- Title:
- AEO vs SEO: is SEO dead?
Today the site panel showed 488 people, 15 AI agents and 1,335 bots. And zero tool calls on the MCP server I left there precisely for agents.
You can ask whether SEO is dead, now that a lot of people ask ChatGPT instead of searching on Google. Before answering, it is worth looking at who actually reads your site. Mine is read by people, by search engines and by agents, and each one wants a different format. The numbers from 15 days show something I did not expect: for an AI agent, HTML is the format it reads least on my site.
SEO is for whoever clicks, AEO is for whoever answers
SEO (search engine optimization) is making your page rank well when someone searches, so that person clicks and reaches you. AEO (answer engine optimization) is making an assistant like ChatGPT, Perplexity or Claude read your content, understand it and use it in the answer it gives, often without the person ever opening your site. You will see the same ground under other names, like GEO (generative engine optimization, a term from a 2023 academic paper) and LLMO. I use AEO throughout so I do not mix them up.
| SEO | AEO | |
|---|---|---|
| Who reads first | the search crawler | the crawler or agent of an assistant |
| What it wants | to know what the page is about and whether it deserves a position | clean content, to cite or answer |
| Who arrives at the end | a person, via the click | sometimes nobody: the answer is enough |
| How you measure | position and clicks from search | reads of machine routes and citations |
My reason for building the second column was a bet. When I rewrote the site in 2026, my work over the previous year was building AI agent tooling, and I wanted the site to be readable by an agent. AI already takes part in screening candidates: according to a June 2026 report by Corp! Magazine, 44% of HR professionals surveyed by SHRM use AI to review resumes, and 42% of recruiters in the Bullhorn survey use it to screen candidates. That is resume screening. An agent visiting the candidate's site before a human reads anything, I did not find measured anywhere, so that reason is still my own hypothesis, and the numbers below are what I have to test it.
What shrank was the click
The drop in clicks has numbers, and they come from more than one place. Pew tracked the real browsing of 900 US adults in March 2025: when the search showed an AI summary, 8% of visits ended in a click on a traditional result, against 15% without the summary. Only 1% clicked a link inside the summary. And 26% ended the session there, against 16% (Pew, July 2025). Ahrefs looked at 300,000 keywords and, in February 2026, estimated that the presence of an AI Overview comes with a 58% lower CTR in the first position, against 34.5% in its April 2025 study (Ahrefs). And SparkToro, with Similarweb data, counted 68.01% of Google searches ending without a click in the US in early 2026, with a warning from the author himself that the 2024 number (60.45%) came from a different panel (SparkToro, June 2026).
Not every study agrees. Semrush tracked the same keywords before and after they got an AI Overview and saw zero-click fall from 33.75% to 31.53%, which means people clicked a little more (Semrush, December 2025). Their explanation is that those searches were already informational and that the AI Overview spread to other types of search. I go with the cautious version. For whoever depends on the informational click at the top, the average effect is negative, and its size varies with the type of search and with the month. All of these studies are from the US. I did not find the equivalent for Brazilian search.
What I can say about Brazil is the volume. Between January and August 2025, Google had 38.57 billion visits in the country and ChatGPT had 1.97 billion, according to Cadastra and Similarweb (Exame, November 2025). ChatGPT grew 71.9% in the period, and search is still about twenty times bigger. Worldwide, Similarweb counted in June 2025 1.13 billion visits coming from AI platforms to the top thousand sites, against 191 billion coming from Google (Similarweb, July 2025). The click shrank. Search is still the bulk of the traffic.
On the publishers' side, Cloudflare measured the bill this leaves. In July 2025 Anthropic's crawler pulled 38,000 pages for every visitor the company sent back, and Perplexity's, 194 (Cloudflare, August 2025). Most of that crawl is for training, which is not meant to send visits. Even so, it is the "sometimes nobody" column of the table that these numbers describe.
SEO is still there, and it is the base for both
Before anything for AI, the site has the usual SEO: sitemap.xml, an absolute canonical URL on every page, a share image generated per page, RSS and an seoTitle when the post title is too long for the search result.
