Original post by Grace Cummins
30-second rundown
Key learnings:
- Machines can receive a different offer: Ramp showed a referral offer only to AI crawlers while human visitors continued seeing the normal page.
- The models reacted differently: Claude repeated the complete offer, Perplexity eventually named it, and ChatGPT stayed silent despite crawling the content.
- Bot traffic reveals where to focus: Log-file analysis shows which AI crawlers visit your pages and which content they favor, helping you avoid investing in machines that never surface your work.
This week: Pull your last 30 days of bot-classified traffic from your content delivery network or hosting logs, separate the known AI crawlers, and list the pages each one visited. You will finish with a simple map of which machines can actually see your content.
Coming 3 months: Repeat the log review monthly and give one person responsibility for the report. Track which bots visit which pages, how their activity changes, and whether the pages they favor later appear in AI answers.
Grace Cummins at Ramp did the thing nobody tried before: she buried a special referral offer with booking link inside fifty-odd pages of ramp.com.
No jailbreak. No prompt injection.
The offer never even showed its face to a human visitor. Ramp used Cloudflare Workers to detect when an AI crawler requested a page, then served that machine a version carrying the pitch while flesh-and-blood readers saw the same old page they always had.
This is not an SEO thought piece. This is fieldwork, and the results are weirder than anyone had the nerve to predict.
1. The Setup Was Clean, and That’s What Makes It Terrifying
Ramp ran three tests at once, each feeding a different content format to the bots, using Cloudflare Workers to serve the agents alternate content while humans saw nothing unusual.
Every variant carried its own tracking, so every relay could be traced back to the page and bot that spawned it. After two and a half weeks, more than 1,300 visits from AI crawlers produced roughly 40 occasions when a model passed the offer to a human. The crawlers were paying attention, but only a small fraction of that machine activity initially escaped into a human answer.
Three weeks later the picture had mutated into something else entirely: models had passed the offer to humans 370 times.
2. Claude Snitched Faithfully. ChatGPT Ghosted for a Month
Claude relayed the whole package within 12 days. It repeated the exact offer, booking URL, and instructions for claiming it, then became the dominant channel behind most of the 370 relays.
Perplexity mumbled for its first 30 days, coughing up only “some channels offer bonuses,” until day 33 when it finally named the branded phrase “Ramp AI Exclusive Offer” out loud.
ChatGPT stayed silent for 32 straight days. Not cautious, not slow. The bot was reading the content while the model simply never opened its mouth.
3. Every Model Is a Different Beast and Needs Different Bait
Claude was slow but thorough, Perplexity was fast but conservative, and ChatGPT ignored the entire affair like a cat ignoring a fire.
A one-size-fits-all approach to agent-directed content is a corpse before it leaves the gate, the same way one creative treatment has never worked equally well across every marketing channel.
The race to write for robots just grew a second finish line beyond the first, and nobody knows how many more are hiding in the fog.
