Across 90 product searches in 30 categories, walking the search term from a broad category down to a long-tail product raised the marketplace share of the top three organic results from 19.3% to 48.3% and cut the share held by brands and specialist shops from 44.3% to 23.6% — the opposite of the direction predicted before measuring, and the opposite of what "go long-tail" advice assumes.
On 2026-08-13 we ran 90 live searches. The frame is 30 consumer product categories — running shoes, mattresses, chef knives, headphones, dog beds, cordless drills, fountain pens and the rest — drawn three from each of 10 retail sectors so that no sector can carry the result on its own. The list was written out in full and committed before any search ran.
Each category carries three search terms of increasing specificity, fixed in advance: the broad category a shopper would first type, a named sub-type within it, and that sub-type plus one or two qualifying attributes — the "long-tail" term the standard advice tells a small retailer to build a page for. The template around the term never changes: the word "buy", the term, the word "online". That is the entire design. The term's specificity is the only thing that moves, so anything that moves with it is attributable to the term rather than to the phrasing.
For each query we recorded the first three organic results — title and URL, verbatim — giving 270 result slots across 113 distinct hosts. Every slot was then classified on two axes defined in advance and kept deliberately independent. The first asks who controls the domain: a marketplace where third parties list, a retailer selling across many unrelated categories, a specialist retailer inside one product world, the brand whose own product it is, a publisher, or a community site. The second asks what kind of page it is: one product, a shoppable list, an article, a homepage, or a forum thread.
Keeping the two axes independent is the whole design, because several of the findings below are a cross-tabulation of them. If the domain were allowed to decide the page type — assuming a marketplace result must be a product page, or a publisher result must be an article — the study would have manufactured the correlation it claims to have found. Both rules therefore read only the title string and the URL path, every row records the named rule that classified it, and where the rules genuinely do not decide, the slot is recorded as low confidence and dropped from that axis rather than guessed at.
Six numbered predictions were written into the protocol before measuring, and they are scored on this page rather than quietly forgotten: 1 of 6 held. So were 6 conditions that would have stopped publication — a minimum number of complete categories, a minimum sample, a ceiling on unclassifiable slots on each axis, and a requirement that the specificity ladder had genuinely changed the results rather than returning the same pages three times. All 6 passed.
This is the question the study was built to answer, and the prediction was written down in advance because the advice is so universal: you cannot beat the marketplaces on a head term, so go long-tail. We committed to marketplace share falling by at least five points between the broad category term and the long-tail term. It rose by 29 points.
On the broad category term, marketplaces held 19.3% of top-three organic slots. On the named product term, 28.4%. On the long-tail term — the one the advice recommends building for — 48.3%, about two and a half times the category-term share. The move is monotonic across all three rungs.
The mirror of that number is the one a small retailer cares about. Slots on a domain an independent seller could own — a brand's own site, or a shop specialising in that product world — were 44.3% of the category-term results, 26.1% of the product-term results and 23.6% of the long-tail results, a fall of 20.7 points. We had predicted a rise of at least eight. The broad-range retailers, meanwhile, gave ground: from 34.1% to 25.8%. The long tail did redistribute the page — it just redistributed it to the marketplaces.
The obvious objection is that the engine ignored the longer terms and returned the same pages three times, in which case the ladder would be measuring nothing. That was pre-registered as a condition that would change the headline, and it did not happen: 0% of long-tail slots carried a URL that also appeared in that same category's category-term results. Every rung returned genuinely different pages. Why the pattern runs this way is beyond this instrument, and the limits above say so. What can be said is that across these categories on this day, the advice pointed the wrong way.
| Search term | Marketplace | Broad-range retailer | Specialist retailer | The brand itself |
|---|---|---|---|---|
| Category term | 19.3% (17) | 34.1% (30) | 27.3% (24) | 17% (15) |
| Product term | 28.4% (25) | 43.2% (38) | 10.2% (9) | 15.9% (14) |
| Long-tail term | 48.3% (43) | 25.8% (23) | 10.1% (9) | 13.5% (12) |
The second axis moved even harder than the first, and in the direction the advice assumes. On the broad category term, a page about one single purchasable product was 0% of results: not one slot in 90. The category term returns lists — 84.4% of those slots were a shoppable list of many items, and 12.2% were somebody's homepage.