Whoever decides this on Google's side says the same. In the guide to optimizing for generative AI features, published in May 2026 and updated in July, the answer to "does SEO still matter for AI search?" is "yes". The AI features rely on the same ranking and quality systems as search. On AEO and GEO, the text says that, from Google Search's point of view, optimizing for AI search is optimizing the search experience, and so it is still SEO. Danny Sullivan had said something similar in December 2025: SEO for AI is still SEO.
Structured data comes in here, and it is where I changed my mind the most while researching. I use a JSON-LD block (the schema.org format that describes the page in fields, not in prose) in the <head>. The site declares who I am (Person), what the site is (WebSite), what each page is (WebPage) and each portfolio project (SoftwareApplication when it has a live link, CreativeWork when it does not). I used to think this helped the agent avoid having to guess who wrote what. Google says the opposite in the same guide: structured data is not a requirement for AI answers, and there is no special schema.org markup you need to add for that. It still counts as part of SEO, because it enables rich results in search. So my site's JSON-LD is there for the usual SEO, and I do not treat it as a shortcut to show up in an AI answer. That position is Google's, about Google's features. For ChatGPT, Perplexity and Claude, I found no study that measures the effect.
One detail that costs a single line and avoids pain: every JSON-LD goes through one function that escapes <. Without that, a title with </script> closes the tag in the middle of the content, and a test fails anyone who tries to reimplement the escape elsewhere.
Allow AI crawlers by name
The robots.txt with User-agent: * and Allow: / already lets everyone in. Even so, mine lists the AI crawlers one by one:
export const ANSWER_ENGINE_CRAWLERS = [
'GPTBot',
'OAI-SearchBot',
'ChatGPT-User',
'ClaudeBot',
'Claude-User',
'Claude-SearchBot',
'PerplexityBot',
'Perplexity-User',
'Google-Extended',
'Applebot-Extended',
'CCBot',
] as const;
The comment in the code explains why: the list documents the intent, and survives the day I tighten the default rule. If * ever becomes Disallow, the eleven still get in, and nobody has to remember that they existed.
Listing them by name also forces you to know what each one does, because training, search and user requests are different bots. OpenAI's documentation separates GPTBot, which collects content that may become training data, OAI-SearchBot, which decides whether the site shows up in ChatGPT search, and ChatGPT-User, which visits a page when someone asks. About that last one, the documentation itself warns: because the action is started by the user, robots.txt rules may not apply. Anthropic's page has the same split: ClaudeBot collects for training, Claude-User fetches pages at the request of whoever is chatting with Claude, and Claude-SearchBot indexes content to improve search. Blocking only the training one and leaving the rest open is a choice robots.txt allows. Mine allows all of them. That is a personal site decision, not a recommendation for everyone.
Just do not confuse that with security. robots.txt is a polite request, and a rude crawler ignores it. In August 2025 Cloudflare showed that Perplexity, when blocked, swapped its declared user agent for a generic Chrome, and that in the same test ChatGPT-User stopped when it found a Disallow (Cloudflare). Perplexity disputed the attribution. The real protection is having nothing sensitive on public routes, and the site has a test that scans the machine routes for email, keys, absolute paths and private IPs.
Give every page a Markdown version
That was the technique that got the most agent reads on my site, and I have the number.
Every page on the site has a Markdown sibling: /tratando-dotfiles-como-software has /tratando-dotfiles-como-software.md, with the clean text, a header with title, URL and date, and no menu, footer or script. There is also an /ia version, the same thing rendered for a person to see how the agent sees the page. Both come from a single table (path to Markdown), and a page outside it returns 404. The HTML announces the sibling in the <head>:
<link rel="alternate" type="text/markdown" href="/tratando-dotfiles-como-software.md" />
The site analytics classifies each visit by user agent into people, AI agents, search engines and other bots, and measures on the server, with no browser script. From October 14 to 28, content reads by AI agents looked like this:
| Format | Agent reads |
|---|---|
.md |
429 |
/ia |
376 |
| HTML | 245 |
Three in four reads were on the two versions built for machines, and the most read of all is plain Markdown. On my site, when clean text exists, the agent reads the clean text, and it finds the version on its own through the link in the <head>.
