What Ramp learned when they marketed to AI agents

Marketing built for humans meets agents

Ramp’s framing starts from a simple observation: websites were designed for people scrolling and clicking. As more decisions get delegated to machines, companies need to know how to get agents’ attention, speak to what matters to them, and incentivize the right actions.

That is not the same problem as classic SEO. It is a comprehension and format problem: can an agent that lands on your domain parse what you sell, and does the format you serve help or hinder that?

Ramp set out to measure that in production, not as a ranking contest, but as a practical test of how agent-facing content behaves when real LLM products are asked prospect-style questions.

What Ramp tested

Across roughly 50 marketing pages, Ramp ran three concurrent content-format tests with identical tracked offer text and per-variant tracking:

• Markdown
• Stripped HTML
• Schema-heavy variants

They used Cloudflare Workers to conditionally serve bot-targeted content, then queried major LLM products and watched whether the tracked offer surfaced in responses.

Their headline format result: Markdown was the only tested format that reliably surfaced in LLM responses.

Findings summarized from: Ramp Builders, Marketing to AI agents

Model behavior diverged sharply

Ramp’s results also carry an important caveat we preserve: model behavior was not uniform.

In their published observation, Claude relayed the tracked offer with specificity, Perplexity surfaced it vaguely, and ChatGPT did not surface it over the 32-day observation window despite bot reads.

Agent visibility is model-specific. A single score or a single snapshot does not explain every AI tool, which is why progress should be tracked over time.

Ramp’s own published result also changed materially between an earlier checkpoint and a later view. Time-series measurement was more useful than treating one point in time as the whole story.

What this validates, and what it does not

This supports a careful claim: agent-readable infrastructure is worth testing, and Markdown-first agent content is a strong empirical signal from a real B2B SaaS production experiment.

It does not support guarantees of AI citations, rankings, revenue, or conversions, and we do not make those guarantees either. Ramp’s study is one company’s experiment; it should not be read as proof that every site in every industry will see the same outcomes.

One more boundary matters for how we deliver: Ramp’s methodology used conditional bot-targeted content. AgentReadyWebsite installs publicly accessible static files at well-known paths. We do not recommend or implement cloaking or bot-only pages. Their research is useful evidence; their delivery model is not ours.

Where this overlaps our package

The strongest overlap with Ramp’s content finding is the Markdown-first surface: llms.txt and agenticweb.md give agents curated, readable entry points. Companion files such as robots.txt and sitemap.xml support crawl access and URL inventory.

We also ship HTML and schema surfaces publicly (llm-info.html, schema.jsonld) as complementary formats, not as mutually exclusive bot-only variants. Markdown helps LLM parsing; Schema.org JSON-LD still supports structured facts.

On the product side, the Agent Readiness Dashboard is designed around before/after progress over time (snapshots, deltas, what changed). That matches how Ramp’s published results moved across weeks.

For a complementary measurement angle from documentation-heavy software, see our SerpApi llms.txt case study.