1,123 responses, 5 models, 45 Dutch questions: the baseline AI-visibility snapshot of the Netherlands bicycle and e-bike market

Vertical: Bicycles & e-bikes (fietsen & e-bikes), Netherlands
Method: A single point-in-time study of 1,125 responses across five large language models (Claude, Gemini, GPT, Perplexity, Grok), each asked 45 Dutch-language questions five times
Collection window: 5 September 2026

Key terms: AI visibility, AI Visibility Score, bicycle brands, e-bike brands, fietsen, elektrische fietsen, bakfiets, longtail, speed pedelec, Dutch bicycle brands, LLM brand recommendations, ChatGPT/Gemini/Claude/Perplexity/Grok brand recommendation, Generative Engine Optimization (GEO).

This report reproduces the real Dutch-language questions people ask AI assistants about bicycles and e-bikes, and measures how often, in what order, and across how many models five large language models name each entity. Because the same questions can be read with or without an origin restriction, because the market divides into genuinely different product formats, and because a question about a motor is not the same as a question about a bike, the entities are ranked in nine separate cuts, defined in Chapter 2. It is a single point-in-time study and should be read within the limitations in Chapter 11.

How to read this report: visibility is not quality

Every number here measures whether an AI assistant names an entity, and how early. It does not measure whether the product is well made, whether it lasts, whether the price is fair, or whether the recommendation was accurate. A brand can be highly visible and mediocre, or excellent and invisible. Read the leaderboards as a map of findability in a new discovery channel — nothing more.

1. Executive Summary

One-sentence takeaway: Gazelle leads the Dutch bicycle shelf in seven of nine cuts, but the single most-named entity in the entire study is a motor manufacturer that sits outside the ranking by design.

• Gazelle leads decisively. It tops the general leaderboard at 94.31, appearing in 594 of 1,123 analyzed responses (52.89%) and named by all five models. Cube (73.04, on 407 of 1,123) and Batavus (69.27, on 394 of 1,123) follow, and the gap from first to second is 21.26 points — wider than the gap from second to eleventh.
• 19 entities qualified for the general ranking. Of 222 dictionary entities, 218 matched at least one response, and 19 Core entities cleared the 5% mention threshold across all 45 questions: 10 Dutch and 9 foreign. Their average score is 53.8, and 17 of the 19 are named by all five models.
• The most-named entity is a component brand. Bosch appears in 626 of 1,123 responses (55.74%), more than any bicycle brand in the study. It is classified Adjacent, reported in Chapter 8, and deliberately excluded from every leaderboard. Shimano (460 responses, 40.96%) would rank second on the same basis. Ranking motors against bicycles would have made the leaderboard a table about drivetrains.
• Origin depends entirely on which questions you count. On the 34 origin-neutral questions (848 responses), 7 of 20 qualifiers are Dutch and Dutch brands take 40.0% of the top ten. On the 11 questions that explicitly ask for a Dutch brand (275 responses), 15 of 19 qualifiers are Dutch and Dutch brands take 90.0% of the top ten — by construction of the question set, not as a finding. The two figures are printed together everywhere in this report and neither is quoted alone.
• The origin-neutral shelf is genuinely contested. Cube is named in 382 of 848 origin-neutral responses against Gazelle’s 381 — a one-response difference — and Gazelle keeps the top score only on position and breadth. This is the one cut in the study where the domestic leader does not lead on reach.
• Position can outrank reach. Urban Arrow ranks fourth in the general cut on a mention rate of 7.12% (80 of 1,123), because when it is named it is named first: it holds the general-cut position maximum at an MRR of 0.6744. The composite is working as designed, but the underlying numbers should be read alongside the score.
• The models differ far more in how much they say than in whom they name. Grok names 9.10 Core entities per response against GPT-5.6’s 3.25, and Grok mentions Gazelle in 78.7% of its answers against Claude Haiku’s 24.4%. Every model puts Gazelle first among Core entities. Grounding runs from 41.8% for Gemini to 100% for three models, against an overall rate of 85.6%.
• Six of the nineteen ranked entities were in or through insolvency. Accell Nederland B.V. was declared bankrupt on 13 August 2026, 23 days before collection, leaving continuity unresolved for Batavus (3rd), Koga (9th) and Sparta (10th); VanMoof, QWIC and Stella are all post-bankruptcy. The assistants had not caught up. Visibility is not viability.
• Three segments, three leaders, one caveat. Gazelle leads everyday & commuting (100.00), Cube leads sport & recreational (97.74) and Urban Arrow leads cargo, family & special formats (85.79). All 15 attribute questions fall in the everyday segment and the other two carry none, so segment leaderboards are indicative only and entity counts must not be compared across them.

Why it matters. Discovery is moving from search engines and marketplaces toward AI assistants, and the Netherlands is an unusual place to watch that happen: bicycles are ordinary infrastructure here rather than a specialist purchase, and the questions people ask reflect that — commuting distances, battery range in headwind, whether parts will still be available in five years. This edition asks a narrow question about it: when someone types a Dutch question into an assistant, whose name comes back, and how early. It is the first (baseline) snapshot, so there is no prior edition to compare against and none is implied anywhere in this report. One reminder: the numbers describe visibility and information availability, not which bicycles are best.