That number has limits, and the first is that it measures reading, not citation. I do not know how many of those 1,050 reads became an answer for someone. Dries Buytaert, who ran the same experiment on his blog, announced the .md with the same rel="alternate" and looked at a month of Cloudflare logs. The crawlers found the Markdown through the link, and none of them used the Accept: text/markdown header. The share of requests for .md varied a lot: 34.8% for GPTBot, 22.7% for OAI-SearchBot, 2.1% for ClaudeBot and 0.1% for ChatGPT-User. His crawler traffic went up about 7%, because the bots request both the HTML and the Markdown. He could not demonstrate a gain in citations. His conclusion: if it is work, spend the effort on the content.
I cannot say that we are measuring different audiences. The "AI agent" category in my analytics also mixes training crawlers (GPTBot, ClaudeBot, CCBot), search crawlers and user agents, like his logs. Even so, on my site three in four reads were on the machine versions, and on his the HTML got most of the requests. With what I have, I cannot explain that difference.
There are also people against the idea. In February 2026, Google's John Mueller called it a "stupid idea" to convert pages to Markdown for bots. His target was a middleware that detects GPTBot or ClaudeBot and swaps the HTML response for a text file, with no menu and no internal links. Google's 2026 guide also says you do not need AI files or Markdown to show up in Google Search, because it does not use them. And Profound ran a controlled experiment, on 381 pages across 6 sites, serving Markdown to bots on half of them: it measured about one extra visit at the median, with no statistical significance (Profound, February 2026).
None of that describes my site. I serve HTML to the human, announce the sibling with rel="alternate" and do not swap the response by user agent. Profound measured the volume of bot visits, and my panel measures which URL the agent picks when both exist. Those are different questions. What I claim is the second one, on my site, over 15 days. That Markdown increases citation, I do not claim.
The llms.txt almost nobody read
Every AEO guide tells you to create llms.txt, a Markdown index at the site root, with a summary of who you are and annotated links to the pages that matter (Jeremy Howard's proposal is here). Mine has existed since launch, along with an llms-full.txt with everything in a single file, and both are generated from the same data layer as the pages. A hand-written llms.txt goes stale on the first portfolio change.
In the same period, llms.txt got 6 AI agent reads. And 65 from visits classified as people (I suspect curious developers, but the analytics does not keep the user agent, so there is no way to know).
The rest of the internet saw something similar. Ahrefs looked at 137,000 domains in May 2026: 28% publish an llms.txt, and 97% of those files got no requests in the month (Ahrefs, June 2026). Of the requests that did exist, 96% were from bots, and only 19.5% from named AI tools. AI bots also do not go looking for the file where it does not exist. On Dries's blog there were 52 requests in a month, all from SEO audit tools, and he summed up llms.txt as a solution looking for a problem. Google says it ignores the file and that keeping it neither helps nor hurts.
AGENTS.md is the same idea for another audience: a hiring brief, with services, availability and how to call the endpoints. It got 10 agent reads and 116 from people. Both stay up, because they cost nothing to maintain. But if I had to pick only one thing to do today, I would do the Markdown version of every page before llms.txt.
The MCP nobody calls
The site also has an MCP server (the Model Context Protocol, a standard way for an assistant to call functions on a service) at /api/mcp, with five tools: list projects, detail a project, profile, resume and availability. And a /api/hire, so an agent can send a proposal on someone's behalf. I covered how I built that in the MCP post.
On the 17th, the panel showed agent access under "Live" and zero under "Tool calls". I thought it was a counting bug. It was not. The raw event showed only the connection sequence: server/discover, initialize, notifications/initialized, tools/list. That is the handshake Claude Code does when it opens a session with MCP configured. The tool is only called if the model decides to use one, and nobody had decided. To prove it, I called get_profile in production and the counter went from nothing to 1.
Over the 15 days, there were 149 connections and 2 tool calls. One of them was that test of mine. And to make it worse (for me), some of the connections are mine: the cookie that marks my visits does not ride along on an MCP call, and the 19 connections from the claude-code client are my own sessions.
The protocol is growing, just somewhere else. On its first anniversary, on 25/11/2025, the official registry had close to 2,000 servers, 407% more than in September (MCP blog). That speaks about servers and clients, and my panel speaks about a portfolio that an agent visits in passing. I did not find any survey of the tool call rate on public sites, so the 149 and the 2 are a well-measured anecdote, not a statistic.
MCP works. It is just that today an agent that reaches the site reads the Markdown and leaves, instead of connecting and asking.
