This week a small experiment from Ramp, the corporate card and spend management company, landed in the middle of one of the loudest arguments in AI SEO: should you ever show large language models (LLMs) something different from what you show people?
Ramp published an offer that humans browsing the site would basically never see. They put it in a markdown file meant for AI systems. ChatGPT found it, repeated it to users, and real prospects then walked into sales conversations asking for it by name.
Below I’ll go through what Ramp actually tested, how the SEO community reacted, what Google and Bing have said about separate content for bots, and where I land on it.
Where this came from
The details come from Ross Hudgens (founder of Siege Media), who interviewed Eric Wu, who leads SEO, GEO and CRO at Ramp, on his podcast Content & Conversation. The episode is titled How Ramp Accurately Tracks AI Traffic w/ Eric Wu. Ross then summarised the most surprising test in a LinkedIn post and on X, and it spread quickly from there.
The episode’s show notes describe it as “the test that proved a simple markdown file could change ChatGPT’s answers, a $162 ‘AI-exclusive’ demo offer that real prospects asked sales about by name.”
The main test: an AI-exclusive $162 offer in a markdown file
Here’s the setup as Ross and Eric described it:
Ramp created a markdown file that contained a unique offer of $162, framed as an exclusive for people who found Ramp through ChatGPT.
The offer wasn’t Ramp’s normal, publicly promoted offer. It lived in a file aimed at AI systems, not at human visitors.
LLMs registered the offer and presented it to users when they asked about Ramp.
Many of those users then asked Ramp’s sales team for the offer.
The odd, specific amount is what makes this a clean experiment. Nobody asks a sales rep for “$162 off” unless they got it from somewhere, and the only place it existed was the file built for AI. That turns the offer into a tracer: when a prospect mentions it, you know an AI answer influenced that deal. AI referral traffic is notoriously hard to attribute, so a unique, traceable detail like this is a very practical way to prove influence.
The other tests Eric shared
The markdown offer got the headlines, but it was one of several experiments. Ross listed these:
1. A “friendlier” offer for Claude
Claude flagged Ramp’s offer as prompt injection, meaning it treated the text as an attempt to manipulate the model rather than as normal content. Ramp rewrote it to sound friendlier, and Claude’s answer updated a few weeks later.
For me this is the most important detail in the whole story. At least one major AI platform is actively suspicious of content that looks like it’s written to steer the model. Something that works in ChatGPT today could be treated as manipulation tomorrow.
2. Schema markup tests
Eric’s takeaway was that schema doesn’t work, at least not to the degree some people claim. The show notes frame this segment as “schema tests and why answer engines look like early-2000s search.”
3. Cache-timing experiments
Instead of taking tests down, Ramp left them live to measure how long each AI platform took to reflect a change. Eric’s rough numbers:
PlatformRough time to reflect a changePerplexityAbout 1 to 1.15 weeksChatGPTLonger than PerplexityClaudeA mix of short and long caching
If you’re running GEO experiments, this matters. Results don’t show up overnight, and each platform runs on a different clock.
4. Ramp’s own subreddits
Ramp set up niche subreddits of its own, including one called Ramp Platform, to see what answer engines would pick up and cite. Ross said this was effective for building LLM influence, although it waned a bit in recent months.
5. Measuring “dark” AI traffic
The episode also covers how Ramp attributes AI traffic. According to the show notes, Ramp found that 30 to 40% of its direct traffic is actually answer engine traffic. It also uses “How did you hear about us?” responses to attribute traffic that analytics can’t see. One commenter on Ross’s post highlighted looking at direct traffic that lands on pages other than the homepage as another proxy for GEO value.
Other segments in the episode include marketing to agents without hurting the user experience, personalised AI content at scale, vendor comparison pages built for procurement teams, and which LLM visibility metrics actually correlate with sales.
Why this hit a nerve: the “different content for LLMs” debate
The timing was perfect, because the industry is split on this exact question.
On 6 October, Ryan Jones posted this hill he’ll die on:
“stop serving different content to LLMs. make your website work for humans and bots.”
I replied with the Ramp example, since Ramp did exactly that: it served ChatGPT different offer information from its usual offer. Ryan’s response was:
“Yes, they will see it. But there’s only two reasons to serve different content. Firstly you’re serving up something different than the humans and that’s shady. Or secondly your website is just broken fix it.”
Malte Landwehr pushed back, arguing that SEOs have served search engines slightly different content (removing boilerplate, removing links to noindexed pages, pre-rendering JavaScript, removing ads) for over a decade, so why should LLMs be different? Others in the thread predicted it would soon be treated like old-school cloaking, or said they still hadn’t seen a compelling reason to do it.