2. Methodology

One-sentence takeaway: Five large language models were each asked 45 Dutch questions five times with no system prompt, producing 1,125 responses of which 1,123 were analyzed; entities were identified by alias matching plus human-gated research and ranked with a three-component (45/30/25) score computed separately within each of nine cuts.

2.1 Scope and models

The study queried five large language models directly via API with identical questions. All models received only the user question. No system prompt was used, temperature was not set manually (provider default), and no region or locale parameter was defined. Because the questions are in Dutch, the models infer the Netherlands market from language alone.

45 questions × 5 models × 5 repeats produced 1,125 responses collected. Two responses returned a completed status with empty text and are excluded from every denominator, giving 1,123 responses analyzed. Denominators are computed per (question, model) cell, never as questions × models × repeats, which is why the per-model denominators are not all 225.

2.2 Web-search configuration and grounding

No model was instructed to search. Grounding is whatever each provider does by default, and the spread is the widest this series has recorded: 85.6% of all 1,123 responses were web-grounded, but that average conceals Claude Haiku, Perplexity and Grok at 100%, GPT-5.6 at 86.2%, and Gemini at 41.8%. Because grounding differs so sharply by model, and so does answer breadth, breadth carries only 25% of the composite score. Chapter 5 treats the spread as a finding in its own right.

2.3 Questions and cuts

Questions are written the way Dutch consumers actually ask them, including colloquial phrasing, English loanwords, abbreviations and imperfect grammar. No question was rewritten and none was removed because its answers were inconvenient: new_text equals original_text for all 45 questions, and every quotation in this report reproduces the question as asked. All 45 appear in Dutch with English glosses in Chapter 12.

Nine cuts are published. There is no trend cut in this edition, because there is no prior edition to trend against.

Every question belongs to exactly one segment, exactly one question type, and exactly one origin scope, so each of the 45 questions contributes to four of the nine cuts: general, one of local/open, one segment and one type. Question 45 — “Zijn er goede Nederlandse merken voor een elektrische bakfiets of longtail waarmee ik twee kinderen en boodschappen kan meenemen?” — is a local, use-case, cargo question, and its 25 responses sit in the general, local, use-case and cargo denominators and nowhere else.

The denominators reconcile in three independent ways: local (275) plus open (848) equals general (1,123); the three segment cuts (724 + 200 + 199) sum to 1,123; and the three question-type cuts (375 + 375 + 373) sum to 1,123. Every entity’s response counts decompose the same way.

2.4 Entity resolution

Candidates were extracted from the analyzed responses after stripping markdown link targets, bare URLs, domain-style anchors and citation markers, so that source domains could not inflate brand mentions. 9,157 raw surface forms were reduced to 2,137 candidates reaching a floor of ≥3 responses. Mechanical curation sent 211 of those to human-gated research and dropped 1,926, each with a stated reason and the statistic the rule fired on.

The drop list is worth reading, because it shows what a Dutch-language corpus does to capitalization heuristics. Nederlandse was dropped at 415 responses and Nederland at 342 as nationality terms; Wh at 406 and Nm at 253 as unit abbreviations; Bosch Performance at 213 as a product designation sitting under its parent. Riese Müller was dropped at 202 responses as a markdown artefact with no clean-text occurrence.

Three backstop passes then recovered entities the generator had missed, and the failures fall into four classes:

1. Sub-brands demoted as model names. Rockrider and B’TWIN were dropped as Decathlon product designations — Rockrider because 86% of its occurrences directly followed another proper name — while Elops and Van Rysel survived. Sub-brands count independently in this series; the rule was wrong and both were restored.
2. Brand names that are domain forms. Bakfiets.nl, Bol.com, Fiets.nl and Fietsenwinkel.nl were removed by URL stripping before candidate generation. They are now protected before stripping and matched as aliases.
3. Lowercase-styled and vocabulary-colliding names. biXbi and iMove never surfaced as capitalized runs; Mantel collides with an ordinary Dutch noun.
4. Names containing an ampersand. Riese & Müller never existed as a candidate at all. It entered the dictionary only as a promoted merge target, inheriting the response sets of Riese, R&M and Supercharger, and finished 8th in the general cut on 225 responses. A top-ten entity was invisible to the extractor for a reason that had nothing to do with the market.

The largest single retailer miss was 12GO Biking, named in 46 responses. It does not clear the general-cut threshold (4.10%) and qualifies only in the attribute, discovery and everyday cuts.

251 candidates were researched in total: 212 Keep, 32 Merge, 7 Drop. Eleven merge targets did not exist as candidates and were promoted to canonical entities inheriting the merged response sets. The final dictionary holds 222 entities carrying 437 aliases — Core 166, Adjacent 32, Out-of-scope 24.

Four dictionary entities matched zero responses, which is why 218 entities appear in the metrics and 222 in the dictionary. All four are Out-of-scope non-brands, so no leaderboard, cut or origin share is affected. Three of them are publications whose domains are cited heavily even though their names are never written in the answer text: elektrischefietser.nl is cited in 61 responses, bicycling.nl in 57 and bestebike.be in 4. Bikefixr is neither named nor cited. The gap between “cited as a source” and “named as an entity” is real and this is what it looks like.