I keep doing both
The numbers on my site pull both ways. In 15 days, search brought few people: 17 visits from Google, 2 from Bing and 2 from DuckDuckGo, against 168 from LinkedIn. That is a small personal site, and it says nothing about the industry. And agents read the site more than a thousand times, without me knowing how many of those reads became an answer for someone.
But everything I did for agents sits on top of what SEO already asked for: the sitemap.xml with every page, the canonical URL that says which version counts and the JSON-LD that says who wrote it, which serves SEO and which I do not sell as an AI shortcut. Dries reached the same conclusion by his own path: what keeps working is clear writing, content with authority and publishing on time. Google's guide puts unique, useful content at the top of the list, and says it weighs more, in the long run, than any other suggestion in the text.
My answer to the title is that SEO is not dead. Whoever publishes now has two audiences, the one that clicks and the one that answers in place of whoever would click, and the part that shrank was the informational click. I keep doing both, in the same code. Every new page is born with the HTML for the human, the Markdown version for the agent and the structured data for search.
User-agent: *
---
title: AEO vs SEO: is SEO dead?
url: https://nandomoreira.dev/en/aeo-vs-seo-is-seo-dead
date: 2026-10-28
---
Today the site panel showed 488 people, 15 AI agents and 1,335 bots. And zero tool calls on the MCP server I left there precisely for agents.
You can ask whether SEO is dead, now that a lot of people ask ChatGPT instead of searching on Google. Before answering, it is worth looking at who actually reads your site. Mine is read by people, by search engines and by agents, and each one wants a different format. The numbers from 15 days show something I did not expect: **for an AI agent, HTML is the format it reads least on my site.**
## SEO is for whoever clicks, AEO is for whoever answers
<abbr title="Search Engine Optimization">SEO</abbr> (search engine optimization) is making your page rank well when someone searches, so that person clicks and reaches you. <abbr title="Answer Engine Optimization">AEO</abbr> (answer engine optimization) is making an assistant like ChatGPT, Perplexity or Claude read your content, understand it and use it in the answer it gives, often without the person ever opening your site. You will see the same ground under other names, like GEO (<i lang="en">generative engine optimization</i>, a term from a [2023 academic paper](https://arxiv.org/abs/2311.09735)) and LLMO. I use AEO throughout so I do not mix them up.
| | SEO | AEO |
|---|---|---|
| Who reads first | the search crawler | the crawler or agent of an assistant |
| What it wants | to know what the page is about and whether it deserves a position | clean content, to cite or answer |
| Who arrives at the end | a person, via the click | sometimes nobody: the answer is enough |
| How you measure | position and clicks from search | reads of machine routes and citations |
My reason for building the second column was a bet. When I rewrote the site in 2026, my work over the previous year was building AI agent tooling, and I wanted the site to be readable by an agent. AI already takes part in screening candidates: according to a June 2026 report by Corp! Magazine, [44% of HR professionals surveyed by SHRM use AI to review resumes, and 42% of recruiters in the Bullhorn survey use it to screen candidates](https://www.corpmagazine.com/industry/human-resources/strategic-ai-use-with-guardrails-necessary-for-recruiter-success/). That is resume screening. An agent visiting the candidate's site before a human reads anything, I did not find measured anywhere, so that reason is still my own hypothesis, and the numbers below are what I have to test it.
## What shrank was the click
The drop in clicks has numbers, and they come from more than one place. Pew tracked the real browsing of 900 US adults in March 2025: when the search showed an AI summary, 8% of visits ended in a click on a traditional result, against 15% without the summary. Only 1% clicked a link inside the summary. And 26% ended the session there, against 16% ([Pew, July 2025](https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/)). Ahrefs looked at 300,000 keywords and, in February 2026, estimated that the presence of an AI Overview comes with a 58% lower CTR in the first position, against 34.5% in its April 2025 study ([Ahrefs](https://ahrefs.com/blog/ai-overviews-reduce-clicks-update/)). And SparkToro, with Similarweb data, counted 68.01% of Google searches ending without a click in the US in early 2026, with a warning from the author himself that the 2024 number (60.45%) came from a different panel ([SparkToro, June 2026](https://sparktoro.com/blog/in-2026-less-than-one-third-of-google-searches-still-send-a-click/?hl=en-US)).
