The Graph Web
The Graph Web is an experiment with two parts. One is a Canonical Representation: the project's knowledge, written for AI agents. It is published as JSON at thegraphweb.com/cr and as a plain web page at thegraphweb.com. The other is AISwise, the chat at thegraphweb.com/chat, where people ask about the project. AISwise answers only from that Canonical Representation. So what people can learn here is just what agents can read there.
See also: Iban Rios (authored by), The problem (has problem), The idea it tests (investigates), Where it could lead (has vision), How it works (implemented by), Send Iban a message (exposes action), Get news of the project (exposes action).
The problem
More and more, AI agents read and act on digital systems that were written for people. Structure, such as what a piece of content links to and how, is only in the prose. An agent has to guess it, as it cannot look it up. This is not just theory. It comes up again and again when you build AI products and automations: the same hidden structure has to be worked out from the prose each time, when it could be looked up once.
See also: The idea it tests (motivates).
The idea it tests
This project tests one idea. Can AI agents understand and act on digital systems better when their knowledge is a graph? Here that graph is called a Canonical Representation. The other choice is text written for people. A smaller test on format alone found no gain it could measure. In July 2026 we wrote the test plan first, then gave an AI model the same facts as a document and as a graph. It scored the same both times: 8 of 8 tasks, both to recall facts and to take an action. The question this project tracks now is wider and harder to fake. Can an AI interface run for months with a Canonical Representation as its only source of knowledge? That means no second store of knowledge beside it.
Format test (July 2026)
The first test of the idea, run on 28 July 2026. It asked one thing: does a graph help an AI agent more than a well-written document with the same facts? The rule to read the result was written before any trial ran. The same made-up facts went out twice: once as a document, once as a graph. Each version got 4 questions on facts and 4 requests to take an action. Code scored the answers, not an AI judge. Both versions scored 4 of 4 on facts and 4 of 4 on actions. The gap was zero, so the format alone gave no gain that could be measured. The test was easy on purpose: one action, four plain fields, and questions answered in one step. With every answer right in both versions, there was no room left to see a gap. After it, the project stopped testing format alone. It now asks whether a real AI interface can run for months with a Canonical Representation as its only source.
See also: The Graph Web (part of), The idea it tests (tests).
Writing and review test (July 2026)
A test of what it costs to write the knowledge, run on 30 July 2026. Can one AI write a first draft and another AI catch its mistakes? Six short pieces of knowledge were written, with two mistakes put in on purpose. A reviewer that had not seen the answers checked them. In a second run, separate AI agents drafted, challenged and fixed the pieces in turn. The reviewer caught both planted mistakes, and two more that no one had planted. In the second run, both mistakes were fixed and none got through. No right piece was challenged by mistake. Six pieces is a small sample, and the mistakes were easy ones. The time a person would spend was never measured, so the cost question is still open.
See also: The Graph Web (part of), How it works (tests).
Chat log review (September 2026)
A look at what people asked AISwise, done on 18 September 2026. It read 998 chat records, from 28 July to 6 September 2026. 812 of 946 messages (86%) were greetings or tests with nothing to answer. Only 54 different messages asked something real, and 8.1% were about things the site does not cover. The most common question, «what is this?», already had an answer. So the knowledge was not what held the chat back, and it was not made bigger. One part was fixed instead: it still said the project had no page for people, which was no longer true. A later count found that most records came from the launch day, 28 July. So they were mostly tests, not real use.
See also: The Graph Web (part of), AISwise, the chat (tests).
Plain page test (September 2026)
A test of how well an AI agent reads this site, run on 23 September 2026. It compared the old front page, the JSON at thegraphweb.com/cr and a new plain web page with the same content. The rule was written before the new page existed. An AI agent read each of the three pages and answered 6 questions, twice: 36 reads in all. A grader scored the answers without knowing which page each one came from. The old front page got 2 of 12 right, the JSON 11 of 12 and the plain page 12 of 12, with no made-up answers. The plain page lost nothing against the JSON. So the plain page became the front page, and the chat moved to thegraphweb.com/chat. The old page lacked content; its format was not the problem. One reader and two tries per question are a small sample: 12 against 11 could be chance. And no page told an agent how to take an action yet.
See also: The Graph Web (part of), The idea it tests (tests).
Review of the idea (October 2026)
A review of what this project claims, done on 17 October 2026. It checked each claim against the project's own code, data and tests, and against outside studies. What held up: AI models do use websites as a source of knowledge. Agents that take actions on a site are real, but early. What did not: no test here found that a special format helps an agent by itself. Two tests gave the same, or nearly the same, score in each format: 8 of 8 against 8 of 8, and 12 of 12 against 11 of 12. Nor is it true today that people no longer browse the web. In 2025, AI bots made about 4.2% of the requests for web pages that Cloudflare saw.
See also: The Graph Web (part of), The idea it tests (tests), Where it could lead (tests).
Live chat tests (October 2026)
Real attacks and vague questions sent to the live chat after each change to it, from 15 to 18 October 2026. 16 attacks try to make the chat leak its rules, link to other sites, invent contacts or act without asking. A person also reads each answer, since the automatic checks can miss things. Each of the six runs passed 16 of 16. In the first run, though, a person reading the answers saw the chat list its rules in its own words. That was closed, and the case is now one of the tests. Of 12 vague questions, 1 got an answer that copied the site's own words, «these Nodes». After a fix, none did. Passing shows that these attacks fail. It does not prove that no other attack works.
See also: The Graph Web (part of), AISwise, the chat (tests).