Matching is alias-based, Unicode-normalized (NFC), case-insensitive, URL-stripped and anchored on word boundaries with tolerance for spacing and hyphen variants. Unrestricted substring matching is not used. Dutch compounds are handled only through registered aliases: forms such as Bosch-motor, Gazelle-dealers and Gates-riem were observed in the corpus and added explicitly. All 142 aliases of six characters or fewer were inspected with in-context samples; ART (the Dutch lock certification of Stichting ART), AD (the newspaper Algemeen Dagblad) and NS (the railway operator, in the context of folding-bike luggage rules) were retained as Out-of-scope entities that cannot affect a leaderboard.

2.5 The AI Visibility Score (0–100)

Per entity, per cut:

mention_rate = responses / denom
mrr          = sum_rr / responses      (reciprocal of the first-mention rank)
breadth      = distinct models
qualified    = mention_rate >= 0.05

c_mention  = 100 * mention_rate / max_mention[cut]
c_position = 100 * mrr / max_mrr[cut]
c_breadth  = 100 * breadth / 5
ai_score   = 0.45*c_mention + 0.3*c_position + 0.25*c_breadth

An entity is counted once per response, at its earliest mention.

Scale basis: qualified Core entities within the cut. This is not a market-wide basis, and the reason is the whole shape of this vertical. Bosch appears in 626 responses, more than any bicycle brand. Had the scale been drawn from all entities, every bicycle brand in the Netherlands would have been scored against a German motor manufacturer, and the leaderboard would have measured how close each brand came to a component. Adjacent entities are scored on the same published scale and are therefore free to exceed 100: Bosch reaches 105.39 on c_mention in the general cut. Those values are reported and never clipped; 41 component values across the study exceed 100, the highest being Kymco at 172.33 on c_position in the everyday segment.

Worked example, general cut. Gazelle has a mention rate of 0.5289 against a cut maximum of 0.52894, giving c_mention = 100.00. Its MRR is 0.5464 against a cut maximum of 0.67442, held by Urban Arrow, giving c_position = 81.02. It appears in all 5 models, so c_breadth = 100.00. The composite is 0.45 × 100.00 + 0.30 × 81.02 + 0.25 × 100.00 = 94.31.

The leader does not reach 100 unless one entity tops all three components. In the general cut it does not, and that is expected behavior rather than an error. Gazelle does reach exactly 100.00 in the local, attribute and everyday cuts, where it holds both the mention and position maxima.

2.6 Origin, neutrality and self-exclusion

Origin is trademark ownership, researched per entity and stored as origin_trademark. Local means the trademark country is the Netherlands. Belgium and Germany are Foreign. Benelux is not a rollup and does not appear as one in this report. A .nl domain, Dutch-language marketing, Dutch retail availability, a Dutch importer or a Dutch subsidiary do not make an entity Dutch. Tables use ISO country codes; the mapping from the full country names stored in the data pack was applied at render time, and the full names appear in the entity index. The domestic/foreign rollup is used for shares only.

Both trademark and operator country were collected for every entity, so the classification can be revisited later without repeating entity research. Nine entities have Unknown trademark origin. They are excluded from origin shares and retained in every other metric; none of them qualifies in any cut, so no published share moves either way.

Marketplaces and retailers are ranked in every published leaderboard, because they are what the assistants actually name. The _brandonly companion view strips five marketplaces by name — Alltricks, BikeFair, Bol.com, Marktplaats and Refurbed — and retains own-label and vertically integrated retailers everywhere.

Self-exclusion. The publisher’s own domain was checked for in the cited-source data and does not appear, so no exclusion was applied and no citations were removed.

2.7 What a future edition will and will not be able to compare

This is edition 1. There is no prior Netherlands edition, so this report contains no cross-edition comparison and does not gesture at one. When edition 2 arrives, mention_rate, mrr, breadth and raw response counts will be directly comparable, as will the source mix, the per-model behavior table and the question set, which is held constant. ai_score will not be comparable, because the scale is redrawn from whichever entities qualify in the new cut. If the leader changes, or if a new entity clears the threshold, every score in the table moves without anyone’s visibility having changed.

3. Overall Visibility Leaderboard

One-sentence takeaway: Nineteen entities cleared the 5% threshold across all 45 questions, Gazelle leads by 21 points, and four of the top fifteen are named far more often than their rank on reach alone would suggest.

3.1 General leaderboard: 19 qualified entities

General cut — all 45 questions, 1,123 analyzed responses.

Tiers. The board separates into four. Gazelle stands alone at 94.31. A second tier of Cube (73.04) and Batavus (69.27) sits 21 to 25 points below it. A long third tier runs from Urban Arrow (61.06) down to Decathlon (43.08), thirteen entities inside eighteen points, where rank is decided by fractions and small movements would reorder it. A fourth tier of QWIC (40.77), Stella (33.87) and Upway (33.79) sits at the threshold, and the last two are the only qualifiers not named by all five models.

Seventeen of the nineteen appear in all five models. That is a high-agreement market by this series’ standards: the assistants differ over how many names to give, not over which names are legitimate.

The composition is not what a bicycle-market ranking would look like. Fourteen of the nineteen are conventional brands; three are brand-retailer hybrids that sell their own product directly (Tenways, Canyon, Stella); and two are pure retailers (Decathlon, Upway). No marketplace qualifies in the general cut. The retailers are here because assistants answering “where do I buy this” name a shop, and stripping them would describe a shelf that no user is shown.

