# Most of our visitors don't have eyes
type: article | language: en | date: 2026-10-18 | author: Tony, creator of Contexte Tech | tags: ai, agents, mcp, llms-txt
url: https://contextetech.com/en/articles/site-human-agent-mode
description: On our site, bots and AI assistants outnumber humans. We added a Human | Agent switch: one page, two ways to read it.
other languages:
  fr: https://contextetech.com/articles/site-mode-humain-agent
  es: https://contextetech.com/es/articles/sitio-modo-humano-agente
  de: https://contextetech.com/de/articles/website-modus-mensch-agent
  it: https://contextetech.com/it/articles/sito-modalita-umano-agente
  nl: https://contextetech.com/nl/articles/site-modus-mens-agent
  pt: https://contextetech.com/pt/articles/site-modo-humano-agente
  sv: https://contextetech.com/sv/articles/webbplats-lage-manniska-agent
  pl: https://contextetech.com/pl/articles/strona-tryb-czlowiek-agent
all articles: https://contextetech.com/en/articles

---

On October 16, out of about 8,900 requests to our site, more than half came from bots and AI assistants: PerplexityBot (2,109), GPTBot (971), Googlebot, crawlers, and 778 calls to our MCP server. The same week, our MCP server received 666 connections from 93 different pieces of software. Resource views by humans range from 0 to 17 a day.

In other words, a large share of our visitors never look at the screen. They read text, follow links and call tools.

So we added a switch to the site header: **Human | Agent**.

## What each one sees

In human mode, nothing changes: cards, tabs, charts, quality scores.

In agent mode, every resource page becomes a few lines of text:

```
# Dataset IA : lire la météo marine relevée par les bouées
code: 1495xut | type: dataset | language: fr | licence: CC-BY-4.0 | version: v1
page: https://contextetech.com/datasets/assistant-meteo-marine-bouees-1495xut
mcp: connect https://contextetech.com/mcp then call load_resource with code "1495xut"
json: https://contextetech.com/api/resources/contextetech--assistant-meteo-marine-bouees
download: …/download (56 rows, JSONL)
source: https://www.data.gouv.fr/datasets/archives-annuelles-des-observations-bouees
citation (required by the licence): …
```

Lists become one line per resource: short code, title, type, language, licence, URL.

## One page, not two

The obvious move was to serve bots a different page. We didn't, for two reasons.

- **Caching.** Our CDN keeps HTML for 60 seconds. Two versions behind one URL guarantee that, one day, a human gets the bot's version.
- **Trust.** Showing bots something other than what humans see is exactly what search engines penalise.

Both versions live in the same HTML. The browser shows one or the other based on the saved choice. An agent reading the page source finds the text version, no JavaScript needed.

## What agents taught us

Looking at what assistants actually did on our MCP server:

1. **Most "visitors" are probes.** Out of 666 connections, a handful come from assistants people actually use (Claude, opencode, Claude Code). The rest are directories checking that the server is up.
2. **They search in languages other than ours.** "marine weather" didn't find "météo marine". We added synonyms across French, English, German and Spanish.
3. **An empty search is a signal.** Assistants searched for "rag prompt" and found nothing. We wrote 18 RAG prompts.
4. **A short code beats a pretty URL.** "load 1495xut" works in any assistant.
5. **Licence and source must be in the text.** An agent doesn't click a "Licence" tab.

## For decision-makers: seven concrete steps

Your buyers no longer just type a search: they ask an AI assistant to compare and recommend. The assistant reads websites for them. **What it cannot read, it does not recommend.** Here is where to start, from the easiest to the most committing.

1. **Take the test, in five minutes.** Ask two or three assistants to recommend an offer like yours, in your sector. Are you cited? Are your price and terms right? That is your baseline.
2. **Decide which bots you let in.** Your `robots.txt` file allows or blocks GPTBot, ClaudeBot, PerplexityBot. Blocking them protects your content but makes you invisible in their answers. Make that choice deliberately, not by default.
3. **Publish a `/llms.txt`.** A few lines saying who you are, what you sell and where to find the essentials. Half a day of work, and it is the first page an assistant reads.
4. **Spell out prices, terms and licence.** Not in an image, not behind a form. Add structured data (schema.org: organisation, product, offer). An assistant recommends what it can cite without risk of being wrong.
5. **Measure that traffic separately.** Ask your team to count, every month, the AI assistants in your server or CDN logs, and which ones. Your usual analytics often miss them. For us, it was more than half of all requests.
6. **Open your data if it fits.** An MCP server lets an assistant query your catalogue, prices or documentation directly. It is your storefront for agents; then list it in MCP directories.
7. **Choose your own AI tools on numbers.** Model, budget, your own server or an API: we gathered free, sourced tools (model vendors, spec sheets, pricing pages) that compute everything in your browser:
   - [which AI model for which task](/en/decide/ai-model-ranking);
   - [how much an AI workload costs](/en/tools/artificial-intelligence-cost-calculator-simulator);
   - [local server or API: the comparison](/en/tools/local-gpu-server-or-api-artificial-intelligence-cost-comparison);
   - [all decisions by use case](/en/decide): extracting data from documents, customer-service chatbots, automation agents.

The hardest part isn't technical. It takes a change of mindset: for twenty years, we built websites for eyes. Now we have to rewire our brains to build for AI agents too, which read without seeing. That is how you deploy solutions for the future, instead of playing catch-up.

And to meet your future reader, try **Agent** mode at the top of this page.