> It is a bad design trade-off to send a bunch of header fields on every request just to tell the server all of the possible variations of preference held by the user, particularly when there is a very small chance that any of those dimensions are applicable to the target resource. It has been a bad design trade-off ever since the very brief period in 1993-94 when folks didn't know which image format would be usable on all UAs and there was no CSS or javascript to allow for client-side adaptation.
> ...The caching impact of proactive negotiation is far worse than the one extra round trip per site for reactive negotiation, and even that round-trip isn't necessary in formats that support client-side adaptation.
On the caching impact, Simon Willison wrote:
> ...you can’t deploy an application that uses content negotiation via the Accept header behind the Cloudflare CDN — for example serving JSON or HTML for the same URL depending on the incoming Accept header. If you do, Cloudflare may serve cached JSON to an HTML client or vice-versa.
Note: I posted this in another comment with links to those two quotes which I couldn't copy easily now - will add later.
If you do content negotiation, then it’s imperative to send “Vary: accept” in your response. CF and all other CDNs will automatically do the right thing when they see that header.
Content negotiation still has its uses, but most of the time you’re better off using different endpoints.
Many CDNs do not support arbitrary Vary headers. Cloudfront, for example asks you to create a "Cache and Origin Request Policies" that gives you the ability to choose what headers are part of the request that gets sent to your origin. These will get added to the cache key, but it is a static list based on the request to the origin and not the response.
Akamai is another case where Vary is harmful, from their docs [1]:
> As the content in response may be different for the same URL,
> Akamai edge servers don't cache responses that include the Vary header,
> even if the content is cacheable by definition. The only exception is
> the case where the Vary header's value is Accept-Encoding and the
> Content-Encoding header's value is br or gzip – edge servers cache
> such responses, applying the caching rules you defined in your property.
Cloudflare's docs do seem to indicate they support the Vary header as does Fastly. But one should read the docs of their CDN to find out the behavior. Do not assume Vary is supported.
Conceptually, I’m not sure I agree. There’s elegance in clients saying "I want this resource, and I’d like to get your markdown version of it. If you don’t have one, I’ll also take HTML." And if you couple that with optional "file extensions" at the end of the url to force a specific format (say, /foo for automatic negotiation, and /foo.html, /foo.json, or /foo.md for the respective media type,) you have a very easy to use API that adapts to the client; not the other way around.
I take the point that it makes caching harder, but I don’t think that should overrule ergonomics concerns.
In general I think I just don't like the idea of one URL being able to return different content. Forces me to think about what each system I give that URL to may be sending in content negotiation headers. Would rather the HTML is returned and alternatives listed in HTML head.
But for HTML and Markdown in particular, there's been so much useful work done in the semantic HTML space and microformats, that I don't know why anyone interested in this wouldn't just improve their HTML markup and leave it to the agent to do the rest. Convert to markdown or extract the useful HTML before handing it to model.
But it's something I think developers should think about if they rely on caching. If it took Cloudflare this long to support this, there may be other systems which still don't.
> It is a bad design trade-off to send a bunch of header fields on every request just to tell the server all of the possible variations of preference held by the user, particularly when there is a very small chance that any of those dimensions are applicable to the target resource. It has been a bad design trade-off ever since the very brief period in 1993-94 when folks didn't know which image format would be usable on all UAs and there was no CSS or javascript to allow for client-side adaptation.
Doing this with the Accept header is a bad idea, although I think CSS and JavaScripts (in web pages) is not a good solution to this either (they can often make it worse).
My way is the Scorpion conversion file, which must be downloaded explicitly by the end user and the end user must be allowed to override it with their own, and which tells it what to do when it receives a file that it does not recognize, based on the URL or the file type, such as: rewrite the URL, use a uxn program to convert it (to a format that you can use), use a uxn program to display it, etc. Something similar might be possible to add into WWW, by adding a "Interpreter:" response header into HTTP, perhaps using WebAssembly instead of uxn.
