Updated on 19 November 2026, by João Leite, Founder of murmur.
Medium and Substack help you appear in AI as a source, but the citation link leads to the platform where the text is. That is why the article goes first to the company’s domain and, when it goes on to Medium, enters through import. The Medium help center writes that “Imported stories will automatically apply a canonical URL referencing the original source material”, that is, the imported text automatically gets a canonical link to the original. The canonical link is the markup, in the page code, that tells search engines which URL is the original when the same text exists in two places.
- Medium and Substack appear in answers as a source, with the platform’s URL in the link. In murmur’s weekly measurement, on GEO and AI search, pages from the two platforms were a source in 32 of the 2,437 answers with sources, in 21 of the 90 questions, from 4 September to 2 November 2026, above all in Google AIO (14 of 745) and in Perplexity (12 of 787).
- The original stays on the company’s site. On Medium, the copy enters through import, which points the canonical link to the company’s page.
- Medium asks OpenAI’s and Anthropic’s training crawlers to stay out. The medium.com robots.txt, read on 9 November 2026, closes the site to GPTBot, ClaudeBot and four more crawlers, and leaves the others under the general rule.
- Substack leaves the training choice to the author. The substack.com robots.txt has only the general rule, and each publication has an option that asks AI models not to train on the text, with the warning that this can reduce discovery in searches that answer with AI.
- Search crawler and training crawler are separate choices for OpenAI. The site can allow OAI-SearchBot and block GPTBot.
Medium and Substack become a source, and the citation link leads to the platform where the text is
An article on Medium or Substack can become a source for the AI answer, and the citation shows the URL where the text was read, which is the platform’s. Anyone who clicks the source reaches the article inside Medium or Substack, with the platform’s name in the URL, and not the company’s site. The visit, the URL and the name left on the screen belong to the platform.
murmur, a GEO agency in Brazil that places brands in the answers of ChatGPT, Claude, Google AIO and other AI agents, tracks the same 90 questions about GEO, AI search and GEO agencies every week. The questions go to ChatGPT, Claude, Perplexity and Google AIO, the summary written by AI that Google shows at the top of many searches. In 10 rounds, from 4 September to 2 November 2026, they produced 2,437 answers with sources, and Medium and Substack pages were a source in 32 of them, spread across 21 of the 90 questions. By AI model, that was 14 of the 745 answers with sources in Google AIO, 12 of the 787 in Perplexity, 5 of the 442 in Claude and 1 of the 463 in ChatGPT.
The article on the platform, then, can reach the AI answer. Who the cited URL belongs to depends on the order in which the text is published.
Publish first on the company’s domain and import into Medium, with the canonical link to the original
The order that leaves the original with the company is to publish on its domain and then take the text to Medium through import. On the help page about importing, updated on 26 August 2026, Medium says it applies the canonical link by itself on the imported text, pointing to the page of origin. The canonical link serves search: it is the notice to search engines that an original exists and where it is.
The publication moves like this, from the site to the platforms:
- Publish the whole article on the company’s site, at a URL that will not change, with the answer at the beginning and the company’s name in the text.
- Take the text to Medium through import, which marks the copy with the canonical link to the company’s page.
- Send the Substack edition with the link to the article on the site, so that newsletter readers also reach the original.
- Decide the Substack training option after reading its warning, which talks about the publication’s discovery in searches that answer with AI.
The definitive URL in the first step has a reason. The imported copy points to it, and an original that changes URL leaves Medium’s canonical link pointing to a page that no longer exists.
What Medium lets AI crawlers read: training closed and the general rule for the others
Medium closes the site to six crawlers, among them OpenAI’s and Anthropic’s training crawlers, and leaves the others under the file’s general rule. The robots.txt is the site file that tells Google’s crawlers and AI crawlers which pages they can read. AI crawlers (crawlers) are the programs that AI companies use to read sites, such as GPTBot, from OpenAI.