On the long-tail term, single product pages were 43.3% of results, 39 slots. So the mechanism the advice describes is real: name the product precisely enough and the engine stops showing you departments and starts showing you items. The part the advice gets wrong is whose items. Combine the two axes and the long-tail page that ranks is, more often than anything else, a marketplace product listing.
The page-type mix differs sharply by who owns the domain, which is why the two axes were kept independent. Marketplace slots are never anything but commerce: 72.9% shoppable lists and 27.1% single products, with no articles and no homepages at all. Ownable slots are the only ones that arrive as a homepage — 16.9% of them do, which is a small brand or shop ranking on its front door rather than on a page built for the query.
| Search term | One product | A shoppable list | A homepage | An article |
|---|---|---|---|---|
| Category term | 0% (0) | 84.4% (76) | 12.2% (11) | 3.3% (3) |
| Product term | 4.4% (4) | 91.1% (82) | 2.2% (2) | 2.2% (2) |
| Long-tail term | 43.3% (39) | 50% (45) | 1.1% (1) | 4.4% (4) |
Percentages across a whole sample hide the thing a shop owner actually needs, which is whether a given search has a slot in it for them. Counted per search: of 90 searches, 42 returned no ownable slot in the top three at all — not one of the three results was a brand's own site or a specialist shop.
That count follows the ladder. On the broad category term 6 of 30 searches had no ownable slot. On the product term 19. On the long-tail term 17. At the other end, 10 searches returned three ownable slots out of three, and 7 returned three marketplace slots out of three.
Concentration is worth stating plainly. 113 distinct hosts appear across 270 slots, but one of them — amazon.com — accounts for 24.1% of every slot measured, appearing 65 times. The next, walmart.com, holds 13.3%. Those two together are 37.4% of the whole sample. And the leader's share climbs with specificity in step with everything else here: 15.6% of category-term slots, 36.7% of long-tail slots.
| Search term | Searches with no ownable slot | Ownable share of slots |
|---|---|---|
| Category term | 6 of 30 | 44.3% (39) |
| Product term | 19 of 30 | 26.1% (23) |
| Long-tail term | 17 of 30 | 23.6% (21) |
The ladder is the finding, but it is not the largest spread in the data. Ownable share ran from 52% in apparel and accessories down to 14.8% in kitchen — a wider range than the 20.7 points the whole specificity ladder moved.
The sectors where a brand or a specialist shop held the most ground are the ones where the product is itself an identity: apparel, outdoor gear, hobby goods. The sectors where they held least are the ones where the product is a commodity a general retailer already stocks. Nothing here explains that, and the cells are small — the counts are printed beside each figure precisely so that no sector line is read as more solid than it is.
| Sector | Slots | Ownable | Marketplace |
|---|---|---|---|
| Apparel and accessories | 25 | 52% | 36% |
| Outdoor and sport | 26 | 46.2% | 30.8% |
| Baby, kids and office | 27 | 44.4% | 33.3% |
| Hobby, craft and music | 27 | 37% | 33.3% |
| Tools, garden and auto | 26 | 30.8% | 26.9% |
| Home and furniture | 27 | 29.6% | 18.5% |
| Pet supplies | 26 | 23.1% | 42.3% |
| Electronics and audio | 27 | 18.5% | 29.6% |
| Beauty and personal care | 27 | 18.5% | 40.7% |
| Kitchen | 27 | 14.8% | 29.6% |
Six predictions were committed before the first search. 1 held. They are printed here because a study that reports only the predictions it got right is not reporting a method, it is reporting a conclusion — and on this one we were wrong about nearly everything, including the direction of the effect the study exists to measure.
The misses are worth naming individually. We predicted marketplace share would fall with specificity; it rose. We predicted the ownable share would rise; it fell. We predicted the ownable share on category terms would be under forty per cent; it was 44.3%, higher than expected, which is the one direction in which the picture is friendlier than we assumed. We predicted publishers and forums would take over fifteen per cent of long-tail slots on the theory that no shop would match the string; they took 2.2%, because the marketplaces match every string. And we predicted at least ten categories would return no ownable slot on their category term; 6 did.
