Agentic Web | Reading time: 5 minutes | November 2026

From invisible to AI agents to built for them

In July 2026, this translation agency’s website scored 9.2 out of 10 for human SEO and 5 out of 10 for agent readiness, on the same day, on the same page. Three weeks later it had a layer built for AI agents: plain-language discovery files, a read-only API and a live MCP server with 11 tools. Five independent AI platforms have since confirmed the build.

The client is an established Swedish language and translation services provider. The company’s name is withheld to protect client confidentiality.

3/3
Agentic Browsing checks
5/5
AI platforms confirmed it
11
MCP tools, read-only
3 wks
from audit to live

3/3

Google Agentic Browsing checks passed, up from 2 of 2 applicable

5/5

AI platforms independently confirmed the build

11

read-only MCP tools live for AI agents

3 wks

from discovery audit to a live, verified build

Agentic Browsing: Google’s agent-readiness category in PageSpeed Insights.

About the client

An established translation agency with a strong human website

The client is a Swedish language and translation services provider that has worked with businesses for almost three decades. It covers legal, technical and medical translation between Swedish and more than 30 languages, and buyers typically arrive with a document, a language pair and a deadline.

The website was built well for people. It loaded fast, was well organised and scored highly on traditional SEO. What it lacked was any layer built for the AI agents that increasingly do the research before a person ever visits.

Opportunity

Be the source an AI cites when buyers ask who can translate it

Buyers increasingly start supplier research in ChatGPT, Claude, Perplexity or an agentic browser: “Who can translate a legal contract into German?” The assistant reads a handful of sources and writes one answer. If it cannot read a translation agency’s services cleanly, it guesses, or leaves the agency out of the answer altogether.

In July 2026 we audited the site the way a growing share of the web now experiences it: as software trying to extract a clean answer. The two disciplines came back very differently.

Before · July 2026 · ChatGPT audit
Same site, same day, out of 10
Human SEO
9.2
LLM readability
8.5
AI discoverability
7.5
Agent readiness
5.0
Overall LLM friendliness: 7.8/10. Traditional SEO had been solved for years. The second discipline had barely started.

Google’s Agentic Browsing audit told the same story. It passed 2 of 2 applicable checks, because without an llms.txt the third check was not even being run.

Together with the client, we set three goals. Discoverable: canonical company and service facts in the files AI agents look for first. Readable: services, FAQs, case studies, team and articles available as clean, structured data. Accurate: explicit rules for what an AI may cite, how it describes pricing and what it must never invent. And no rebuild: the human website stays exactly as it is.

Deliverables · one site, two audiences
Discovery audit → live on 28 July 2026
Incoming request
ChatGPTClaudePerplexityGeminiGrok
“Who can translate a legal contract into German?”|
Website for humans
Unchanged
Imagery, layout, brand and conversion paths stay exactly as they were. Visitors see no difference.
JavaScriptMediaLayoutForms
Website for AI agents
New agentic layer
# Translation agency
Legal, technical and medical translation, Swedish and 30+ languages.
## Services
– structured, token-light, citable
llms.txtllms-full.txtai.txtrobots.txt/agentJSON APIOpenAPIMCP · 11 tools
No JavaScript, no media weight: agents parse the layer in one pass.

Strategy

Three layers. Read, understand, cite.

We built the agentic layer in three parts, each answering a question an AI agent asks when it arrives at a website. Everything is read-only: nothing writes back into the client’s systems.

Three layers, one agentic web
Layer 01 · Read
Discover and read · llms.txt, llms-full.txt, ai.txt, robots.txt
Google Agentic Browsing: 3/3
Layer 02 · Understand
Understand the services · JSON API, OpenAPI, MCP server
MCP server found by all five AI platforms
Layer 03 · Cite
Cite it correctly · /agent page, citation and pricing rules
/agent page confirmed by 4 of 5 platforms

All three layers sit alongside the existing website.

Workstream · Read
01

Discover and read

Before, agents parsed pages built for people and found no llms.txt. Now machine-readable summaries and explicit crawler rules hand them the canonical facts first.
llms.txtllms-full.txtai.txtrobots.txt
✓Google Agentic Browsing: 3/3
Workstream · Understand
02

Understand the services

Services, FAQs and case studies used to live inside page layout. Now a read-only JSON API and a live MCP server return them as structured data.
JSON APIOpenAPIAPI catalogueMCP · 11 tools
✓MCP server found by all five AI platforms
Workstream · Cite
03

Cite it correctly

Nothing used to tell an agent what it could quote. Now a dedicated /agent page sets what an AI may cite, how to describe pricing and what it must never invent.
/agentCitation rulesPricing guidance
✓/agent page confirmed by 4 of 5 platforms

All three layers sit alongside the existing website. The human website was not rebuilt.

