# What agents take from an agent-first site, and what they send back (slim)

> Quynh Tran-Thanh tests what small businesses argue about when they build for AI agents (does ChatGPT send customers, do agents read Markdown, does llms.txt do anything, is an MCP server worth it) against six weeks of Anchor Terminal's own traffic and the changes we shipped in that time.

- Full: https://www.anchorterminal.com/blog/what-agents-send-back.md (~4,550 tokens) · this version ~1,980 tokens · JSON https://www.anchorterminal.com/blog/what-agents-send-back.json · canonical https://www.anchorterminal.com/blog/what-agents-send-back
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-11-20
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

By Quynh Tran-Thanh, 20 November 2026. 2,184 words, about 9 min to read.

If you run a small business in 2026, someone has told you to get ready for agents.

We did all of it.

## The research

I started with what people argue about.

- Does ChatGPT send customers? "Tons of people click from ChatGPT! It's now the top referrer on most of my websites" [[2]](#ref-2)
- Does anyone read llms.txt? "LLMs are not reading llms.txt nor AGENTS.md files from servers" [[3]](#ref-3)
- Do agents want Markdown? One person logged the Accept header of ChatGPT's, Claude's and Perplexity's fetchers and wrote "None of them use text/markdown." [[4]](#ref-4)
- Is an MCP server worth building? "Why on earth would you make an MCP to let agents query your docs specifically?" [[5]](#ref-5)

There is some data on the first one.

So I used ours.

- 1,197,467: visitors whose browser said it was a browser
- 11,992: of them that ran our page script, the people
- 285,419: crawler visitors
- 16,256: agent visitors

Anchor Terminal's traffic, 5 October to 19 November 2026. Source [8].

The server sorts every visitor by what it says it is.

I also lined the numbers up against the changes we shipped in those six weeks, since we changed a lot.

## What the data says

### The assistants read a lot and send almost nobody

For each company I counted its bots' requests over the 30 days and the visitors its product sent back [[8]](#ref-8).

| Company | Its bots' requests | Visitors it sent | Requests a visitor |
| --- | --- | --- | --- |
| OpenAI (GPTBot, ChatGPT fetching for users) | 97,336 | 21 | 4,635 |
| Perplexity (PerplexityBot) | 135,338 | 16 | 8,459 |
| Microsoft (Bingbot) | 41,047 | 94 | 437 |
| Apple (Applebot) | 198,247 | 0 | none sent |
| Google (Googlebot) | 4,323 | 19,150 | 0.23 |

Source [[8]](#ref-8).

OpenAI's ratio is five times Cloudflare's 887, and Perplexity's is seventy times its 118 [[6]](#ref-6) [[8]](#ref-8).

That's a little unfair to the AI side, and I want to say how.

I think this is the nature of an answer engine.

### Who reads the Markdown depends on who you call an agent

Every listing, comparison and category page on our site has a Markdown copy, linked from the page itself, and llms.txt tells agents to append .md to any address to get it [[15]](#ref-15).

Comparisons are where agents take the Markdown most.

But the agents that name themselves are a small part of who reads the Markdown.

The clearest sign came from a bug fix.

### Files for agents are barely read, a listing where agents look is

llms.txt, llms-full.txt, our /agents page and the files under /.well-known/ together were 0.43 per cent of agents' requests, and 0.05 per cent of crawlers' [[8]](#ref-8).

The files agents did look for and didn't find were more telling.

The change that moved agent traffic was the MCP registry.

A day, before and after the MCP registry entry on 31 October

| Measure | 5 to 30 October | 1 to 19 November |
| --- | --- | --- |
| MCP requests | 3 | 3076 |
| Agent visitors | 234 | 514 |

Each series has a zero-based scale.

Agent visitors more than doubled and agents' downloads went from about 1 GB a day to 4.6 GB [[8]](#ref-8).

### Google is still the one that sends people, and one mistake cost most of it

This one we learnt the hard way.

Two things in the data stood out.