Not every study agrees. Semrush tracked the same keywords before and after they got an AI Overview and saw zero-click fall from 33.75% to 31.53%, which means people clicked a little more ([Semrush, December 2025](https://www.semrush.com/blog/semrush-ai-overviews-study/)). Their explanation is that those searches were already informational and that the AI Overview spread to other types of search. I go with the cautious version. For whoever depends on the informational click at the top, the average effect is negative, and its size varies with the type of search and with the month. All of these studies are from the US. I did not find the equivalent for Brazilian search.
What I can say about Brazil is the volume. Between January and August 2025, Google had 38.57 billion visits in the country and ChatGPT had 1.97 billion, according to Cadastra and Similarweb ([Exame, November 2025](https://exame.com/marketing/ia-generativa-leva-6-mi-ao-e-commerce-e-tem-mercado-livre-na-lideranca/)). ChatGPT grew 71.9% in the period, and search is still about twenty times bigger. Worldwide, Similarweb counted in June 2025 1.13 billion visits coming from AI platforms to the top thousand sites, against 191 billion coming from Google ([Similarweb, July 2025](https://aisearch.similarweb.com/blog/ai-referral-traffic-winners/)). The click shrank. Search is still the bulk of the traffic.
On the publishers' side, Cloudflare measured the bill this leaves. In July 2025 Anthropic's crawler pulled 38,000 pages for every visitor the company sent back, and Perplexity's, 194 ([Cloudflare, August 2025](https://blog.cloudflare.com/crawlers-click-ai-bots-training/)). Most of that crawl is for training, which is not meant to send visits. Even so, it is the "sometimes nobody" column of the table that these numbers describe.
## SEO is still there, and it is the base for both
Before anything for AI, the site has the usual SEO: `sitemap.xml`, an absolute canonical URL on every page, a share image generated per page, RSS and an `seoTitle` when the post title is too long for the search result.
Whoever decides this on Google's side says the same. In the [guide to optimizing for generative AI features](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide), published in May 2026 and updated in July, the answer to "does SEO still matter for AI search?" is "yes". The AI features rely on the same ranking and quality systems as search. On AEO and GEO, the text says that, from Google Search's point of view, optimizing for AI search is optimizing the search experience, and so it is still SEO. Danny Sullivan had said something similar in December 2025: [SEO for AI is still SEO](https://searchengineland.com/google-danny-sullivan-seo-for-ai-is-still-seo-466368).
Structured data comes in here, and it is where I changed my mind the most while researching. I use a JSON-LD block ([the schema.org format](https://schema.org) that describes the page in fields, not in prose) in the `<head>`. The site declares who I am (`Person`), what the site is (`WebSite`), what each page is (`WebPage`) and each portfolio project (`SoftwareApplication` when it has a live link, `CreativeWork` when it does not). I used to think this helped the agent avoid having to guess who wrote what. Google says the opposite in the same guide: structured data is not a requirement for AI answers, and there is no special schema.org markup you need to add for that. It still counts as part of SEO, because it enables rich results in search. So my site's JSON-LD is there for the usual SEO, and I do not treat it as a shortcut to show up in an AI answer. That position is Google's, about Google's features. For ChatGPT, Perplexity and Claude, I found no study that measures the effect.
One detail that costs a single line and avoids pain: every JSON-LD goes through one function that escapes `<`. Without that, a title with `</script>` closes the tag in the middle of the content, and a test fails anyone who tries to reimplement the escape elsewhere.
## Allow AI crawlers by name
The `robots.txt` with `User-agent: *` and `Allow: /` already lets everyone in. Even so, mine lists the AI crawlers one by one:
```ts
export const ANSWER_ENGINE_CRAWLERS = [
'GPTBot',
'OAI-SearchBot',
'ChatGPT-User',
'ClaudeBot',
'Claude-User',
'Claude-SearchBot',
'PerplexityBot',
'Perplexity-User',
'Google-Extended',
'Applebot-Extended',
'CCBot',
] as const;
```
The comment in the code explains why: the list documents the intent, and survives the day I tighten the default rule. If `*` ever becomes `Disallow`, the eleven still get in, and nobody has to remember that they existed.
