Research desk · Original study
Keyword search misses one product in six — even when you search its own title
A baseline measurement of naive keyword retrieval on live US retailer feeds, using the friendliest possible query: the product’s own words.
Some links here are affiliate links — if you buy, the retailer may pay us a commission at no extra cost to you. We have earned $0 from them so far. Here’s how it works.
Part of: Research
Sample: 120 live US retailer listings. Method: query each listing’s own title back against the same catalogue and check whether that listing comes back. Date: 2 September 2026. Result: the listing was ranked first 69.2% of the time and appeared anywhere in the top ten 84.2% of the time. 15.8% of the time it did not come back at all.
That is the generous case. A real product search does not have the retailer’s own wording to work from, so this figure is an upper bound on naive keyword matching, not an average one. It measures a baseline technique against public catalogue data; it does not measure any particular matching system.
The headline table
| Measure | Result | Count |
|---|---|---|
| Exact listing returned at rank 1 | 69.2% | 83 of 120 |
| Exact listing anywhere in top 10 | 84.2% | 101 of 120 |
| A listing with the same title at rank 1 | 80.0% | 96 of 120 |
| A listing with the same title in top 10 | 85.8% | 103 of 120 |
| Queries returning nothing at all | 0% | 0 of 120 |
| Query errors | 0% | 0 of 120 |
120 queries against live US retailer catalogues, 2 September 2026. Median rank when the target was found: 1.
The quotable sentence, for anyone who wants one: searching a live retail catalogue for a product using that product’s own exact title fails to return it about one time in six.
Method
- We pulled 4,600 live listing rows from US retailer catalogues across twelve product families on 2 September 2026, and kept the 3,726 that had a title, a link and at least three content words in the title.
- We drew a stratified random sample of 120, ten per family, with a fixed seed (20260902) so the draw is reproducible.
- For each sampled listing we took the first eight content words of its own title — dropping stopwords and generic retail words like set, pack, color and size — and used them as the keyword query.
- We requested the top ten results for each query, in stock, US advertisers, and recorded whether the original listing came back and at what rank.
The catalogue query tokenises a keyword list and matches loosely, which is exactly the behaviour under test. This is what “search the feed for the product” does when nobody has built anything on top of it.
Where it fails
The average conceals the finding. Six of the twelve families returned the right product every single time; two families produced almost all the failures.
| Product family | Found | Rate |
|---|---|---|
| Holiday gifts | 10 of 10 | 100% |
| Tupperware | 10 of 10 | 100% |
| Dinnerware sets | 10 of 10 | 100% |
| Fiesta dinnerware | 10 of 10 | 100% |
| Christmas | 10 of 10 | 100% |
| Stocking stuffers | 10 of 10 | 100% |
| Loungefly | 9 of 10 | 90% |
| Disney gifts | 9 of 10 | 90% |
| Fiesta, any stock state | 9 of 10 | 90% |
| Advent calendars | 8 of 10 | 80% |
| Blue Willow china | 4 of 10 | 40% |
| Carnival glass | 2 of 10 | 20% |
Same 120 queries, broken out by the family the target listing came from.
Both failing families are replacement china and collectible glass: catalogues built from thousands of near-identical variants that differ by pattern, colour or form rather than by product. Fourteen of the nineteen failures came from one specialist retailer of exactly that kind.
The failure mechanism
Every failure we inspected has the same shape. The search returns the right maker and the right form, and gets the distinguishing word wrong.
| What we searched for | What came back first |
|---|---|
| Imperial Glass-Ohio Pansy White Carnival 9″ Pickle Dish | Imperial Glass-Ohio Beaded Block Pink Pickle Dish |
| Smith Glass Quintec White Carnival Creamer | Smith Glass Quintec Clear Creamer |
| Block Carnival Wine Glass | Block Carnival Brandy Glass |
| Harmony House China Pussy Willow Dinner Plate | Harmony House China Harmony Rose Dinner Plate |
| Waterford China Willow Flat Cup & Saucer Set | Waterford China Lavaliere Flat Cup & Saucer Set |
| Physical Disney Gift Card — Up | Physical Disney Gift Card — Disney+ |
Six of the nineteen failures, quoted verbatim from the query and the top-ranked result.
