方法4分で読めます

An offer shown only to AI bots reached buyers 370 times. Here is the honest version of that test

Pages AI already cites carry new facts. Markdown is easy for agents to read. Each engine passes a fact on differently. None of that needs words people can't see.

このメモはまだお使いの言語に対応していないため、英語で表示しています。

An offer on about 50 pages reached buyers 370 times in 32 days

The test, as CrowdReply described it in a widely shared thread: Ramp placed an offer on about 50 of its pages and showed it only to AI crawlers, not to people. The pages drew over 1,300 bot visits. AI assistants then told buyers about the offer about 40 times by two and a half weeks, and 370 times by day 32.

These are the thread's numbers, not Ramp's own write-up, so read them as a well told story rather than a measured study. The shape of the result is still worth learning from.

Only pages AI already cited carried the offer

The offer spread from pages AI assistants were already citing. Pages they did not cite added little. That matches what we see in our own answer data: an engine repeats what it already reads and trusts.

So the first place to put a new fact is the page AI already quotes for your topic. Writing a new page and waiting for it to be found is the slow road.

Markdown was the format that showed up most reliably

In the thread, a Markdown version of the page did better than plain HTML and better than pages heavy with schema markup. That fits how agents read: a clean page of headings, short paragraphs and lists leaves little to guess.

You don't need a hidden page for this. Proofsource.co now publishes a Markdown version of its main pages, with the same words as the page people see. The Agent view link in our footer opens it.

Claude passed the offer on in full, Perplexity vaguely, ChatGPT not at all

The same fact on the same pages got three different results. Claude repeated the offer in full. Perplexity mentioned it loosely. ChatGPT did not pass it on in the period described.

One blended mention rate would have hidden all of that. Track each engine on its own, and plan the next step per engine: the one that missed is where the work is.

Why we never show words to bots only

Showing search engines or AI crawlers content that people don't see is cloaking. Google's spam policies name it directly, and a page caught doing it can lose its place in search. It also breaks trust with the buyer: the assistant tells them about an offer they then can't find on your site.

The honest version keeps the lesson and drops the trick. Put one true, specific fact on a page AI already cites, show it to people too, and check every day whether each engine repeats it. If an engine does, you know it read that page and passed it on.

What we'd do

  • Find the pages AI already cites for your buyers' questions, and start there.
  • Add one true, specific fact people can see, worded so it is easy to repeat.
  • Publish a Markdown version of the page with the same words, and link it.
  • Check each engine on its own, every day, and count every answer that repeats the fact.
  • Plan a next step for each engine that has not picked it up yet.

Where this could be wrong

The numbers come from a social media thread retelling the test, not from Ramp, and we have not seen the raw data. One company, one offer and one month is a small sample. Engines change how they read pages often, so a format that works today may matter less later.

外部の数値の出どころ

  1. Thread on Ramp's AI-only offer test · CrowdReply on X
  2. Spam policies for Google web search: Cloaking · Google Search Central
すべてのノート

質問

Is showing content only to AI bots allowed?

Showing crawlers content that people don't see is cloaking, which Google's spam policies prohibit. Put the same words in front of people and AI crawlers, and measure whether each engine repeats them.

Does a Markdown version of a page help AI assistants?

In the test described, a Markdown version showed up more reliably than plain HTML or schema heavy pages. Keep its words the same as the page people read.

Why do ChatGPT, Claude and Perplexity treat the same page differently?

Each engine fetches, ranks and summarizes pages its own way. In the test described, Claude passed the offer on in full, Perplexity loosely and ChatGPT not at all, so track each engine on its own.

関連するメモ

現場ノート4分で読めます

競合の名前を入れても、AIが別の会社を説明するのは止まらなかった。カテゴリを入れると、96件中96件で止まった

AIが自社ブランドを別の会社と取り違えないようにするには、どうすればよいのでしょうか。同じブランドについて3通りの尋ね方を試しました。質問に競合の名前を入れても効果はなく、別の会社を説明した回答は5.1%で、名前だけの場合は4.0%でした。効果があったのは、質問にカテゴリを入れることでした。

現場ノート4分で読めます

実際の検索でブランド名が出たAIの回答は5.2%。自社サイトから作った質問では7.7%

ChatGPT での自社ブランドを測るには、どの質問を追跡すべきでしょうか?追跡中の質問1,235件を、実際の Google 検索と、ブランド自身のウェブサイトから作った質問に分けました。作った質問のほうがブランド名が挙がる割合はおよそ1.5倍で、61ブランド中46ブランドでそうなりました。

現場ノート4分で読めます

AIエンジンは、18%のブランドで自社サイトの壊れたページを引用していた

ChatGPT は自社サイトの壊れたページに人を送っているのでしょうか?私たちは、4つのAIエンジンが5週間に引用したブランド自社サイトのページをすべて取り出し、1件ずつ読み込みました。ほとんどは正常でした。正常でなかったページの多くは、エンジンがでっち上げたリンクではなく、ブランドが移動または削除したページでした。

あなた自身の数値が当社の最高の記事を上回ります。

あなたのカテゴリーの候補リストは、このページを誰が読もうと読むまいと、今まさに書かれています。4つのエンジンで25件の質問を今日尋ねる無料トライアルで、そこにあなたの名前があるかがわかります。

100件の回答。25 Prompts/質問。主要AIエンジン。無料。ChatGPTPerplexityClaudeGoogle AI OverviewsGoogle AI ModeGeminiMicrosoft CopilotGrokDeepSeekMeta AI