Execution

From a buyer’s question to a documented answer

Most websites leave an AI agent to guess. Here, an agent that lands on the site is pointed straight to a documented answer. The MCP server exposes 11 callable tools covering services, FAQs, case studies, team and articles, all read-only by design.

Request flow · read-only end to end
01
A buyer asks
“Who can translate a legal contract into German?”
→
02
The agent discovers
llms.txt and the /agent page point it to the documented layer
→
03
The agent calls
A read-only API endpoint or MCP tool returns the service data
→
04
The agent cites
Services, languages and the next step, from the source

There is no JavaScript to execute and no layout to interpret: the agent parses structured, token-light content in one pass. The human website still serves visitors exactly as before, with its imagery, brand and conversion paths untouched.

Example: an agent answers a buyer
Who can translate a legal contract from Swedish into German?
I’ll check the translation agency’s own service data rather than guess. One moment.
Great. And how do we get a quote?
→ MCP tool call: service lookup · read-only
✓Legal translation, Swedish to German, answered from the agency’s documented services
✓Next step taken from the /agent page: request a quote on the agency’s website
Illustrative, anonymised example. Wording simplified; tool names and company details removed.

The next milestone
Read today. Act next.
ReadSHIPPED
An agent can describe a service, pull a case study or explain the pricing model accurately.
→ 11 MCP tools, all read-only by design
↓ same infrastructure ↓
ActNEXT
Turning a subset of tools actionable, so an agent can transact instead of describing how a person would.
→ Submit a quote request directly
→ Kick off a document upload

The result

Discoverable, readable and verified, measured three ways

We measure whether the work succeeded, not just whether it was delivered. Three independent checks answer three different questions.

Can an agent navigate the human website?
Google’s Agentic Browsing audit in PageSpeed Insights.

Do AI platforms find and use the agent layer?
Five independent LLM audits.

Is it safe to let agents in?
Read-only by design, across the API and all 11 MCP tools.

Measure 1 · Google Agentic Browsing

Lighthouse now adds a fifth category to every PageSpeed Insights report, next to Performance, Accessibility, Best Practices and SEO. It checks whether an agent can navigate and read the page.

Before · 3 Jul 2026
2/2
applicable checks, one not run
→
After · 10 Aug 2026
3/3
checks passed
Check
Before
After
Well-formed accessibility tree
Passed
Passed
Low Cumulative Layout Shift
Passed
Passed
llms.txt follows the recommendations
Not run
Passed

Source: Google PageSpeed Insights, 3 July and 10 August 2026. The llms.txt check does not run when the file is absent.

Measure 2 · Five AI systems, one question

Each model assessed the finished build independently, with no knowledge of the others’ answers. Verdicts are in their own words; ticks show what each report explicitly confirmed.

Claude
Best-in-class
ChatGPT
9.5/10
Grok
Strong yes
Perplexity
8.5/10
Gemini
Confirmed
llms.txt discovery file
✓
✓
✓
✓
✓
Machine-readable /agent page
✓
✓
✓
✓
—
Public REST API + OpenAPI
✓
✓
✓
✓
—
ai.txt permissions file
—
✓
✓
✓
—
AI-crawler-aware robots.txt
—
✓
✓
—
—
Live MCP tool server
✓
✓
✓
✓
✓
Clear company details
✓
✓
—
✓
—

✓ explicitly confirmed · — not mentioned, which is not the same as failed. Highlighted rows were found by all five.

Measure 2 · Found by every model

Two signals, five out of five

Of seven agent signals checked, only two were found independently by every model.

5/5
llms.txt
5/5
MCP server
Measure 3 · Read-only by design
11
MCP tools live
0
write actions

Since 28 July 2026, AI agents can call the API and MCP server directly: services, FAQs, case studies, team and articles. Every tool returns data; none of them writes back into the client’s systems.

What it means

What the build showed

The project showed what an agentic layer can do for a service business that already wins on traditional SEO.

01
No rebuild needed
The agentic layer sits alongside the existing website, which visitors see unchanged.
02
Measured, not asserted
Scored by Google’s own audit and confirmed independently by five AI platforms.
03
Fast to ship
From an audited gap to a live, verified build in about three weeks.
About the author
Gaurav Khetan
Gaurav Khetan Head of Development Gaurav is Head of Development at GO MO Group.

Is your website ready for AI agents?

Tell us your domain and we’ll come back with what AI agents see today, and what it takes to make your site readable, citable and actionable.

Get AI summary of GO MO Group