3.2 The same cut, ranked by mention rate

Reordering by reach alone moves four entities substantially. Urban Arrow falls from 4th to 15th, Tenways from 6th to 11th and Canyon from 12th to 14th, while Sparta rises from 10th to 7th and Kalkhoff from 14th to 10th. The pattern is consistent: the entities that gain from the composite are ones the models name early in a narrower set of answers, and the entities that lose are ones named often but late in a list. Neither ordering is more correct; they answer different questions, and both are published for that reason.

3.3 Brand-only view (marketplaces removed)

Removing the five named marketplaces changes the general leaderboard not at all: no marketplace qualifies in the general cut, and none qualifies in seven of the other eight. The single exception is Marktplaats, which qualifies 20th of 37 in the cargo segment on 14 of 199 responses. The brand-only companion tables are published for consistency with the rest of the series, and the honest summary of them in this edition is that they are identical to the published boards everywhere except one segment.

That is itself a finding. In the Turkish markets this series has covered, marketplaces dominate the assistants’ answers. In the Dutch bicycle market they barely appear: the retail names the models reach for are specialist bike shops and e-bike webshops, not general marketplaces.

3.4 Above the threshold but off the board

Four Out-of-scope entities clear 5% in at least one cut and are correctly excluded from every leaderboard, because they are not entities a shopper can buy:

ANWB appears in 217 of 1,123 responses (19.32%) and on responses alone would sit just below Koga, ninth of the nineteen, if it were rankable; Consumentenbond appears in 133 (11.84%). Both are consumer organizations whose bicycle tests the models cite constantly, and their presence says something about the Dutch information environment that a brand leaderboard cannot: when an assistant is asked which e-bike is reliable, it often reaches for a testing body before it reaches for a manufacturer.

Accell is the awkward one, and it became more so three weeks before collection. It is classified Non-brand / Out-of-scope because consumers do not buy an Accell, but it owns Gazelle, Batavus, Koga and Sparta — four of the ten Dutch entities on the general board — and Accell Nederland B.V. was declared bankrupt on 13 August 2026. It qualifies in the local and attribute cuts on two of five models. Chapter 11 sets out what that means for reading this board. Readers who want a corporate view of the market should note that four leaderboard rows share one owner; this report ranks the names the assistants say, not the groups behind them.

3.5 Below the threshold

The 19 qualifiers are drawn from 166 Core entities, so 147 Core entities were named at least once and did not clear 5% of 1,123 responses. They are published in full in the data pack rather than summarized away, because the tail is where a baseline edition is most useful: it is the list against which edition 2 will be read.

Two properties of that tail are worth stating now. First, an unqualified entity is scored against a scale it did not help set, so its components can exceed 100 without anything being wrong — the qualification threshold decides who defines the scale, not who is measured on it. Second, the tail is long and shallow: no unqualified Core entity in the general cut reaches 5% of 1,123 responses by definition, which means every one of them sits below 57 responses, and most sit far below it. A brand appearing in 40 answers out of 1,123 is genuinely present in this channel and genuinely invisible in it, and both halves of that sentence are true at once.

4. Origin: the constrained and origin-neutral shelves

One-sentence takeaway: Dutch brands take 40.0% of the origin-neutral top ten and 90.0% of the constrained top ten, and only the first of those figures says anything about visibility.

Eleven of the 45 questions explicitly ask for a Dutch brand — “ik zoek een degelijk Nederlands fietsmerk”, “Wat is een goed Nederlands e-bikemerk als prijs-kwaliteit belangrijker is dan luxe?” — and 34 do not. The two sets are reported separately and always together, because the constrained set cannot be used to argue that Dutch brands are more visible: it was built to return them.

4.1 The origin-neutral shelf (34 questions, 848 responses)

Twenty entities qualify: 7 Dutch and 13 foreign, with Dutch brands taking 40.0% of the top ten. This is the valid origin comparison in this report, and it describes a market where the domestic industry leads but does not dominate.

The detail that matters most is at the top. Cube is named in 382 of 848 origin-neutral responses and Gazelle in 381 — a single response apart, and the mention maximum in this cut belongs to Cube, not to Gazelle. Gazelle holds the top composite (89.66 against 86.99) on position and breadth alone. Take away the origin clause and the Dutch leader’s advantage over a German competitor comes down to one answer in eight hundred.

Below the top two, the foreign presence is broad rather than concentrated: Trek, Giant, Riese & Müller, Specialized, Canyon, Kalkhoff, Tern, Decathlon, Cannondale, Merida and Fiido all qualify, spanning the United States, Taiwan, Germany, France and China. Nineteen of the twenty are named by all five models.

4.2 The constrained shelf (11 questions, 275 responses)

Nineteen entities qualify and 15 are Dutch. Gazelle reaches 100.00, holding both the mention maximum (213 of 275, 77.45%) and the position maximum. Four foreign brands leak through the constraint — Riese & Müller, Cube, Giant and Reany — which is what the models do when the Dutch shelf runs short of options.

The constrained cut is more useful for what it surfaces than for its shares. Five Dutch entities qualify here that do not qualify in the general cut: Veloretti, Bakfiets.nl, Dutch ID, Lovens and Santos. These are brands the assistants know but rarely volunteer. Asking for a Dutch brand does not merely reweight the list — it produces a partly different list.

