranking can order by it and no shopper can filter on it, which is why the audit comes before this page in our own product order.
Ranking a set of results well is a different problem from matching them, and it is the one nobody solves by swapping search engines.
Matching is not ranking
Your search engine already found the right 200 results.
The problem is the order they came back in, and that is a different problem with a different fix.
- Results returned200your existing engine, doing its job
- A shopper will read10the first screen, and rarely past it
- Separable on recorded fields6the rest are indistinguishable in the order they arrive
Illustrative of the shape, not a measurement of your shelf. Swapping search engines changes the first row and leaves the third one alone.
What it looks like
Why this result came first.
Base relevance, persona, affinity and price band, per result, with the held-back arm beside it.
Engine
- 1ThinkPad X1 Carbon
- 2MacBook Pro 14
- 3Dell XPS 15
Held back
- 1MSI GS66 Stealth
- 2ThinkPad X1 Carbon
- 3Razer Blade 15
A share of sessions always sees your own order. That is the comparison.
What happens
How the ranking works
Four factors, computed per request, and every one of them readable.
- Match01
Relevance, as now
Query intent is parsed and matched against the catalogue. This part is not the clever bit and we do not claim it is; it is the baseline every result starts from.
- Session02
Weighted by the visit
What this person has browsed, searched and filtered in this visit moves persona, affinity and price-band fit. No cookie and no login, so it works on a first visit.
- Always03
A share never sees it
A held-back arm, stable for the session, ranks the same query your original way. That is what makes any number anyone reports about this measurable rather than asserted.
What you get
What you get
Four factors, named
Base relevance, persona fit, attribute affinity and price band. Each contributes a visible amount to the score, and they are the same four in the console and in the API response.
A decision record per request
Every ranked response carries an id and the factors behind it. When somebody asks why a product was third, the answer is retrievable rather than reconstructed.
A control arm by default
Not a setting you turn on when you want to measure. Some of your searchers are permanently served your existing ranking, and the two are reported side by side.
Before you commit
What search will and will not do
What this does
- Rank results against what the current visit implies, and show the factors behind every score
- Run alongside your existing search rather than replacing it
- Serve some searchers your existing ranking permanently, so there is always something to compare against
- Return your existing result order rather than an empty page if the time budget is missed
What it does not do
- Come with a conversion figure. Nothing has been measured on your traffic yet, and a result from another merchant's shelf is not evidence about yours
- Rank on an attribute your catalogue does not record. Nothing can; that is what the audit is for
- Replace your search infrastructure, your storefront or your platform
- Work from customer records. It reads the session, not a profile
Questions
The ones that decide it.
- Do we have to replace our search engine?
- No. It ranks the result set your search returns rather than producing that set from scratch, so your existing index, your synonyms and your infrastructure stay where they are. If you switch it off, your own order is what renders, because that order is the fallback path.
- How fast is it?
- No model call and no database read on the hot path, because the work is a rank over a set your own engine already retrieved. The console shows the whole distribution, since a median tells you nothing about the slowest search in fifty.
- Why can you rank without a cookie?
- Because it infers from what is happening in the current visit rather than from who the person is. Browsing, searching and filtering in this session are enough to move the ranking, which is also why it works on a first visit when a profile-based system has nothing to go on.
- How do we know it is doing anything?
- The held-back arm. A share of your traffic always sees your original ordering, stable per session, so the two arms can be compared on your own traffic rather than against a number we quote you.
- What if our catalogue has gaps?
- Then ranking helps less than fixing them would, and we will say so. No ordering can use an attribute nobody recorded. That is why the Decision Audit is first in our product order and this is not.
Start with the catalogue, then rank it.
Search is worth turning on once your shelf can answer the questions people ask of it. One category export settles that.
Request a Decision AuditNothing is priced until a pilot has run on your own searches.