Read on 9 November 2026, Medium’s robots.txt gathers Bytespider, ClaudeBot, FacebookBot, GoogleOther, GPTBot and meta-externalagent in one group, with the rule that closes the entire site, and then opens a few institutional pages to that group. GPTBot collects text that can go into the training of OpenAI’s models, and ClaudeBot is the training bot of Anthropic, the company that makes Claude. Bytespider is ByteDance’s bot.
The bots the file does not name follow the general rule, which closes only some folders of the site. OAI-SearchBot, which fetches pages for ChatGPT’s answers, and Claude-SearchBot, which does the same for Claude, are not in the group of six. The robots.txt is a request from the site to the bots, and Medium can change it.
What Substack lets bots read: the general rule and the option to ask AI models not to train
Substack does not single out any bot in the robots.txt of the main address, and the decision about training sits in an option for each publication. Read on 9 November 2026, Substack’s robots.txt has a single rule, for all bots, which closes three folders: /action/, /publish and /sign-in.
The option in each publication makes the request that the file does not make. According to the Substack help center, updated on 8 July 2026, it tells AI tools such as ChatGPT, Claude and Gemini that their models should not be trained on the published content. The same Substack page warns, in the explanation of the option, that “blocking training may limit your publication’s discoverability in tools and search engines that return AI-generated results”. In plain terms: blocking training can reduce the times the publication is found in tools and search engines that answer with AI.
The newsletter author decides with that warning in front of them. Anyone who wants the publication as a source for AI answers has, in Substack’s own text, a reason to think before turning the option on.
Search bot and training bot are separate choices for OpenAI
For OpenAI, allowing the search bot and blocking the training bot are two different decisions. OpenAI’s bots documentation writes that the site owner can allow OAI-SearchBot to appear in search results and block GPTBot to indicate that the collected content should not go into the training of its models. How each bot appears in the robots.txt, and what the llms.txt file does or does not change, is in the guide about llms.txt and AI bots.
Medium and Substack or the company’s site: who decides the bots and which URL goes in the link
Medium, Substack and the company’s site differ in two things that count for AI: who writes the robots.txt and which URL appears when the text becomes a source.
| what is compared | Medium | Substack | company’s site |
|---|---|---|---|
| who writes the robots.txt | Medium | Substack | the company |
| GPTBot and ClaudeBot, training bots (file read on 9 November 2026) | blocked, in a group with four more bots | under the general rule, which closes /action/, /publish and /sign-in | the company decides |
| bots the file does not name, such as OAI-SearchBot | under the general rule, which closes only some folders | under the same general rule | the company decides |
| request to AI models not to train | made by Medium, in the robots.txt | an option in each publication, with a warning about discovery in AI searches | made by the company, in the robots.txt |
| URL the citation link shows | Medium’s | the publication’s on Substack | the company’s domain |
| who it serves best | the author who already has readers on Medium | the newsletter that reaches subscribers | the original, with the company’s name in the URL |
The company website is the only one of the three where the company decides what the bots read and where the citation link carries its name. Medium and Substack come afterward, as a copy, to reach people who already read on those platforms.
When Medium and Substack are worth it: the author with readers on the platform and the newsletter
Medium and Substack are worth it when there are already people reading there: the author who has followers on Medium and the company that sends a newsletter through Substack. In those cases, the platform takes the text to readers who would not reach the website, and the copy points to the original. In murmur’s measurement, pages on the two platforms were a source mainly in Google AIO and Perplexity, across the 90 questions about GEO and AI search.
For the company that has no readers on those platforms, the effort goes into the page on the website, which keeps the company’s name in the URL and answers the buyer’s question. Publishing on a brand subdomain, such as guia.suaempresa.com.br, is another way to keep the original on the company’s own domain, and the guide about content on a brand subdomain covers that choice.
Other channels outside the website weigh differently in each AI. Press has its own guide, about the weight of press relations in each AI, and so does LinkedIn, about LinkedIn as a source for AI models. Which domains ChatGPT uses most in answers in Portuguese is in the guide about the sites that most influence ChatGPT in Brazil.
murmur publishes on the brand’s domain, and the text stays with the client
murmur publishes each client’s content on the client’s own brand domain: the content goes out on a subdomain of the client’s brand, not on murmur’s domain. When an AI uses one of those pages as a source, the citation URL carries the brand name, and anyone who clicks lands on the brand’s website. The content belongs entirely to the client, even after the contract ends, and that is why the cited URL stays with the client.