I'll do that once any of the top 4 AI chatbots says they'll start making requests with this header. Before that it's just a neat idea with no adoption.
I also think it's exceedingly unlikely that any of the top 4 chatbots would choose to load websites this way. Too many risks for a 0.01% adoption even years later.
Claude Code’s been doing this for at least a year already and I assume Claude Cowork does it too as a result. That’s already millions of users, so it’s not “just a neat idea with no adoption”.
I mean, it's exceptionally easy for them to add it. Even with low adoption. Somewhere in the bowels of chatgpt there is code that queries for html, and then parses that to markdown or something similar. All this would be is adding the logic to request markdown if available, then bypass the html->markdown step if you get markdown back.
I agree. I think a lot of people here are assuming that the full HTML retrieved has to go into the LLM eating up tokens. But why wouldn't the agent try to clean up first and remove bloat and convert to markdown itself, before feeding into LLM. Semantic HTML would make that easier.
Not my problem. Currently my problem is the high traffic of bots that kills our web apps and don't respect robots.txt or meta tags . We are currently deploying Anubis.
Is that even true? I most often use HTML. HTML is about 5%-20% more tokens than a similar Markdown. As a rule of thumb, the number of tags/structural tokens doubles, when going from markdown to html, while the rest don't change much. On the other hand, I can view HTML without any extra/unusual tools. And composing HTML when I need a bit of structure is far easier than composing markdown.
> HTML is about 5%-20% more tokens than a similar Markdown.
I'm not seeing this at all. I've got a small search engine I made that strips HTML back to Markdown for its full-text indexing. HTML is typically 10x bigger than the Markdown of the actual content, but that's because the majority of HTML out there is truly terrible.
I personally like HTML, and my own webpages are all hand-coded HTML. In that case, it's probably a closer ratio to what you describe. I'd suggest it's much higher than 20% more, but it's not likely more than double. But that's assuming someone paying attention to the efficiency of the HTML, and most people / websites just don't.
Markdown is even more readable without tools than HTML - it's essentially a plain text document - but I agree that HTML is better for actual semantic structure.
> On the other hand, I can view HTML without any extra/unusual tools. And composing HTML when I need a bit of structure is far easier than composing markdown.
This is kind of the opposite of reality no? Markdown is just plain text and meant to be human readable. You don't need XML tags to read and write it, opposed to html where you do and you need a browser to properly view it.
HTML I can view in any browser / webview / etc. Good markdown viewers are fewer / more special, or end up translating md to html for display.
And by composing, I didn't mean writing by hand. We are talking about prompting, right? Or that is what I thought we are talking about. Composing HTML "components" into a final prompt HTML is easier than composing markdown snippets into the final prompt. That is because with HTML there are several ergonomic libraries to parse HTML to AST and to format AST back to HTML. The libraries (for parsing to AST and back to strings) are more limited with markdown.
That's highly dependent on what sites you're visiting. Take a look around at a lot of modern sites, there's a sea of divs and spans. Markdown conversion helps LLMs a lot.
Aah, I thought we are talking about prompting or providing information to AI agents in either html or md form, and comparing the two.
Assuming that is what we are talking about, HTML is easier to work with than Markdown, unless you are writing it by hand. That is, composing semantic HTML is more ergonomic than composing a Markdown formatted document from components / snippets, programmatically. The libraries are just better and more versatile in most programming languages. Typically you go from HTML or Markdown to AST, then you compose them to end up with the final tree, then you format the tree to HTML or Markdown. LLMs treat them basically identically (context in HTML or context in Markdown), so I have ended up forming complex prompts / context parts using HTML.
That used to matter to me back in the days when the best models still only accepted ~32,000 tokens, but these days even the models that run on my laptop are happy with ~100,000 and the hosted models I use take ~200,000 or more.
The trick there is to use a subagent to read the HTML page and extract the relevant information, than dumping all that HTML into your top-level session.
That's effectively using an LLM as an HTML to markdown converter, which is both absurdly wasteful and also surprisingly inexpensive (if you use a model like GPT-5.6 Luna.)
