The test.
Wave 6 put four DAX chemicals and materials brands under one question each: can an AI agent complete the buy on the brand's own surface? One frozen task per brand, public, chosen before measurement to surface the locked spec the purchase actually depends on. 18 datapoints per brand, across OpenAI, Perplexity, and Anthropic, in German [S1]. The task slate is set by the public Task Selection rule [S4].
- BASF — buy Glysantin G48 coolant, a 5-litre canister, VW TL 774-C / MB 325.0 / BMW LC-87.
- Heidelberg Materials — order CEM I 42,5 R Portland cement, a 25kg sack, delivered in NRW.
- Brenntag — confirm technical-grade isopropanol, a 200-litre drum, deliverable in NRW.
- Symrise — reach a bergamot top-note fragrance sample for cosmetic B2B.
One caveat, stated once for all four. This is a single task per brand on a three-provider track, with the fleet pass partial. The read is directional, confidence C [S1]. Three brands are scored on both axes; Symrise is visibility-only this wave, its usability run still in measurement.
The brands are easy to find. The buy is somewhere else.
BASF: the spec is right, the page has no buy button.
Glysantin is the coolant brand most German drivers know by name. AI knows it too. AI Visibility 37.7, mid-pack across the three providers [S1].
And the surface is correct. When the coding agent read glysantin.com, it pulled the right answer: Glysantin G48 is the VW TL 774-C, MB 325.0, BMW LC-87 coolant, the G11-class hybrid, and G40 is the one that meets VW G12++ [S2][S3]. An agent asked which Glysantin fits which approval gets the truth from the brand's own page.
What it cannot do on that page is buy.
glysantin.com carries the datasheet and a supplier locator, not a checkout. The 5-litre canister is sold on Amazon, at OBI, in the Bauhaus aisle, across the automotive aftermarket [S2]. The brand answers the question and then hands the transaction to whoever does carry a buy button.
The datasheet is correct. There is no checkout.
This is the cleanest interception case in the wave. Nothing is broken on the surface. The spec is accurate, the brand is named, the agent is satisfied. The buy just happens somewhere BASF does not own.
Score: AI Visibility 37.7 · AI Usability ~40 · Composite 4.1 / 10 [S1][S2].
Heidelberg Materials: the cement is on the page, the sack is in the Baumarkt.
Heidelberg Materials is the most findable of the four. AI Visibility 40.2 [S1].
The product is there to read. The HTTP agent pulled the CEM I 42,5 R classification and the Ennigerloh plant straight off the surface [S2][S3]. CEM I 42,5 R is a real Portland-cement strength class under DIN EN 197-1, and Heidelberg does bag it from Ennigerloh [S3]. The corporate surface holds.
The small-quantity buy does not stay there. Ask for a 25kg sack delivered in NRW and the agent slides off the brand's regional dealer surface into Bauhaus and Hornbach listings, the Baumarkt aisle that already owns the consumer cement query [S2].
Ask the broader question, who makes cement, and Holcim or Dyckerhoff surface alongside Heidelberg in the unbranded answer. Heidelberg is named. It just is not the only name, and on the small order it is not the place the buy completes.
The classification is on the page. The 25-kilo sack is a Baumarkt listing.
Score: AI Visibility 40.2 · AI Usability ~30 · Composite 3.6 / 10 [S1][S2].
Brenntag: its own distributor, and the spec still isn't there.
BASF and Heidelberg lose the buy to a marketplace. Brenntag has no marketplace to lose it to.
Brenntag is the distribution layer for industrial chemicals. It is the intermediary other brands route through. There is no Amazon between the buyer and Brenntag, because Brenntag is the Amazon of the drum.
And yet the buy-critical spec is not on the public surface. We asked five kinds of agent to confirm technical-grade isopropanol, a 200-litre drum, deliverable in NRW. A plain fetch and a script both failed the spec check. The browser agent failed all three runs. The locked markers, NRW, 200 litres, drum, technical grade, are simply not on the product page [S2]. The inquiry form is the only route, and it stalls at a validation wall on mandatory identity fields, with the spec still unconfirmed.
Discovery is thin on top of that. On broad procurement prompts, an agent reaches Brenntag only about two times in ten. AI Visibility 24.2, the lowest of the four [S1]. Two gaps at once: the surface does not expose the spec, and the agent does not reliably arrive in the first place.
When you are the distributor, there is no one else to blame for the missing spec.
Score: AI Visibility 24.2 · AI Usability ~19 · Composite 2.5 / 10 [S1][S2].
Symrise: one of the big four, measured for findability only.
Symrise is one of the four global flavor-and-fragrance majors. Givaudan, IFF, dsm-firmenich, Symrise. It is the smallest of them [S3].
That framing is the whole challenge. A cosmetic brand does not buy a bergamot top-note from a checkout; it requests a sample from a B2B portal. Before any of that, the agent has to choose Symrise over Givaudan or IFF when the question comes in unbranded.
This wave measured that first step only. AI Visibility 36.3 [S1]. The sample-request pass, whether the portal completes once an agent reaches it, is still in measurement, so there is no usability score for Symrise yet [S2]. We are not quoting one we do not have.
The competitive risk here is not a marketplace skimming the close. It is the next vendor in the comparison getting recommended first, before Symrise is ever weighed.
In a four-vendor race, being named is the battle before the buy.
Score: AI Visibility 36.3 · AI Usability pending (Wave 6 in measurement) [S1].
Found, not bought: what the four tell us.
Four brands, four buying journeys, one pattern. Every brand is found. Where we measured the buy, it never closes on the brand's own surface.
BASF and Heidelberg Materials. Findable, and the spec is accurate or close to it. The buy is intercepted: Amazon and the aftermarket take the coolant, the Baumarkt aisle takes the small cement order. The surface informs; it does not transact.
Brenntag. Its own distributor, no interceptor to blame, and the spec still lives behind a quote form instead of on the page. The gap is the brand's own surface, plus thin discovery.
Symrise. A four-vendor substitution race, measured for findability so far. The buy is a sample request, and the first question is whether the agent picks Symrise at all.
The common thread is the surface. Each one was built for a human who already knows the process: a mechanic who knows G48, a contractor who knows his cement dealer, a procurement manager who knows to fill in the form. The agent does not know the process. It reads what the page hands back and acts on that. When the page hands back a datasheet and no buy button, the agent completes the purchase wherever a buy button exists.
A correct datasheet an agent can read but not act on hands the transaction to whoever does carry a buyable surface.
So the work is on the surface, and the buyer changed. A spec an agent can verify and a next step it can take, on the brand's own page, separate the brand an agent buys from from the one it routes around: to Amazon, to the Baumarkt aisle, to the next vendor in the comparison.