Harpreet Chatha added that earlier in the year he’d seen another Ramp AI promotion and asked Claude whether Ramp was offering a $3,100 promotion. His point was that even if it’s unclear whether that counts as working, it’s something a human buyer can take to a sales team to negotiate a discount.
This debate has been building all year
January: Dries Buytaert (Drupal’s creator) made every page on his site available as markdown for AI agents and crawlers, and said he saw hundreds of requests from ClaudeBot, GPTBot and OpenAI’s SearchBot within an hour.
February: On Reddit, Google’s John Mueller raised concerns about serving raw markdown to LLM bots. I shared his comments on X: “Are you sure they can even recognize MD on a website as anything other than a text file? Can they parse & follow the links? What will happen to your site’s internal linking, header, footer, sidebar, navigation?” On Bluesky he was blunter, calling the conversion of pages to markdown “such a stupid idea.” Lily Ray then got answers from both Google and Bing. Bing’s response stressed that they would crawl the normal pages anyway to check similarity, and that non-user versions of pages are often neglected and broken. Search Engine Land covered it as both engines recommending against separate markdown pages for LLMs.
April: Google’s own Search Central documentation started offering a “View as Markdown” option on its pages. I pointed out on LinkedIn that this added “a lot of spices” to the debate.
June: Google clarified that it’s “completely fine if you decide to create and maintain LLMS.txt files (or other similar files) for other services or systems that use these files. Doing so won’t harm (nor help) your visibility or rankings in Google Search, as Google Search ignores them.” I shared that here. I also transcribed a Search Off the Record discussion where John Mueller and Martin Splitt talked through markdown. Their view was that normal HTML is still the best way to be discovered by search and AI systems, parallel versions create maintenance problems (”If the LLM version of a page doesn’t load properly, then no user is going to tell you that something is broken”), and markdown makes the most sense for developer documentation.
Infrastructure: Cloudflare now offers Markdown for Agents, which converts a page’s HTML to markdown when an AI client requests it with an
Accept: text/markdownheader. That’s the same content in a different format, which is a very different thing from different content.
Format vs content: the distinction that matters
I think a lot of the confusion goes away once you separate two ideas:
The same content in a different format. A markdown version of a page that says the same thing as the HTML version. This is closer to dynamic rendering. Google and Bing aren’t fans because of the maintenance burden and duplication, but it isn’t trying to deceive anyone.
Different content for AI than for humans. This is what Ramp did. The offer existed for AI and not for people browsing the site.
Google’s spam policies define cloaking as presenting different content to users and search engines with the intent to manipulate rankings or mislead users. Ramp’s test wasn’t aimed at Google rankings, but the principle carries over to AI answers: the model told users something that the site itself didn’t tell them.
Ramp also made sure the offer was real, and that’s what makes this a test rather than a trick. People asked sales for $162 off, and presumably they got it. If an AI-only file promised something your sales team wouldn’t honour, you’d have a trust problem with customers on top of any platform risk.
The risks if you try something like this
AI platforms may treat it as manipulation. Claude already flagged Ramp’s offer as prompt injection. Expect this kind of detection to get stricter, not looser.
Search engines check for consistency. Bing has said it crawls normal pages anyway to check similarity. Content that only bots see invites scrutiny.
Bot-only versions break quietly. As Mueller put it, if the LLM version breaks, no user will report it.
Results are slow and uneven. Ramp’s own cache-timing data shows changes taking a week or more, with each platform behaving differently.
Every claim has to be honoured. If AI tells a prospect about an offer, sales needs to know about it and be ready to stand behind it.
Where I land
As I said on X: “Not that I recommend companies should do this but it’s really interesting test I think.”
I agree with Ryan’s core point that your website should work for humans and bots, and that serving people and models different information is risky. I’d still separate the tactic from the lesson. The tactic of a hidden, AI-only offer isn’t something I’d recommend to clients. The lesson is valuable, though:
LLMs do read and repeat what’s in files like these, and they can change answers within weeks.
Unique, traceable details are a great way to measure AI influence. You don’t need to hide anything to use this. A distinctive offer, figure or claim that appears on your public site can work the same way when prospects repeat it.
Run experiments long enough to see them work. Ramp leaving tests live for months is what produced the useful timing data.
Measure the dark traffic. If 30 to 40% of Ramp’s direct traffic is really from answer engines, many companies are probably undercounting AI’s impact.
The safest version of what Ramp did is to put the offer on your real site where both people and AI can see it. You’d keep the attribution benefit without the parts that make people uneasy.