### Most of your visitors aren't people

Of the visitors whose browser said it was a browser, a median of 0.96 per cent a day ran our page script [[8]](#ref-8).

Names lie too.

## Key takeaways

- AI assistants read a site built for them thousands of times for every visitor they send. Google sent us more visitors than every AI assistant together.
- Agents that name themselves took the Markdown for about a third of pages in October and a sixth in mid November. Most of the Markdown went to crawlers and to clients calling themselves browsers.
- llms.txt and the other files for agents were under half a per cent of agents' requests.
- Being listed in the MCP registry took agent traffic from a trickle to thousands of requests a day overnight.
- One sitemap with too many pages cost us most of our Google visits in a week.
- Only about one in a hundred "browsers" in our logs was a person.

## What it means if you want agents to find your business

Build for agents because they read you, not because they'll send you customers.

Go where agents look for tools.

Make the HTML readable and keep the Markdown.

Be careful with Google.

Count people by what they do, not by what they say they are.

What I couldn't settle is whether agents that read you without sending anyone still bring you business.

---

More in [From the data](https://www.anchorterminal.com/blog/from-the-data/index.md) · Previous from the data [Two handshakes.

- Older: [OpenAI's agents now finish half its two-to-four-hour research tasks](https://www.anchorterminal.com/blog/openai-agents-half-the-research.md)
- Newer: [What to build that lasts. The jobs agents ask for that nobody finishes](https://www.anchorterminal.com/blog/the-jobs-nobody-finishes.md)
- All posts: https://www.anchorterminal.com/blog/index.md

Anchor Terminal grades the tools AI agents run on, with a source for every score.

## References

[1] Anchor Terminal, "How to make your site and tool ready for agents", 15 October 2026. Source for our advice to write an llms.txt and a Markdown copy of every page.

[2] Hacker News, comment by timpera, 13 October 2026. Source for the quote on ChatGPT as the top referrer.

[3] Hacker News, comment by reconnecting, 18 February 2026. Source for the quote on llms.txt not being read.

[4] Hacker News, comment by startages, 20 April 2026. Source for the quote on the fetchers' Accept header.

[5] Hacker News, comment by drbscl, 20 November 2026. Source for the quote on building an MCP server for docs.

[6] Cloudflare, "A deeper look at AI crawlers: breaking down traffic by purpose and industry", 28 August 2025. Source for the crawl-to-refer ratios of about 50,000 for Anthropic, 887 for OpenAI and 118 for Perplexity.

[7] Ahrefs, "AI Chatbot Traffic: What It Is, and How to Get More", 15 May 2026. Source for AI chatbots' 0.28 per cent of traffic across 74,752 sites, and AI visitors' 0.5 per cent of Ahrefs' visitors and 12.1 per cent of its signups.

[8] Anchor Terminal, the server's daily traffic summaries on the admin page, 5 October to 19 November 2026, read on 20 November 2026, not published.

[9] Anchor Terminal, "How to build an agent-first website, from our first six weeks and 5 million page views", 19 November 2026. Source for how we tell people, agents and crawlers apart.

[10] Anchor Terminal, the hypotheses for this post, written and committed on 20 November 2026 before the analysis ran, not published.

[11] Anchor Terminal, deploy notes and commits of 31 October, 8 November and 15 November 2026, not published.

[12] SE Ranking, "LLMs.txt: Why Brands Rely On It and Why It Doesn’t Work", 15 October 2026. Source for 10.13 per cent of nearly 300,000 domains having the file and no effect on citations.

[13] Google Search Central, "AI Features and Your Website", updated 10 December 2025. Source for Google saying no machine-readable or AI text files are needed.

[14] Hacker News, comment by ChillyCapy, 12 August 2026. Source for the quote on fake Googlebot visits.

[15] Anchor Terminal, llms.txt, read on 20 November 2026. Source for the instruction to append .md to an address, or send Accept: text/markdown, to get a page's Markdown copy.