In the FotoRegistro case study, an e-commerce business for photo printing and personalization, ChatGPT cited the brand’s guide for the first time on 24 July 2026, 30 days after the guide launched. The guide was on a FotoRegistro subdomain, and it was that page, at the brand’s URL, that made it into the answer.
murmur’s work continues after publication. A specialist dedicated to the brand studies the company and the sector, looks at what ChatGPT, Claude, Perplexity and Google AIO answer to the buyer and who appears in its place, and tracks the same questions week by week.
Frequently asked questions
Does republishing the same text on Medium hurt the original page?
Republishing through Medium’s import keeps the website page marked as the original: Medium automatically applies a canonical link to the source URL on the imported text. The canonical link is the notice to search engines of which URL is the original. When an AI uses the text as a source, the citation link leads to the URL where it read the text, which may be the Medium URL. That is why the page on the company website comes first, and Medium receives the imported copy.
Is it worth turning on the Substack option that asks AI models not to train?
It is worth turning on only after weighing Substack’s own notice, which says that blocking training can limit the publication’s discovery in tools and search engines that answer with AI. The option tells tools such as ChatGPT, Claude and Gemini that their models should not be trained on the published text. For those who want the newsletter as a source for AI answers, the notice weighs against turning it on. For those who do not want the text in model training, the option makes that request on behalf of the publication.
Can I publish on Medium before the website?
You can, but the text originates at the Medium URL, and that is the URL that goes in the link when the AI uses the article as a source. The order that leaves the original with the company is the reverse: the page on the website first and the import on Medium afterward, with the canonical link to the company page. That way the company name stays in the URL of the original.
Medium or LinkedIn: where should the company article be published?
The company article goes first on its website, and Medium and LinkedIn come afterward, to reach those who read on each platform. On Medium, the import marks the copy with the canonical link to the original. What LinkedIn lets crawlers read and how to publish there are in the guide about LinkedIn as a source for AI models.
Where to start
Before choosing between Medium, Substack and the website, find out which pages AI models already use as a source when someone asks about what your company sells. The free diagnostic from murmur takes your website URL and shows, without sign-up and in less than two minutes, how AI models answer the questions in your market; the full report arrives by email. murmur’s sector studies show which brands AI models recommended in the sectors it measured.
Read also
- llms.txt: What It Is, What It Is For, and Whether It Works to Get Cited by AI
- Press coverage helps you show up in ChatGPT: the weight of press in each AI model
- LinkedIn as a source for AI models
- Content on a brand subdomain and AI answers
References
- Medium, help center, “Import a post”, updated on 26 August 2026. The imported text automatically carries a canonical link to the original. https://help.medium.com/hc/en-us/articles/214550207-Import-a-post (accessed on 9 November 2026).
- Medium, robots.txt of medium.com. The group with Bytespider, ClaudeBot, FacebookBot, GoogleOther, GPTBot and meta-externalagent has the rule that blocks the site, and some institutional pages are allowed afterward. https://medium.com/robots.txt (accessed on 9 November 2026).
- Substack, robots.txt of substack.com. A single rule, for all crawlers, that blocks /action/, /publish and /sign-in. https://substack.com/robots.txt (accessed on 9 November 2026).
- Substack, help center, page for the option that asks AI tools not to train on the publication, updated on 8 July 2026. The request to the models and the notice about discovery in tools and search engines with AI answers. https://support.substack.com/hc/en-us/articles/20382615953556 (accessed on 9 November 2026).
- OpenAI, “Overview of OpenAI Crawlers”, documentation for OpenAI’s crawlers. The site can allow OAI-SearchBot and block GPTBot for training. https://developers.openai.com/api/docs/bots (accessed on 9 November 2026).
- murmur, FotoRegistro case study. https://murmur.marketing/en/cases/fotoregistro/ (accessed on 6 November 2026).