Listing them by name also forces you to know what each one does, because training, search and user requests are different bots. [OpenAI's documentation](https://developers.openai.com/api/docs/bots) separates `GPTBot`, which collects content that may become training data, `OAI-SearchBot`, which decides whether the site shows up in ChatGPT search, and `ChatGPT-User`, which visits a page when someone asks. About that last one, the documentation itself warns: because the action is started by the user, `robots.txt` rules may not apply. [Anthropic's page](https://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler) has the same split: `ClaudeBot` collects for training, `Claude-User` fetches pages at the request of whoever is chatting with Claude, and `Claude-SearchBot` indexes content to improve search. Blocking only the training one and leaving the rest open is a choice `robots.txt` allows. Mine allows all of them. That is a personal site decision, not a recommendation for everyone.
Just do not confuse that with security. `robots.txt` is a polite request, and a rude crawler ignores it. In August 2025 Cloudflare showed that Perplexity, when blocked, swapped its declared user agent for a generic Chrome, and that in the same test `ChatGPT-User` stopped when it found a `Disallow` ([Cloudflare](https://blog.cloudflare.com/perplexity-is-using-stealth-undeclared-crawlers-to-evade-website-no-crawl-directives/)). Perplexity disputed the attribution. The real protection is having nothing sensitive on public routes, and the site has a test that scans the machine routes for email, keys, absolute paths and private IPs.
## Give every page a Markdown version
That was the technique that got the most agent reads on my site, and I have the number.
Every page on the site has a Markdown sibling: `/tratando-dotfiles-como-software` has `/tratando-dotfiles-como-software.md`, with the clean text, a header with title, URL and date, and no menu, footer or script. There is also an `/ia` version, the same thing rendered for a person to see how the agent sees the page. Both come from a single table (path to Markdown), and a page outside it returns 404. The HTML announces the sibling in the `<head>`:
```html
<link rel="alternate" type="text/markdown" href="/tratando-dotfiles-como-software.md" />
```
The site analytics classifies each visit by user agent into people, AI agents, search engines and other bots, and measures on the server, with no browser script. From October 14 to 28, content reads by AI agents looked like this:
| Format | Agent reads |
|---|---|
| `.md` | 429 |
| `/ia` | 376 |
| HTML | 245 |
Three in four reads were on the two versions built for machines, and the most read of all is plain Markdown. On my site, when clean text exists, the agent reads the clean text, and it finds the version on its own through the link in the `<head>`.
That number has limits, and the first is that it measures reading, not citation. I do not know how many of those 1,050 reads became an answer for someone. Dries Buytaert, who ran the same experiment on his blog, [announced the `.md` with the same `rel="alternate"`](https://dri.es/giving-agents-a-markdown-version-of-my-blog) and [looked at a month of Cloudflare logs](https://dri.es/markdown-llms-txt-and-ai-crawlers). The crawlers found the Markdown through the link, and none of them used the `Accept: text/markdown` header. The share of requests for `.md` varied a lot: 34.8% for `GPTBot`, 22.7% for `OAI-SearchBot`, 2.1% for `ClaudeBot` and 0.1% for `ChatGPT-User`. His crawler traffic went up about 7%, because the bots request both the HTML and the Markdown. He could not demonstrate a gain in citations. His conclusion: if it is work, spend the effort on the content.
I cannot say that we are measuring different audiences. The "AI agent" category in my analytics also mixes training crawlers (`GPTBot`, `ClaudeBot`, `CCBot`), search crawlers and user agents, like his logs. Even so, on my site three in four reads were on the machine versions, and on his the HTML got most of the requests. With what I have, I cannot explain that difference.
There are also people against the idea. In February 2026, Google's John Mueller [called it a "stupid idea"](https://www.searchenginejournal.com/googles-mueller-calls-markdown-for-bots-idea-a-stupid-idea/566598/) to convert pages to Markdown for bots. His target was a middleware that detects `GPTBot` or `ClaudeBot` and swaps the HTML response for a text file, with no menu and no internal links. Google's 2026 guide also says you do not need AI files or Markdown to show up in Google Search, because it does not use them. And Profound ran a controlled experiment, on 381 pages across 6 sites, serving Markdown to bots on half of them: it measured about one extra visit at the median, with no statistical significance ([Profound, February 2026](https://www.tryprofound.com/blog/does-markdown-increase-ai-bot-traffic)).