4.3 Reading the two together

The honest summary is that the origin clause changes the answer a great deal here, unlike in some other markets this series has measured. Dutch entities hold 40.0% of the origin-neutral top ten and 90.0% of the constrained top ten. Anyone quoting one figure without the other, or without the question count behind it, is describing a question set rather than a market.

5. Differences Between Models

One-sentence takeaway: All five models put Gazelle first among Core entities, but they disagree by a factor of nearly three about how many entities an answer should contain and by nearly 60 points about whether to search the web at all.

5.1 Per-model behavior

Core entities named per response, by model.

Grok names 9.10 Core entities per response and mentions 127 distinct Core entities across its 225 answers. GPT-5.6 names 3.25 and touches 62. Claude Haiku is the narrowest on Core entities at 2.87. These are answer-format differences, not disagreements about the market, and they are the reason breadth carries only a quarter of the composite: an entity that appears in Grok’s long lists and nowhere else has a very different visibility profile from one all five models volunteer.

5.2 Grounding: one model barely searched

Three models grounded 100% of their answers, GPT-5.6 grounded 86.2%, and Gemini grounded 41.8% — meaning roughly 131 of its 225 answers were produced without retrieval. The overall rate of 85.6% is arithmetically correct and individually misleading for one model in five.

Web-grounded share of each model’s own answers, against the 85.6% overall rate.

This matters for interpretation rather than for the scores. Gemini’s leaderboard is not obviously stranger than the others: it names Gazelle in 50.7% of its answers, Cube in 32.9% and Batavus in 30.7%, an ordering close to Perplexity’s. What differs is where the information comes from. For the ungrounded portion, Gemini is describing what its training data recorded about Dutch bicycle brands, not what the Dutch web says today — and a brand that changed hands, launched or failed recently is exactly the kind of entity that reads differently under those two regimes.

5.3 Each model’s most-named Core entities

Gazelle is first in all five models, but the spread of that first place is enormous: Grok names it in 78.7% of its answers (177 of 225) and Claude Haiku in 24.4% (55 of 225). A reader who used only Claude would see a market in which the leader appears in one answer in four; a reader who used only Grok would see one in which it appears in four answers in five. Both are this study’s data.

The second and third places diverge more. Batavus is second in Claude, Perplexity and Grok; Cube is second in Gemini; Giant is second in GPT-5.6 on 75 of its 223 responses (33.6%), a position it holds in no other model. Sparta is fourth in Perplexity and Grok and does not reach the top eight in Claude, Gemini or GPT-5.6.

5.4 When a rank is really one model’s habit

Ten entity-cut combinations qualify while appearing in only two of five models. Their visibility is model-specific and should not be read as market-wide. They are published in the leaderboards with their breadth column intact for exactly this reason.

Two cases are worth naming. Puch qualifies 21st of 22 in the attribute cut on 23 of 375 responses from two models — a legacy marque the models associate with reliability questions, not a live presence on the Dutch shelf. Lapierre and Rose both qualify in the sport segment on two models each, in a cut of only 200 responses where 11 responses is enough to clear the threshold. No entity in the study qualifies anywhere on fewer than 10 responses, and none qualifies with an MRR above 0.9, so the thin cases are thin in breadth rather than in volume.

6. Question-Type Ownership

One-sentence takeaway: Eleven entities qualify across all three question types, and the entities that qualify in only one reveal what each kind of question is actually asking for.

The 45 questions divide evenly into three types: 15 discovery (“which brands should I look at”), 15 attribute (“which e-bikes have belt drive / a big range / good service”) and 15 use case (“I commute 25 km each way in headwind”). Denominators are 375, 375 and 373.

Gazelle leads all three. Eleven entities qualify in all three types: Gazelle, Cube, Batavus, Trek, Giant, Riese & Müller, Koga, Sparta, Kalkhoff, Specialized and Tenways. That group is the durable core of this market’s AI visibility — entities the assistants will name whether asked for a recommendation, a specification or a scenario.

The single-type qualifiers are more informative:

• Discovery only — Decathlon, Veloretti, Flyer, and the Decathlon sub-brands Elops and Van Rysel. These are names that surface when someone asks an open “where do I start” question and disappear the moment the question acquires a constraint.
• Attribute only — Cowboy, Dutch ID, ENGWE, Fiido, Puch and Stevens. Direct-to-consumer and specification-led brands: they appear when the question is about belt drive, a quiet motor or a removable battery, and not when it is about a commute.
• Use case only — ADO, Bakfiets.nl, Brompton, Lovens, Phatfour and Stromer. Every one of these is tied to a specific journey: a folding bike for the train, a cargo bike for two children, a speed pedelec for 35 km each way. Brompton reaches an MRR of 0.7792 in the use-case cut — the highest MRR of any qualifying entity anywhere in the study — on 20 of 373 responses.

There is also a structural point buried in the attribute cut. Because all 15 attribute questions sit in the everyday segment, the attribute leaderboard and the everyday leaderboard share half their evidence, and the two cuts should not be read as independent confirmations of each other. Gazelle reaching exactly 100.00 in both is one result, not two.

Discovery qualifies the most entities (28 of 375 responses) and attribute the fewest (22). Open questions spread mentions across a wider field; constrained ones concentrate them.