These are the most dangerous kind of wrong answer. A search that returns nothing is obviously a failure. A search that returns the same maker, the same object type and a different pattern looks like a success to anyone not checking the pattern name — which is precisely the person a keyword match is being run on behalf of.
Longer titles did worse
The counter-intuitive result, and the one with the clearest practical consequence.
| Query length | Sample | Target found |
|---|---|---|
| Three content words or fewer | 31 | 29 (94%) |
| Six content words or more | 56 | 41 (73%) |
Same 120 queries, split on the number of content words in the query.
More words made retrieval worse, not better, because the keyword list is matched loosely: each additional word pulls in another set of partial matches that can outrank the exact one. A precise, specific title is a worse query than a vague one, which inverts the intuition that everyone brings to a search box.
Keep reading
What this study cannot show
- It is not a measure of any product-matching system, including ours. It measures the raw behaviour of keyword retrieval against catalogue data, which is the baseline such systems are built to beat.
- It is one catalogue network on one day. A different network, or the same one next month, could produce a different number.
- Twelve families, ten each. A per-family rate computed on ten queries has wide error bars; the 40% and 20% figures are directional, and the aggregate over 120 is the number we would stand behind.
- It does not measure whether a human would be satisfied with the near-miss result. Somebody shopping for “a pickle dish” might be. Somebody replacing a specific pattern would not.
- It says nothing about images, prices or stock — only about whether the right row comes back.
Why the number matters
Product feeds are matched by keyword constantly — by affiliate tooling, by price comparison, by anyone building a shopping list from a spreadsheet. This measurement puts a floor under how often that goes wrong on the easiest possible input, and shows the error is concentrated in exactly the catalogues where precision matters most: replacement china, collectible glass, and anything sold as a pattern rather than a product.
If you take one operational rule from it: never accept a keyword match on a patterned or variant product without checking the distinguishing word — the pattern, the colour, the form. That is where all nineteen of our failures lived.
Frequently asked
How often does keyword search find the right product?
In this study, 69.2% of the time at rank one and 84.2% within the top ten, when queried with the product’s own exact title. That leaves 15.8% not returned at all.
Is that good or bad?
It is an upper bound. Real searches do not have the retailer’s own wording, so performance on ordinary queries will be lower than this.
Which products fail most often?
Replacement china and collectible glass. Blue Willow returned the target 4 times in 10 and carnival glass twice in 10, against 100% for six other families.
Why do longer, more specific queries perform worse?
Because keyword lists are matched loosely. Each extra word adds more partial matches that can outrank the exact one. Queries of three words or fewer found the target 94% of the time; queries of six or more, 73%.
Can I cite this?
Yes. It is a measured finding with a stated sample, method and date, published for that purpose.
Data and reproduction
- Source: CJ product API shopping-product queries, US advertisers, in-stock filter, run on 2 September 2026.
- Population: 4,600 rows pulled across twelve product families; 3,726 eligible after requiring a title, a link and three or more content words.
- Sample: 120, stratified ten per family, random seed 20260902.
- Query construction: first eight content words of the listing title after removing stopwords and generic retail terms.
- Matching: a hit requires the returned listing’s link and title to match the target’s. A looser title-only match is reported separately above.
- All 120 queries completed; no errors, no empty result sets.
Last verified 2 September 2026 against 120 live keyword queries run against US retailer catalogues during this run
Keep reading
How this guide was made. We research and draft these guides with AI, then a person checks every price, link and factual claim against the source before it publishes. We work this way because it lets us re-verify prices across hundreds of guides in a day, which is what keeps the numbers here current; it does not decide what we recommend. Anything we could not verify is labelled as unverified rather than filled in.