The single prediction that held is the least interesting one: marketplaces are at least a quarter of everything, at 32.1% across the whole sample.
| Prediction | Threshold | Measured | Held? |
|---|---|---|---|
| Marketplaces hold at least 25% of all classified slots. | ≥25% | 32.1% | Yes |
| Marketplace share falls from S1 to S3 by at least 5 points. | fall ≥5 pts | +29 pts | No |
| Ownable is under 40% of S1 slots. | <40% | 44.3% | No |
| Ownable rises from S1 to S3 by at least 8 points. | rise ≥8 pts | -20.7 pts | No |
| Publisher plus community is over 15% of S3 slots. | >15% | 2.2% | No |
| At least 10 of 30 categories return zero ownable slots on their S1 query. | ≥10 of 30 | 6 of 30 | No |
Six conditions were written into the protocol before measuring, any one of which would have stopped this study being published. They are listed because a floor that is never shown is a floor nobody can check. The one that matters most here is the last: had the long-tail searches returned the same URLs as the category searches, the headline would have had to become that instead, and the ladder predictions would have been recorded as untestable rather than failed.
| Gate | Floor | Measured | Passed? |
|---|---|---|---|
| Categories complete across all three rungs | ≥24 of 30 | 30 of 30 | Yes |
| Queries measured | ≥72 | 90 | Yes |
| Slots measured | ≥216 | 270 | Yes |
| Low confidence on seller type | ≤15% | 1.9% | Yes |
| Low confidence on page type | ≤15% | 0% | Yes |
| Ladder did not collapse (S3 URLs repeating S1) | ≤40% | 0% | Yes |
This was not in the protocol, so it is labelled and it is not the headline. A pattern noticed after seeing the data is weaker evidence than one committed to beforehand, however clean it looks.
The ownable slots are not held by the businesses the word implies. Of the 83 ownable slots in the sample, a substantial share sit on large national chains — an outdoor co-op, a pet superstore, a national tool chain, a shoe megastore — all of which are correctly classified as specialist retailers under a rule fixed in advance, and none of which is the independent shop the word conjures. "A domain of the kind a small seller could own" and "a domain a small seller does own" are not the same statement, and this study measured the first. Separating them needs a classification of business size that was not pre-registered here, so it is a study for another week rather than a number for this page.
Pick a product you actually sell. Search "buy", the broad category, "online" — then the same thing with a narrow long-tail version of the term. Ask two questions of each of the first three results: who controls this domain, and what kind of page is it. A dozen searches is enough to see which end of this range your own category sits at, and it costs nothing but ten minutes.
The protocol was committed before the first search, with the category list, all three terms per category, the query template, both classification schemes, the publication floors and the 6 predictions written down in advance. The instrument probe that fell outside the frame was discarded and logged rather than kept. The raw rows are committed with the named rule behind every classification, and the script that turns those rows into every figure on this page is committed beside them. Re-run it and you get these numbers, or you get a bug report worth having.
Or start from ours. Every figure on this page is published as JSON at /functions/serveResearchData, under CC BY 4.0 — free to chart, re-cut or quote, with attribution, and with the method and limits travelling alongside the numbers. New studies arrive by RSS.
We build and rank websites, so read this one knowing that. The honest reading is uncomfortable for our trade. The advice we are all expected to give — go long-tail, build the deep product pages, you cannot fight the marketplaces on the head terms — pointed the wrong way in this sample: the marketplace share of the top three rose from 19.3% to 48.3% as the term got more specific, and in 42 of 90 searches there was no slot in the top three that a small seller could own at all. That does not mean nothing can be done; it means the sector you are in and the specific terms you chase decide far more than a general rule does, and a plan built on the general rule is aimed at something that frequently is not there. We cannot tell you from this data what will rank for you, and we are not going to imply it. What we can tell you is what the results for your own products actually contain today, because that is a measurement rather than an opinion, and we will show you the rows.