7. Segment Leaderboards

One-sentence takeaway: All 15 attribute questions fall in one of the three segments, so these leaderboards are indicative only and their entity counts must not be compared with each other.

Read the limitation first. The segments are a cut of the question set, not an attribute of the entities, and the cut is unbalanced by design rather than by accident:

All 15 attribute questions sit in Everyday & commuting, because attribute questions in this vertical ask about generic e-bike properties — motor, battery, belt drive, brakes, service — rather than about a product format. The other two segments carry eight questions each and zero attribute coverage. Three consequences follow, and they apply to every number in this chapter: segment leaderboards are indicative; distinct_entities must never be compared across segments; and the everyday segment’s larger qualifier field partly reflects its 29 questions and 724 responses rather than a broader market.

7.1 Everyday & commuting (29 questions, 724 responses)

Everyday & commuting — indicative.

Gazelle reaches 100.00, holding both maxima on 461 of 724 responses (63.67%). Eleven of the 18 qualifiers are Dutch — the highest domestic share of any segment — and this is the only segment where 12GO Biking, a retailer, qualifies at all (37 of 724, 5.11%, on three models).

7.2 Sport & recreational (8 questions, 200 responses)

Sport & recreational — indicative.

One of the two cuts Gazelle does not lead. Cube leads at 97.74 on 128 of 200 responses (64.00%), ahead of Trek (89.23) and Giant (80.68). Only 4 of the 19 qualifiers are Dutch, against 15 foreign — an inversion of every other cut. Gazelle still places fifth on the position maximum (MRR 0.6217) despite a mention rate of just 16.50%: when road and gravel questions mention it at all, they mention it first, and then move on.

7.3 Cargo, family & special formats (8 questions, 199 responses)

Cargo, family & special formats — indicative.

Urban Arrow leads at 85.79 on 72 of 199 responses, narrowly ahead of Gazelle (84.08) which is named more often (100 of 199) but later. Twenty-one of the 37 qualifiers are Dutch, and the specialist Dutch cargo names — Carqon, Bakfiets.nl, Lovens, EMQ, Brekr, Dolly, Babboe, BOHLT — appear here and almost nowhere else.

37 qualifiers on 199 responses is arithmetic, not market structure. The qualification threshold is a rate, so a 199-response cut needs only 10 responses to clear 5%, against 57 in the general cut. The cargo segment is not twice as diverse as the general market; its threshold is five and a half times easier to reach. Compare mention rates and response counts across segments, never qualifier counts.

8. Components and Adjacent Brands

One-sentence takeaway: The most-mentioned entity in this entire study is a motor manufacturer, and it is deliberately not on any leaderboard.

Thirty-two entities are classified Adjacent: motors, drivetrains, brakes, tires, batteries and accessories. They are matched, measured and scored on the same published scale as everything else, and they are excluded from every leaderboard, because a shopper choosing between a Gazelle and a Batavus is not choosing between them and a Bosch.

Bosch appears in 626 of 1,123 responses (55.74%) — more than Gazelle’s 594. Shimano appears in 460 of 1,123 (40.96%). Gates, the belt-drive manufacturer, appears in 204 (18.17%), which on responses alone would place it ninth among the nineteen qualified Core entities. Enviolo, the Dutch hub-gear maker, appears in 188 of 1,123 (16.74%). Five Adjacent entities clear the 5% threshold.

Because the scale is drawn from qualified Core entities, Bosch’s c_mention is 105.39 — above the ceiling by construction, reported and never clipped. Reading this correctly matters: it does not mean Bosch is better than the best bicycle brand at anything. It means Bosch is named in 626 of 1,123 responses against the 594 of 1,123 that define 100 on that axis, and that a scale built on bicycle brands cannot contain it.

What this says about how assistants answer Dutch e-bike questions. In this market the models answer specification questions through the drivetrain. Asked which e-bike has a large real-world range, or which needs the least maintenance, the assistants reach first for the motor and transmission and only then for the bicycle around them — Bosch is named in 131 of Gemini’s 225 answers and 169 of Grok’s. The consequence for brands is structural rather than competitive: a substantial share of the specification-level authority in AI answers about Dutch e-bikes accrues to two or three component suppliers that no consumer buys directly.

One reclassification is worth stating. Maxi-Cosi was moved from Out-of-scope to Adjacent. It makes no bicycle product, but it is named as a child-seat fitting on cargo bikes, which is the same relationship a motor has to the vertical. It appears in 22 responses.

9. The Source Ecosystem

One-sentence takeaway: Retailers and consumer-testing publications, not brands, supply most of what AI assistants cite about Dutch bicycles — brand-owned sites account for 15.6% of citations, the lowest share this series has recorded.

750 distinct domains were cited across the 1,123 analyzed responses. 479 reached the floor of ≥3 responses, carrying 95.6% of all citation volume, and 100% of those were classified. The classified set carries 7,561 cited responses in total.

Retailers supply 41.4% of citations from 150 domains, and editorial 34.4% from 180. Brand-owned sites account for 15.6% — the lowest brand-owned share this series has measured in any market. Marketplaces contribute 2.2% from 13 domains, consistent with their near-total absence from the leaderboards. The other bucket sits at 2.6%, well inside the 25% ceiling that would have triggered a reclassification pass.

