What The Direct Index is, who runs it and who pays for it

The Direct Index measures whether AI assistants can find, read and act on a business's own website, and where they send people instead.

What it does

A scanner for any business website. The scanner visits a site the way an AI agent would and checks four things: whether an agent can reach it, read what the business offers and book or order there, and whether assistants send people to it. You get a score from 0 to 100, with each finding explained and a fix for each one, in English and in the languages of the business's country. It's free and takes under a minute.

City indexes. A city index takes every business of one kind that has its own website, from OpenStreetMap, and runs the same scan on each. It publishes leaderboards, the share of sites that turn crawlers away, the booking and ordering vendors and the website builders in use, and how all of that changes from month to month. Cities can be compared and ranked, and the data is open.

What the assistants say. ChatGPT, Claude and Gemini get the questions people ask them, and the index records which businesses they name and where they send people: the business's own site, a marketplace or a directory. For restaurants the questions go out every month, in the city's language every other month. Other kinds of business are asked every quarter. Because the questions repeat on that schedule, the index builds up a public record of where agent traffic goes.

The direct record. For every business in the index, there's a record of what an agent needs to go direct: the booking link, the ordering link, the menu, the hours, the phone, the vendor behind each one and whether that vendor's terms let an agent use it. Owners can correct the record by claiming their page, and a browser checks any corrected booking or ordering link. Agents read the record through the API, the MCP server, the A2A agent and a daily feed for each city.

The numbers so far

Counted on 18 October 2026 from the index's own statistics, which update every hour. Dead domains don't count as scanned. Scanning started on 13 October 2026. The assistants' questions reach each index on a schedule, so the third number grows one index at a time. Every page gives the sample and the date behind its figures, and a score stays provisional until all four of its parts are measured.

Who runs it

I'm Antonio Curado. I live in Brussels and build this on my own.

If something here is wrong, email hello@directindex.org or message me on X (@ab_curado).

I wrote most of the code with AI coding tools. A score is never decided by an AI. It comes from fixed rules, written down on the methodology page. Known mistakes are on the corrections page.

How it is paid for

I pay for it myself. There are no investors and no ads, and no money comes from the marketplaces, booking vendors or website builders the index measures.

Reading is free: every report, every city page, the comparisons, the rankings, the search, the open data, the API, the MCP server and the A2A agent, with no account. Four things are sold, and their prices are small on purpose. Each question sent to ChatGPT, Claude or Gemini costs a few cents and the servers run every day, so the parts that ask the assistants for someone, or do extra work for them, cost a few euros. That money pays those costs:

• The deep report, 5 EUR once: ten questions to each assistant about one business, their full answers, the nearest similar places, a fix plan and a PDF.
• The owner plan, 2 EUR a month or 19 EUR a year: the assistants asked again every month with the differences, weekly rescans, alerts, the menu as a page and a monthly PDF.
• The agency plan, 1 EUR a month per client site, from 10 sites: a client list, bulk scans, white-label reports, a lead finder and an API key with a higher scan limit.
• Vendor verification, invoiced per vendor: the action test on every site using a booking or ordering vendor, and a "works with agents" mark once it passes, stating the test and the date.

The score is computed the same way for everyone. What a plan buys is more frequent measurement, weekly scans and the assistants asked every month, which can move a score from provisional to measured sooner. Nothing paid changes how points are counted, a rank or a listing, and no placement is for sale.

Where the data comes from

• Business listings come from OpenStreetMap contributors, under the ODbL licence, and from the Overture Maps Foundation's places (overturemaps.org), under the CDLA-Permissive 2.0 (Foursquare's records are under the Apache License 2.0, and AllThePlaces under CC0). When both list a business, the OpenStreetMap listing is kept and Overture adds the website if it was missing. Each report names its sources. The index is published under the ODbL, with the notice its sources require.
• Scan results come from the scanner's own visits. A visit is up to fourteen page requests plus a dozen small requests for well-known files, about thirty in all. Each request carries signed headers naming directindex.org, so any site can verify the visit. A robots.txt rule for DirectIndexBot stops the visit, and the deeper pages are only fetched where robots.txt allows. The home page is fetched the way any visitor's page view is. When the plain fetch is turned away, a real browser takes a second look. The browser also runs the action test, which stops before the confirmation step. The scanner never submits a form, places an order or makes a booking.
• The assistants' answers come from the public APIs of OpenAI, Anthropic and Google, with web search on, and the questions are put the way people ask them. Each answer is stored with the model that gave it and the date.
• Vendors' terms are read by a person, and each verdict gives the page it was read from and the date. If a vendor thinks a reading is wrong, they can write in and it's corrected the same day.
• An AI model is used in five places. It sorts the websites in the assistants' answers into own site, vendor, marketplace, directory or press, once per website. In each round, it writes one dated sentence per business on why an assistant sent people elsewhere, using only the stored answers and the scan's facts. Each monthly snapshot gets four sentences of commentary written from the page's numbers, and every number is checked. It also turns a menu PDF into a page. And the ask page answers questions using the index's own tools. None of this touches a score, and a wrong judgement can be reported by email.
• No data is bought from data brokers. The only personal data kept is the email address someone gives to claim, correct or buy something, and the contact details OpenStreetMap contributors recorded for a business. The privacy page says what happens to it.

If a page is about your business

Every business page has a free rescan and a claim form. A rescan settles most disputes. The dispute button on every report opens the contact form below with the business already filled in. Claiming lets you correct the booking and ordering links, the menu and the hours yourself, through a signed link sent by email, with no account. A browser checks a corrected booking or ordering link before the record calls it verified. The wording on every page stays factual, so a site is "not reachable by agents", never "bad". If you own a site and want its page removed, ask and it will be.

This site

This site publishes everything its own readiness check looks for: a sitemap, crawler rules, Content Signals, Markdown for every page, an API catalog, an MCP server, an A2A agent, published skills and signed requests. It has no accounts, so there's no OAuth to publish. Cloudflare's independent checker rates directindex.org Level 5, Agent-Native. The ask the index page answers questions with the same tools the MCP server gives agents, so the site itself uses the tools it offers.

Principles

1. No marketplace money

   Marketplaces don't pay for it, and neither scores nor placements are for sale.

2. Public method

   The methodology is public and versioned. The data is published under an open licence.

3. Open to correction

   Provisional scores are marked, and pages state their samples and dates. Every page has a free rescan and a dispute form.

4. It passes its own test

   This site meets the standards it checks for, and any site its crawler visits can verify the visit.

Contact