None of that describes my site. I serve HTML to the human, announce the sibling with `rel="alternate"` and do not swap the response by user agent. Profound measured the volume of bot visits, and my panel measures which URL the agent picks when both exist. Those are different questions. What I claim is the second one, on my site, over 15 days. That Markdown increases citation, I do not claim.
## The llms.txt almost nobody read
Every AEO guide tells you to create `llms.txt`, a Markdown index at the site root, with a summary of who you are and annotated links to the pages that matter ([Jeremy Howard's proposal is here](https://llmstxt.org)). Mine has existed since launch, along with an `llms-full.txt` with everything in a single file, and both are generated from the same data layer as the pages. A hand-written `llms.txt` goes stale on the first portfolio change.
In the same period, `llms.txt` got 6 AI agent reads. And 65 from visits classified as people (I suspect curious developers, but the analytics does not keep the user agent, so there is no way to know).
The rest of the internet saw something similar. Ahrefs looked at 137,000 domains in May 2026: 28% publish an `llms.txt`, and 97% of those files got no requests in the month ([Ahrefs, June 2026](https://ahrefs.com/blog/llmstxt-study/)). Of the requests that did exist, 96% were from bots, and only 19.5% from named AI tools. AI bots also do not go looking for the file where it does not exist. On Dries's blog there were 52 requests in a month, all from SEO audit tools, and he summed up `llms.txt` as a solution looking for a problem. Google says it ignores the file and that keeping it neither helps nor hurts.
`AGENTS.md` is the same idea for another audience: a hiring brief, with services, availability and how to call the endpoints. It got 10 agent reads and 116 from people. Both stay up, because they cost nothing to maintain. But if I had to pick only one thing to do today, I would do the Markdown version of every page before `llms.txt`.
## The MCP nobody calls
The site also has an MCP server (the [Model Context Protocol](https://modelcontextprotocol.io), a standard way for an assistant to call functions on a service) at `/api/mcp`, with five tools: list projects, detail a project, profile, resume and availability. And a `/api/hire`, so an agent can send a proposal on someone's behalf. I covered how I built that [in the MCP post](/en/site-that-talks-to-ai-via-mcp).
On the 17th, the panel showed agent access under "Live" and zero under "Tool calls". I thought it was a counting bug. It was not. The raw event showed only the connection sequence: `server/discover`, `initialize`, `notifications/initialized`, `tools/list`. That is the handshake Claude Code does when it opens a session with MCP configured. The tool is only called if the model decides to use one, and nobody had decided. To prove it, I called `get_profile` in production and the counter went from nothing to 1.
Over the 15 days, there were 149 connections and 2 tool calls. One of them was that test of mine. And to make it worse (for me), some of the connections are mine: the cookie that marks my visits does not ride along on an MCP call, and the 19 connections from the `claude-code` client are my own sessions.
The protocol is growing, just somewhere else. On its first anniversary, on 25/11/2025, the official registry had close to 2,000 servers, 407% more than in September ([MCP blog](https://blog.modelcontextprotocol.io/posts/2025-11-25-first-mcp-anniversary/)). That speaks about servers and clients, and my panel speaks about a portfolio that an agent visits in passing. I did not find any survey of the tool call rate on public sites, so the 149 and the 2 are a well-measured anecdote, not a statistic.
MCP works. It is just that today an agent that reaches the site reads the Markdown and leaves, instead of connecting and asking.
## I keep doing both
The numbers on my site pull both ways. In 15 days, search brought few people: 17 visits from Google, 2 from Bing and 2 from DuckDuckGo, against 168 from LinkedIn. That is a small personal site, and it says nothing about the industry. And agents read the site more than a thousand times, without me knowing how many of those reads became an answer for someone.
But everything I did for agents sits on top of what SEO already asked for: the `sitemap.xml` with every page, the canonical URL that says which version counts and the JSON-LD that says who wrote it, which serves SEO and which I do not sell as an AI shortcut. Dries reached the same conclusion by his own path: what keeps working is clear writing, content with authority and publishing on time. Google's guide puts unique, useful content at the top of the list, and says it weighs more, in the long run, than any other suggestion in the text.
My answer to the title is that SEO is not dead. Whoever publishes now has two audiences, the one that clicks and the one that answers in place of whoever would click, and the part that shrank was the informational click. I keep doing both, in the same code. Every new page is born with the HTML for the human, the Markdown version for the agent and the structured data for search.
`User-agent: *`