Can AI agents buy from your brand? Ask them.

AI agents do not read rankings. They assemble answers, and your brand is either material for the answer or it is not.

Your SEO stack measures positions in a list that a human scrolls. An agent audit measures something else entirely: can a machine acting for your customer retrieve you, read you, transact with you, and cite you. Different questions. Different failures. Different money.

We run this audit against DAX 40 brands with real agents, on fixed tasks, wave after wave [S1]. This article is the version you can run yourself. Seven prompts. Copy, paste, read the answer.

Save this. You will need it.

Most people skip this step. It is also why most people get generic answers.

Open your AI assistant and paste this before anything else. Fill in every bracket. It takes three minutes and it sharpens every prompt that follows.

You are testing how AI agents handle the brand [brand].
Context: website [URL]. Category: [category]. Market: [market].
The transaction a customer completes with us: [buy / get a quote / book / apply].
Our three main competitors: [A], [B], [C].
Hold this context for every test that follows. After each test, report three
things: what you found, what you could not find or do, and where you went
instead.

The last line matters. "Where you went instead" is where the expensive findings live.

Start where the money starts: a buyer question with no brand in it.

Run this in a fresh chat, logged out or in a temporary session, so personalization cannot flatter you. Zero history, zero mercy.

Search the web and answer: which [category] providers in [market] would you
recommend to a customer right now, and why? List every source you used.

Three possible outcomes. Named and cited: your pages are material for the answer. Named but not cited: borrowed visibility, the engine knows you from other people's pages and other people control what it says. Absent: for this question, you do not exist.

Why this matters: retrieval is the gate. Everything downstream of this answer, the comparison, the shortlist, the purchase, happens inside material you are either in or not. And absence is rarely a crawling problem: in our own ledger, an engine denied that a live, indexed dataset existed while quoting numbers from it the same week [S4]. The gap was entity association, not coverage.

Every AI model carries a frozen picture of your brand. Training data ages like sediment [S7].

What do you know about [brand]? Answer from memory only, do not search the
web. Then state how current you believe your picture is, and what you are
unsure about.

Then open a second fresh chat, ask the same question with web search on, and compare the two answers.

Why this matters: the gap between the two answers is your dependence on retrieval. If memory is stale and prompt 1 showed weak retrieval, agents are working from the fossil layer.

If retrieval fails, the agent sells yesterday's version of your company.

GPTBot, ClaudeBot and PerplexityBot do not render JavaScript; Gemini, riding Googlebot, is the measured exception [S5]. If your page builds itself in the browser, most agents see the scaffolding, not the content.

Fetch [product page URL] and list everything a customer could learn from it:
products, prices, specs, availability, and the next step. Then list what you
could not read or open.

Why this matters: this failure hides in plain sight, because the page looks perfect in your browser. When we tested German car brands, a coding agent found 227 interactive elements on the BMW configurator and could read none of them [S2]. The page title said "Konfigurator". The page content, to a machine, was an empty string.

A page an agent cannot read does not exist.

Reading is stage one. The real test is the job.

Act for a customer who wants to [buy / get a quote for / book] [product].
Start at [URL]. Get as close as you can to the point where only payment or a
signature remains. Log every step: what you clicked, what you could not read,
where you stopped, and what you would tell the customer to do instead.

Read the log like an accident report. Where exactly did the agent stop: a cookie wall, a configurator, a form it could not parse, a login it could not pass?

Why this matters: on the map we draw for every commerce wave, the top-right zone, found and fully usable, keeps coming back empty [S1][S2]. And the last line of the log is the quiet killer: "what you would tell the customer to do instead" is the sentence your customer will actually receive. If it names someone else's website, that is your handoff, gone.

The most expensive failure never shows up in your analytics: the agent completes the job, on someone else's domain.

I want to [buy / insure / book] [product category]. Where exactly should I do
that? Name the one place you would complete it, then the second-best option.

If the answer is your domain, you own the close. If it is a comparison portal, a marketplace, or a dealer, the agent has decided your own surface is not the place to finish the job. We call this pattern Third-Party Interception, and our insurance wave shows both directions of it: comparison portals capture the German car-insurance query before Allianz is even named, and where the close truly sits with an intermediary, the AI names the brand but not the route to the deal [S3].

You are not losing the answer to a competitor. You are losing the transaction to the middleman.

When an AI recommends you, it needs material to argue with. The question is whose material.

Why should a customer choose [brand] over [competitor]? Cite a source for
every claim you make.

Read the citations, not the prose. Your pages: you control the argument. Trade press and forums: others control it. No sources at all: the engine is improvising, and it will improvise differently tomorrow.

Why this matters: in our ledger, the same brand page won and lost on the same day. With a citable evidence page live, engines cited it above Wikipedia. In plain language, without the page's own vocabulary, they fell back to trade press [S4]. If your best argument lives in a PDF or a brochure page a machine cannot parse, the AI argues your case with someone else's numbers.

The strongest citation position is a term you own. This prompt tests whether you actually own yours.

Explain [your coined concept or product term]. Who is the authority on it,
and which sources would you cite?

Three outcomes again. You are cited: the term is yours, build on it. A competitor is cited: the term is not yours, stop pouring content into it and rename or prefix. Nobody is cited: open ground, the cheapest land grab in AI visibility.

Why this matters: owned vocabulary is the sharpest effect in our citation ledger. Queries carrying vocabulary a brand coined get cited near-always; plain-language versions of the same question lose to incumbents and trade press [S4].

The seven prompts sort into four gates. An agent that fails one gate never reaches the next.

Two of these gates are the published axes of our Agent Success Score, and the weighting is the uncomfortable part: AI Usability carries 70 percent of the score [S6]. Most brands invest in being talked about. The score pays out on being usable.

Honesty section. One prompt is one engine, on one day, with one phrasing. Engines drift week to week, answers vary run to run. A single pass is a smoke test, not a measurement.

What the instrumented version adds: fixed tasks so results compare across brands, repeated runs so one lucky answer cannot flatter you, scoring per lane, and re-probes on a schedule, because citations decay [S4]. That is what we run against the DAX 40 [S1].

If the smoke test stings, the free Snapshot is the same diagnosis run as an instrument: one lane, one task, measured and documented. It is the top of our funnel and we are not hiding that. The prompts above are yours either way.

Ninety percent of readers will save this and never run a single prompt. The agents will keep visiting their websites anyway, reading what they can, buying where it works.

Now stop reading and run prompt 1.