BMW.
The €47,000 318i is configurable in a browser one time in three, and not at all by the simpler agent classes.
Found & recommended by AI agents
Search-class agents touched the close in some runs; the full agent-fleet access profile lands in a later wave.
This page measures the commerce lane: can an agent find BMW, and once it arrives, transact. Talent, after-sales, procurement, investor and press lanes run on different surfaces and are not yet measured.
Your brand isn't measured yet
What does an AI agent do with your website?
The same measurement as BMW, free for your domain. Five agent classes, one real task, your score in 48 hours.
Five agent classes, one full configuration. The price lives in Shadow DOM.
We asked five kinds of AI agent to configure a BMW 318i Limousine at bmw.de. The plain reader got an SPA shell, just JavaScript bootstrap code and font URLs, no prices. The coding agent loaded the configurator, the page title said Konfigurator, document.body.innerText returned an empty string: BMW renders inside Shadow DOM web components, no standard DOM query reaches them. The full browser configured the car correctly once in three runs, blocked the other two times by an accessibility API the surface does not expose over CDP.
One label, many breeds. From a plain reader to an autonomous operator, the kinds behind ChatGPT, Perplexity, and Claude Code:
Blocked somewhere on the path: Plain reader, Coding agent. A customer whose assistant runs on one of those breeds never finishes the task.
Scope. This is one model. BMW offers 40+ models across 1, 2, 3, 4, 5, 6, 7, 8 Series, X-models, M-Sport, i-electric, and Z roadsters, plus financial services, used-car BMW Premium Selection, and the dealer network.
Found, and able to transact?
Two questions, measured separately. A brand can be recommended and still un-buyable, or perfectly buyable and never found.
When someone asks an agent to configure a car, does it route to BMW?
BMW comes up about a third of the time. The rest of the time, an agent recommends an alternative first.
Discoverability · 18-datapoint auditOnce an agent is on BMW's site, can it book a test drive or place an order?
Search-class agents surface config-ready in some runs; the full agent-fleet access profile lands in a later wave.
Evidence · 65 / 100 A measure of how provable and consistent the result is, grounded in cross-method ground-truth agreement (the methods that ran returned the same price), not a separate measured run. A confidence layer on the two scores above, not a third sales axis.
What this means for BMW.
On product discovery, BMW is not lacking a configurator. It lacks an agent-readable surface in front of it. The 318i is configurable, the price is €47,000, the runs that succeed prove it. But the encapsulation choice that makes the experience clean for humans, Shadow DOM web components plus an accessibility surface only the most capable browser can reach, blocks two of the three measured agent classes outright and gives the third a one-in-three pass rate. Discovery is not the gap. Reach into the surface is.
The fix is not unwinding the web-component architecture. It is a parallel agent-readable surface alongside it: Reference Pages that carry the configured price and the variant facts in plain HTML, Answer Pages for the buying questions a customer asks before the configurator opens, an Evidence Layer that lets an agent confirm a configuration without rendering the Shadow DOM at all.
The proof is not a redesigned configurator. It is plain readers and coding agents extracting the €47,000 318i price directly, and browser agents passing three of three instead of one of three, re-measured each wave.
Commission an audit.
Where the BrandScore opens the question, an audit closes it. An interpretive engagement on your full surface, scored under the same methodology.
Get the auditFound with us.
Strategic partnership for brands building agent success as a long-term capability, not a one-off engagement.
ApplyAudit an adjacent property.
The BrandScore covers the primary domain. Get the same methodology applied to an adjacent property: a country site, a sub-brand, a category beyond the DAX-40 slate.
Get a SnapshotThis is BMW. What about your brand?
You just saw how an AI agent treats a DAX 40 brand.
The same measurement runs free against your domain: five agent classes, one real buying task, your Agent Success Score in 48 hours, in the exact format of this page.
No intermediary stands between agents and BMW. The gap is being found, not the channel.
Channel position pending the Wave 4 Cody Gate-1 audit response analysis. The pre-measurement expectation, based on the wider automotive market, is that used-car intent leaks to autoscout24 and mobile.de while new-car configuration stays on bmw.de; the measured Wave 4 task is a new-car configuration, so any displacement found will speak to the configured-buy lane specifically, not the broader brand.
Hyperize-selected tasks.
One task from the public sector grid. Task list is frozen before each wave runs.
BMW 318i Limousine
- Close state
- config-ready
- Bottleneck
- Configurator hidden in Shadow DOM web components; the plain reader and coding agent are blocked, browser passes 1 of 3 runs.