I feel like I am going insane. Who in their right mind would feed an LLM raw HTML in the first place??
HTML is a markup language. User agents present it in a way that makes sense for the user—visually, or through assistive technology. Nothing about adding LLM “users” should change this—their user agent, the harness, should present HTML in a way they can natively understand, by translating it to Markdown.
We are not going to rewrite the entire web because harness developers are too lazy or stupid to pull an HTML to markdown package from npm. If some sites want to do so anyway, good for them, in many cases I’d love to skip the CSS/JS and read the markdown directly (or better, nicely formatted). But don’t blame website authors for your harness wasting your tokens.
Exa also has an API for it that has worked well for me, returning markdown for a URL, which means you don't need to render js or anything yourself. It doesn't need an account for up to 1k requests/month, which is more than I've ever needed.
Agree. I think many people forget that not long ago, HTML markup on many sites was a lot richer than it is today. Making it trivial to produce a good trimmed down markdown version.
The reason it may be more difficult today is because we've lost a lot of that. Some of it because of modern JS frameworks, but some also because publishers simply don't want to make it easy for the useful stuff to be scraped and extracted easily.
I'm not convinced that's changing because of AI agents (it's getting worse in many ways with anti-agent rules). Maybe improving for documentation pages intended for agents. But if it is changing, I think it'd be far easier to improve the HTML and let the agent take care of the rest.
> Who in their right mind would feed an LLM raw HTML in the first place??
I'd even say, is that even a problem ? If your website isn't a front end blob soup and just a relatively clean HTML file, your LLM is totally capable to handle it anyway.
And I'm still wondering myself which incentive do I have to serve my content to LLMs anyway.
i have built 2 html to document pipelines and read it back with pdftotext, and I can say that "just convert it yourself" step is the place where things are getting wrong.
For example Markdown has no support for the subscript and superscript in CommonMark or GFM and if you have <sup>4</sup><sub>3</sub> becomes 43 and water formula becomes H2O so in the output you get not the ugly but the wrong one. And no converter can fix it, because that's a target format which can't express it. extractors like Defuddle don't save from this also. and non of those extractors or converter throws you just get wrong text
> Let's say you want to plan a trip to Thailand with your family. You could use the fancy AI to do it for you, or you could build a stupid frontend with minimal natural language understanding.
The web page does not load for me (I cannot reach the server), but the idea to serve Markdown according to the Accept header is possible, although it might be better to do that by changing the file name and adding a header (or a <meta> command) to indicate this possibility, so that there is a different file name in each case. Some already do this, allowing to change the file name for JSON or Markdown instead of HTML, and is helpful (there are also other common ways that you can change a URL to access it differently e.g. "blob" to "raw" in a git viewer, and I use these things sometimes). However, the reason should not be mainly for AI agents, even if it can be helpful for that purpose too.
Unless I'm missing something here, this is all wrong. Markdown is OK, but for a variety of reasons, HTML is much better.
What you _should_ be doing is ensuring that your page works with screen readers. That simultaneously solves any kind of machine-readability issues and also makes your website accessible for the blind.
Everything is determined by cloudflare. If they supported Gemini on the same hostname I’d serve it everywhere. If they supported text/plain and text/html everywhere I’d serve that. But they don’t cache per content encoding so I just do what works for them.
I don’t disagree with you in principle but there is a difference.
The MD content isn’t intended for human consumption whereas HTML is. So you either have that injection readable to your users, which could be jarring to non-technical readers, or you play the cat-and-mouse game of hoping those pesky crawlers don’t ignore text that’s not human-visible on a rendered page.
Right. And today we have something similar with search engines going on where sites want to offer the full text to the engine for seo. But maybe they want to do the opposite to agent and poison the well. So human agents still get the “real” data and agents the “something plausible but not quite right” data.
Or everything just disappears behind something like cloudflare or a paywall.
It's a good thought, but the problem with it is that it can never be trusted by the bot, except for a finite whitelisted set of trusted sites that are known to serve representative markdown that faithfully represents the HTML page, its content, and its links.