The most-cited single domain is 12gobiking.nl at 373 responses — cited in one answer in three, by a retailer that appears by name in only 46. Its own dictionary entity does not clear the general threshold. A shop can be the web’s answer to a question without being the assistant’s answer to it.

The next two are consumer organizations: anwb.nl at 321 and consumentenbond.nl at 221. Together with productscore.nl (92) and elektrischefietser.nl (61), the testing-and-comparison layer of the Dutch web is doing a large share of the assistants’ work for them.

121 of the 479 classified domains resolve to an entity in this study’s dictionary, carrying 4,056 of the 7,561 cited responses. Of the 89 brand-owned domains, 22 belong to entities that qualify in the general leaderboard, led by gazelle.nl (97), giant-bicycles.com (54), batavus.com (51) and koga.com (50). Every one of those numbers is an order of magnitude below the retailer and testing domains above them.

What this chapter cannot show is influence. A citation is a link the model attached to an answer, not a demonstration that the link shaped the answer, and 41.8% grounding on one model means a substantial share of this study’s text was produced with no citations at all. The source mix describes the documented half of the channel. It is the best available proxy for where AI assistants learn about Dutch bicycles and it is not the same thing.

Two caveats belong with this chapter. 42 of the 479 domains carry a Low classification confidence, and the largest unattributed source is productscore.nl, the 14th most-cited domain in the study at 92 responses, whose operator could not be identified — it is classified editorial on the basis of what it publishes rather than who runs it. And source classification is derived from the entity dictionary wherever an entity owns the domain, so a business that is both a cited source and a named entity is never counted twice.

10. What the Patterns Suggest

One-sentence takeaway: Five readings are available from this data, and each has something in it that cuts the other way.

A concentrated top and a flat middle. Gazelle’s 21-point lead is the largest gap on the board, and below third place thirteen entities sit inside eighteen points. Reading: AI visibility in this market is winner-heavy at the summit and unstructured beneath it. Against it: the flat middle is partly a scaling artefact — composite scores compress once c_breadth saturates at 100 for every entity named by all five models, which is 17 of 19 here.

Origin matters more than the domestic share suggests. Dutch brands take 40.0% of the origin-neutral top ten, which is a strong showing for a market of this size against German, American and Taiwanese competition. Against it: Cube outmentions Gazelle on origin-neutral questions by one response in 848, and five Dutch entities qualify only when the question asks for a Dutch brand. Domestic visibility here is real but conditional, and the condition is the wording of the question.

The specification layer belongs to suppliers. Bosch and Shimano are named in 55.74% and 40.96% of all responses, and the assistants answer attribute questions through the drivetrain. Reading: component brands hold a share of AI-mediated authority disproportionate to their consumer relationship. Against it: this is partly an artefact of question design — 15 of 45 questions are attribute questions, and they concentrate in one segment.

The information layer is not brand-owned. Brand sites supply 15.6% of citations while retailers and consumer-testing publications supply 75.8% between them. Reading: what an assistant says about a Dutch bicycle brand is mostly assembled from what other people wrote about it. Against it: citation share is not influence share — this study measures which domains a model cited, not which shaped its answer, and 41.8% of one model’s answers cited nothing at all.

Retail is the shelf, and it is specialist. Two pure retailers qualify in the general cut and marketplaces qualify in one cut out of nine, while retailer domains supply 41.4% of citations from 150 domains. Reading: in this market the assistants route discovery through specialist bicycle sellers rather than through general commerce platforms, which is close to the opposite of what this series has recorded elsewhere. Against it: the question set contains several explicit purchase and second-hand questions, which invites retailer answers, and one of those retailers — 12GO Biking — supplies the single most-cited domain in the study while appearing by name in only 46 of 1,123 responses.

No recommendations follow from any of this. This is a measurement, not a strategy document.

11. Limitations and Notes

Single collection window. 1,125 responses were collected on 5 September 2026 and 1,123 analyzed. This is a snapshot, not a trend. It is the first Netherlands edition, so no cross-edition comparison exists and none is implied.

Tenways is a boundary case, and it moves a published figure. Its EU class-12 marks were filed in 2021 by a Hong Kong company and transferred to a Dutch B.V. in a proprietor change published on 29 September 2025, eleven months before collection. Under the declared trademark rule Tenways is Local and is published that way. Readers who think of it as an Asian brand are describing its founding, not its current EU trademark ownership; its UK and Swiss marks remain with the Hong Kong entity, so the change is jurisdiction-specific. Sensitivity: the origin-neutral cut has 7 qualified Dutch-origin entities including Tenways and 6 without it.

Amslod moved the other way. Its EU mark passed to a Czech company after the 2025 bankruptcy of the Dutch operating entity, so a brand many Dutch consumers would call Dutch is classified Foreign. It does not qualify in any cut.

One brand the assistants describe as Dutch is not. The answers repeatedly presented Ecobike as a Dutch brand. It is Polish. The Tilburg business the models were describing is Ecobike Fietsen, its Dutch dealer. The classification follows the trademark, not the answer text, and this is a case where the models are confidently wrong about a fact this report measures.

Ten qualifiers rest on two of five models. Listed in full in Chapter 5.4. Their visibility is model-specific. Every one is published with its breadth column intact rather than being quietly dropped, and none rests on fewer than 10 responses.