Fairness note
Wave 4 Q2 2026 audit complete on a single task (BMW 318i Limousine, config-ready close state), scored under the public Task Selection Doctrine. AI Visibility from Cody Gate-1 on an unbranded informational probe matching the Mercedes-Benz Wave-3-re-audit probe verbatim (apples-to-apples cross-brand comparison). AI Usability derived from a fleet wave phases 1-3 (ceiling + http + coding + browser-agent, 2026-03-29); the act-phase was not captured in a fleet wave for BMW, so the full_automation row of the agent_matrix is marked pending and the derived usability is summarised from 3 measured breeds. Fairness Review pending the sector fairness grid.
How the score was produced.
Discoverability is audit-pipeline-derived. 3-provider sample (openai, perplexity, anthropic), 3 query variants per task, 2 runs per variant · 18 valid datapoints scored against a five-state handoff cascade. [S1]
AI Usability is derived from the access-profile above (usability-derivation/v1): how far the best agent reached (close state) modulated by how many agent classes succeeded. The per-class profile is the truth; the score is a reproducible summary of it, not a separate rating. Fleet phases (HTTP / Coding / Browser / ACT) produce the profile. [S2]
Agent Success Score = (AI Visibility × 0.20) + (AI Usability × 0.70) + (Evidence × 0.10)
On a 0–100 scale, displayed 0–10. AI Usability bundles the agent's reach + completion; AI Visibility is audit-derived discoverability. Weighting is public; the per-prompt derivation is not.
Measurement scope
Confidence C · one measured task on a 3-provider track (openai, perplexity, anthropic). Confidence promotes to B with a second task plus a fourth provider on the next wave.
Measurement timeline.
Each wave appends; nothing overwrites. Frozen Wave Rule.
- Entry · 01
29 Mar 2026
Wave · Protocola fleet wave
fleet/measured
First-pass fleet measurement: BMW 318i Limousine (GT €47.000) across ceiling + http + coding + browser-agent breeds. Found Custom Web Components Shadow DOM blocking plain readers and coding agents; browser-agent achieved 1 of 3 successes.
- Entry · 02
24 May 2026
Wave · ProtocolWAVE-Q2-2026-W4-AUTOMOTIVE
ars-methodology/v1.1
Wave 4 kickoff. Cody Gate-1 audit running on unbranded informational probe (same probe as the Mercedes-Benz Wave 3 re-audit, for apples-to-apples discovery). AI Usability carried from a fleet wave access profile, no act-phase data for BMW in that wave so the full_automation row is marked pending.
Evidence and provenance.
Public methodology references and internal evidence pointers behind every claim above.
- [S1]Accessed · 24 May 2026
Gate-1 audit run · BMW Wave Q2 2026 (Cody, unbranded informational)
Internal · Hyperize evidence
- · AI Visibility score (audit-derived, 18 datapoints, 3 providers)
- · AI platforms queried (openai/perplexity/anthropic, 3-provider track)
- · the close state reached (config_ready)
- [S2]Accessed · 29 Mar 2026
a fleet wave automotive wave · BMW phase 1-3 (Giorgio repo)
Internal · Hyperize evidence
- · the per-breed access profile (text/code/browser observations)
- · the Shadow DOM finding (Custom Web Components, accessibility API unavailable)
- · the browser-agent pass rate (1 of 3 runs)
- [S3]Accessed · 29 Mar 2026
Ground-truth ceiling run · BMW 318i Limousine (€47.000)
Internal · Hyperize evidence
- · Ground-truth price €47.000 (Schwarz uni 0€, 17-inch Serienfelgen 0€)
- · 12-step konfigurator path
- Accessed · 24 May 2026
Public · hyperize.ai
- · fairness declaration
- · Third-Party Interception framing
Read the doctrine. Challenge the score. Extend the slate.
Task Selection.
The fairness doctrine behind the slate above. Five failure modes, six criteria, public before each wave.
Read ChallengeDisagree with this score.
Send evidence under public Fairness Review. Failed reviews are documented with the named failure mode.
Challenge ExtendSubmit your own task.
Open Surface Run · additive measurement. The Hyperize-selected slate stays frozen; your task gets the same methodology.
SubmitEditorial coverage
The DAX 40 Agent Success Index is a point-in-time snapshot of the agent-success of public digital touchpoints. Results are not statements about product quality, company performance, service quality, or the legal obligations of the brands named. Brand names and logos remain the property of their respective owners and are used solely for identification and reporting purposes in the context of editorial coverage (§ 23 MarkenG, Art. 5 GG).
Brands wishing to respond, engage, or correct a factual error may contact hello@hyperize.ai. Responses received are published in full alongside the findings. Full methodology and editorial-coverage notice: coverage statement.