> It is a bad design trade-off to send a bunch of header fields on every request just to tell the server all of the possible variations of preference held by the user, particularly when there is a very small chance that any of those dimensions are applicable to the target resource. It has been a bad design trade-off ever since the very brief period in 1993-94 when folks didn't know which image format would be usable on all UAs and there was no CSS or javascript to allow for client-side adaptation.
> ...The caching impact of proactive negotiation is far worse than the one extra round trip per site for reactive negotiation, and even that round-trip isn't necessary in formats that support client-side adaptation.
On the caching impact, Simon Willison wrote:
> ...you can’t deploy an application that uses content negotiation via the Accept header behind the Cloudflare CDN — for example serving JSON or HTML for the same URL depending on the incoming Accept header. If you do, Cloudflare may serve cached JSON to an HTML client or vice-versa.
Note: I posted this in another comment with links to those two quotes which I couldn't copy easily now - will add later.
Content negotiation still has its uses, but most of the time you’re better off using different endpoints.
Akamai is another case where Vary is harmful, from their docs [1]:
Cloudflare's docs do seem to indicate they support the Vary header as does Fastly. But one should read the docs of their CDN to find out the behavior. Do not assume Vary is supported.https://techdocs.akamai.com/property-mgr/docs/rm-vary-header
I take the point that it makes caching harder, but I don’t think that should overrule ergonomics concerns.
But for HTML and Markdown in particular, there's been so much useful work done in the semantic HTML space and microformats, that I don't know why anyone interested in this wouldn't just improve their HTML markup and leave it to the agent to do the rest. Convert to markdown or extract the useful HTML before handing it to model.
It's different content representations. A text in a markdown file is conceptually the same content as the same text in HTML (or PDF).
Wrong: https://developers.cloudflare.com/cache/concepts/vary/
Simon wrote that in 2023: https://simonwillison.net/2023/Nov/20/cloudflare-does-not-co...
But it's something I think developers should think about if they rely on caching. If it took Cloudflare this long to support this, there may be other systems which still don't.
Link to Roy Fielding comment:
https://lists.w3.org/Archives/Public/ietf-http-wg/2013JanMar...
Doing this with the Accept header is a bad idea, although I think CSS and JavaScripts (in web pages) is not a good solution to this either (they can often make it worse).
My way is the Scorpion conversion file, which must be downloaded explicitly by the end user and the end user must be allowed to override it with their own, and which tells it what to do when it receives a file that it does not recognize, based on the URL or the file type, such as: rewrite the URL, use a uxn program to convert it (to a format that you can use), use a uxn program to display it, etc. Something similar might be possible to add into WWW, by adding a "Interpreter:" response header into HTTP, perhaps using WebAssembly instead of uxn.
I also think it's exceedingly unlikely that any of the top 4 chatbots would choose to load websites this way. Too many risks for a 0.01% adoption even years later.
https://acceptmarkdown.com/status
It was already optimized for bots and search engines (which are bots) and it has been used for decades. Why we need to serve in markdown now?
There are also many parts of the HTML, like navs, that are useful for bots and AI and may be removed in the markdown version.
Which React package is this?
There's no "agent". It's a few wrappers around API calls in a trenchcoat.
I'm not seeing this at all. I've got a small search engine I made that strips HTML back to Markdown for its full-text indexing. HTML is typically 10x bigger than the Markdown of the actual content, but that's because the majority of HTML out there is truly terrible.
I personally like HTML, and my own webpages are all hand-coded HTML. In that case, it's probably a closer ratio to what you describe. I'd suggest it's much higher than 20% more, but it's not likely more than double. But that's assuming someone paying attention to the efficiency of the HTML, and most people / websites just don't.
Markdown is even more readable without tools than HTML - it's essentially a plain text document - but I agree that HTML is better for actual semantic structure.
This is kind of the opposite of reality no? Markdown is just plain text and meant to be human readable. You don't need XML tags to read and write it, opposed to html where you do and you need a browser to properly view it.