Segment design is unbalanced. All 15 attribute questions fall in Everyday & commuting; the other two segments have zero attribute coverage and eight questions each. Segment leaderboards are indicative and distinct_entities must not be compared across segments. This is the most significant structural limitation in the edition.

Small cuts qualify more entities. The 199-response cargo cut qualifies 37 entities against 19 in the 1,123-response general cut, because the threshold is a rate. That is arithmetic.

Low mention rate with high MRR ranks surprisingly well. Urban Arrow holds the general-cut position maximum on 80 responses at a mention rate of 7.12%. The composite is working as designed, but the raw numbers should be read alongside the score, which is why both are printed in every table.

Sub-brands may structurally under-count. Models frequently name a parent where a manufacturer or a consumer would name a sub-brand. Decathlon’s sub-brands are visible here (Elops, Rockrider, Van Rysel) only because a backstop pass restored them after the extraction rules demoted them.

One extraction gap was structural. Riese & Müller was never generated as a candidate, because its name contains an ampersand. It entered the dictionary as a promoted merge target and finished 8th. Similar names in future editions carry the same risk, and this is now a known check.

Four dictionary entities matched nothing. All are Out-of-scope non-brands; three of them are publications whose domains are cited but whose names are never written in the answers.

Nine entities have Unknown trademark origin. They are excluded from origin shares rather than counted as Foreign, and none of them qualifies in any cut, so no published share is affected.

Two identical response sets were inspected and left separate. Csepel and FabricBike each appear in the same 3 responses, and are unrelated brands.

Six of the nineteen ranked entities were in or through insolvency at collection. The entity dictionary records the corporate status of every brand it tracks, and the results are uncomfortable next to the leaderboard. Accell Nederland B.V. was declared bankrupt on 13 August 2026 — 23 days before this study’s collection window — and brand continuity for Batavus (3rd), Koga (9th) and Sparta (10th) was unresolved when the questions were asked. VanMoof (15th) was acquired after its 2023 bankruptcy. QWIC (17th) went bankrupt in November 2023 and was relaunched by Ecomotion in 2024. Stella (18th) went bankrupt in November 2024 and relaunched in 2025 with a smaller store network. In the cargo segment, Carqon sits inside the same Accell proceedings and Babboe cargo bikes are not currently sold in the Netherlands at all, yet Babboe qualifies 32nd of 37. Union, which ended new-bike production in 2025, is tracked but does not qualify.

This is a finding about the assistants, not an error in the data. The models named these brands because the web is full of them, and a bankruptcy three weeks before collection is exactly the kind of event a language model’s training data cannot contain and a web index absorbs slowly. It is also the sharpest available illustration of what this report measures: visibility is not viability, and a reader treating this leaderboard as a shopping list would be pointed at several brands whose manufacturer was in insolvency proceedings on the day the questions were asked.

Citation volume and naming are different measurements. 12gobiking.nl is cited in 373 of 1,123 responses while 12GO Biking is named in 46. A domain can carry an answer without the organization behind it being named in that answer, and this report measures the second thing, not the first.

The question set is fixed and partial. 45 questions cannot cover a market. They were written to span discovery, specification and scenario, and three product formats, but they are not a sample of Dutch search demand and no weighting was applied to make them one. A brand strong in a format the question set touches lightly will be under-represented here.

Never compare scaled scores across cuts or editions. A score of 100 in one cut is a different absolute level of visibility from 100 in another, because the scale moves when the leader moves. Cross-edition comparison must use mention_rate, mrr and breadth.

Frequently Asked Questions

12. Appendix & Data

12.1 All 45 questions, as asked

Questions are reproduced in Dutch exactly as they were sent, including colloquial phrasing and imperfect grammar. English glosses are editorial and were not used in collection.

12.2 All tracked entities (222)

Marketplaces and retailers are included in the index and in every published leaderboard. Adjacent entities carry a scopeFit of Adjacent and are excluded from leaderboards by design; Out-of-scope entities carry Out-of-scope and are excluded from both leaderboards and origin shares. The four entities that matched no response are listed separately at the end of the index with their websites, so that a reader can see exactly which names were tracked and returned nothing.

12.3 Data availability

The complete data pack behind this report is published alongside it: the full metrics table (one row per entity × cut), the entity dictionary with aliases and origins, the decisions log, the dropped-candidate list with reasons, the published scale registry, all nine leaderboards, the unqualified appendix, the adjacent-entity table, per-model breakdowns, the source classification, and the machine-readable summary this page reads from. Every figure in this report is reproducible from those files.

12.4 Methodology note

About Herm.io & disclosure

Herm.io publishes quarterly AI Visibility studies measuring how large language models answer real consumer questions in specific markets and verticals. This edition is independent research. No entity named in this report commissioned it, reviewed it before publication, or paid for placement. The publisher’s own domain was checked for in the cited-source data and does not appear, so no exclusion was applied.

Method, data pack and corrections are published in full so that any figure here can be re-derived. Corrections are logged rather than silently amended. Questions and error reports about this edition are welcome.

Author: Mert Can Elkaya · Edition: 1 (baseline) · Collection: 5 September 2026 · Published: 7 September 2026

Written by

Mert Can Elkaya

Contributor

I'm a product builder working at the intersection of product, fintech, and growth. From martech and venture capital to leading product at a proptech platform and co-founding a fintech startup, I help teams—and shoppers—make smarter, more confident decisions.

More from Mert →