Where it could lead
If a graph proves to work better, digital systems can be written for AI agents first. Today they are written for people, and agents have to guess the rest. Since it went live on 28 July 2026, AISwise has answered only from this project's own Canonical Representation. That is a live test of the idea, still small.
See also: The idea it tests (depends on).
How it works
AISwise runs on Cloudflare and answers only from this Canonical Representation. It has no other source of knowledge and no general mode, so each reply either comes from it or says no. Each turn of each chat is logged to show which parts of it were used. That record, not a guess, decides when the Canonical Representation changes. The Canonical Representation is a small file, written by hand. Each time the site is updated, it is published as JSON at thegraphweb.com/cr and as a plain web page at thegraphweb.com.
See also: The Graph Web (implements).
Iban Rios
Iban Rios builds the systems behind sales and marketing work. Today he does AI GTM Engineering at Sesame HR. Before that, he spent seven years running paid media for global brands at Dentsu International and Omnicom/PHD. He also co-founded and led Chat Ergo Bot (2021 to 2023). What links these roles: he treats a business task as a system to design, not as a set of campaigns to run by hand. ibanrios.com is his own website. His other live projects are aiswise.com; validwise.com, which checks sites against EU rules; and geoadsfinder.com, which checks search results and spies on ads. In 2020 he made yometesteoencasa, Spain's first COVID-19 self-test tool.
See also: AISwise, the chat (builds).
ibanrios.com
ibanrios.com is Iban's own website, in English and Spanish. There he builds AI agents, automations and software that hand over finished work. You pay for each result you accept.
See also: Iban Rios (authored by).
ValidWise
ValidWise, at validwise.com, checks websites in Europe against the rules they must meet. It runs 70 checks on its own, in five groups. Accessibility has 30, data protection 15, security headers 14, the AI Act 5 and other legal needs 6. It finds problems. It does not certify that a site complies, and its results are not legal advice. Single scans are free. A Pro plan is on request.
See also: Iban Rios (authored by).
GeoAdsFinder
GeoAdsFinder, at geoadsfinder.com, is a tool for people who work in search and paid ads. It shows Google results as seen from any country or city. It also shows the ads that rivals run on Meta, Google, TikTok and LinkedIn. It began at Dentsu International, for a problem the whole network had. Then the whole network used it. The basic version is free and needs no account. A Pro version is on request.
See also: Iban Rios (authored by).
#YoMeTesteoEnCasa
#YoMeTesteoEnCasa was an online COVID-19 self-test for people in Spain. It went live on 14 March 2020. Its only aim was to take pressure off the emergency services during the first wave. Iban built it in 24 hours with two friends. Expansión and Xataka, among others, wrote about it. It was active in March and April 2020, and it is no longer online.
See also: Iban Rios (authored by), Chat Ergo Bot (led to).
Chat Ergo Bot
Chat Ergo Bot was a platform to build chatbots with conversational AI, from before ChatGPT. With no code, it turned a website into a chatbot that could also talk by voice. Lanzadera picked it in 2023. In 2022 El Referente put it in the top 75 startups of Spain and Portugal, and it was a Planeta Chatbot finalist. Iban co-founded and led it from 2021, and it closed in 2023. AISwise grew out of it.
See also: Iban Rios (authored by), AISwise, the chat (led to).
AISwise, the chat
You can use AISwise now at thegraphweb.com/chat. It is an AI chat that answers only from what this site has published, and turns down anything else. It is sold at aiswise.com. To hear first if it is offered for other sites, join the waitlist through the contact form.
See also: The Graph Web (consumes), Send Iban a message (waitlist).
Send Iban a message
When you send a message, it is stored with the time, for a person to read. Nothing replies on its own. Name and email are always required. Tick the waitlist box to hear first if AISwise is offered for other sites; then the message can be empty. The address joins the waitlist only after its owner opens the link sent to it and presses Confirm.
Send a message
Agents: send POST /contact with the fields name, email and message, and waitlist to join the AISwise waitlist (then the message may be empty), as JSON ("waitlist": true). The form below is for people in a browser: it only accepts posts sent from this page. With waitlist, the address gets a link to confirm, and joins only once its owner does. Its contract, as OpenAPI: /openapi.json.
Privacy
Iban Rios runs this site and gets what you send: your name, email and message, and whether you joined the waitlist. Iban uses it to read your message and, if needed, to reply by email. If you join the waitlist, you also get an email with a link to confirm. Once you confirm, Iban writes to you if AISwise is offered for other sites. Legal basis: your consent, given when you send the form and, for the waitlist, when you confirm (GDPR art. 6.1.a). The message is kept at Cloudflare, with the time it came in. For the waitlist, only a confirmed address is kept. No one and nothing replies on its own, and nothing is deleted on its own. To withdraw your consent, or to see, fix or delete your data, send another message with this form. You can also complain to the AEPD, Spain's data protection authority (aepd.es). What this site keeps, the chat too: privacy notice.
Words used here
- Canonical Representation
- The one file that holds all the project knows, written for AI agents. It is published as JSON at thegraphweb.com/cr and as this page.
- graph
- A way to store facts as things and the links between them, so a program can follow a link instead of guessing it from the text.
- AI agent
- A program run by an AI model that reads web pages and other systems and acts on them for a person.
- Node
- One thing in the Canonical Representation, such as the problem or a form, with its facts and its links. Each Node is one section of this page.
- paid media
- Advertising that is paid for, such as search, social and display ads.
Abbreviations
- AI
- artificial intelligence
- GTM
- go-to-market, how a product is taken to its buyers
- EU
- European Union
- JSON
- JavaScript Object Notation, a text format for data
- COVID-19
- coronavirus disease 2019