HTML I can view in any browser / webview / etc. Good markdown viewers are fewer / more special, or end up translating md to html for display.
And by composing, I didn't mean writing by hand. We are talking about prompting, right? Or that is what I thought we are talking about. Composing HTML "components" into a final prompt HTML is easier than composing markdown snippets into the final prompt. That is because with HTML there are several ergonomic libraries to parse HTML to AST and to format AST back to HTML. The libraries (for parsing to AST and back to strings) are more limited with markdown.
Assuming that is what we are talking about, HTML is easier to work with than Markdown, unless you are writing it by hand. That is, composing semantic HTML is more ergonomic than composing a Markdown formatted document from components / snippets, programmatically. The libraries are just better and more versatile in most programming languages. Typically you go from HTML or Markdown to AST, then you compose them to end up with the final tree, then you format the tree to HTML or Markdown. LLMs treat them basically identically (context in HTML or context in Markdown), so I have ended up forming complex prompts / context parts using HTML.
[1] https://arxiv.org/abs/2307.03172
That's effectively using an LLM as an HTML to markdown converter, which is both absurdly wasteful and also surprisingly inexpensive (if you use a model like GPT-5.6 Luna.)
HTML is a markup language. User agents present it in a way that makes sense for the user—visually, or through assistive technology. Nothing about adding LLM “users” should change this—their user agent, the harness, should present HTML in a way they can natively understand, by translating it to Markdown.
We are not going to rewrite the entire web because harness developers are too lazy or stupid to pull an HTML to markdown package from npm. If some sites want to do so anyway, good for them, in many cases I’d love to skip the CSS/JS and read the markdown directly (or better, nicely formatted). But don’t blame website authors for your harness wasting your tokens.
I've been enjoying https://defuddle.md since learning about it. Works great. Not affiliated.
Most harnesses have a "zero config" plugin that works, eg: https://pi.dev/packages/pi-exa
(no affiliation to either)
The reason it may be more difficult today is because we've lost a lot of that. Some of it because of modern JS frameworks, but some also because publishers simply don't want to make it easy for the useful stuff to be scraped and extracted easily.
I'm not convinced that's changing because of AI agents (it's getting worse in many ways with anti-agent rules). Maybe improving for documentation pages intended for agents. But if it is changing, I think it'd be far easier to improve the HTML and let the agent take care of the rest.
I'd even say, is that even a problem ? If your website isn't a front end blob soup and just a relatively clean HTML file, your LLM is totally capable to handle it anyway.
And I'm still wondering myself which incentive do I have to serve my content to LLMs anyway.
> Why do you need AI then?
(from this HN comment, worth a read in full: https://news.ycombinator.com/item?id=43679585)
It's giving https://news.ycombinator.com/item?id=8863#9224
What you _should_ be doing is ensuring that your page works with screen readers. That simultaneously solves any kind of machine-readability issues and also makes your website accessible for the blind.
A cherry.jpg image/jpeg file is shown as an image by the browser
A logo in svg like https://news.ycombinator.com/y18.svg too
So browsers should show readme.md as text/markdown and clicking on a link inside that file should also redirect to another markdown file, instant wiki
Browsers should implement that asap and we will follow
Practically, Markdown is more of a vibe than a spec and everyone just uses whatever subset makes sense to them.
They’re far too useful for me.
The MD content isn’t intended for human consumption whereas HTML is. So you either have that injection readable to your users, which could be jarring to non-technical readers, or you play the cat-and-mouse game of hoping those pesky crawlers don’t ignore text that’s not human-visible on a rendered page.
Or everything just disappears behind something like cloudflare or a paywall.
The problem is that it’s in the sites interest to be SEOed. But it’s against their interests to be AI crawled.
The open internet deserves a real solution instead of gatekeepers in the form of brightdata, firecrawl, cloudflare and the likes.
Maybe tarpit them into reading gigabytes of Markov Chain nonsense, really slowly?