You never integrate GEFT. GEFTBot reads your site once, passively, to learn its facets and query grammar — then GEFT invites you to claim it. If you want your site AI-ready on your own terms, you verify the domain, see what GEFT discovered, and set the terms. Nothing to install, no endpoint to expose, no WAF rule to write.
How a site becomes AI-ready
- Identify — A named agent, signed. GEFTBot declares itself and signs every request. Bot protection can recognize it without asking GEFT.
- Discover — One passive read. A single discovery crawl of public links, sitemaps, and structured data, honoring robots.txt. No forms, logins, or untried parameter combinations.
- Learn — The site's own grammar. GEFT learns the site's facet set and query grammar. AI proposes, probes test, evidence promotes. Active probing only with explicit owner authorization.
- Invite & claim — An invitation, not a toll gate. GEFT invites the site to claim it. The owner verifies the domain. Nothing to install.
- Authorize — Terms the runtime enforces. The owner sets terms. Settings compile into a machine-checkable grant: default deny, fail closed, restrictions always apply.
- Serve — Translated, not guessed. An agent's canonical query is translated into the site's own query language and executed deterministically. No AI at serving time.
- Verify & return — Evidence, not confidence. A typed result returns with evidence attached. What cannot be supported is reported as unsupported, unknown, denied, or stale — never guessed.
Access is not one permission
Five separate axes. Being publicly reachable authorizes none of the other four.
| Axis | Question | Answer |
|---|---|---|
| Reachable | Can a browser open the page? | Says nothing about what a machine may do next. It is a fact about the network, not a permission. |
| Crawlable | May GEFT read the page as itself? | Governed by your crawl directives, a fixed identity, a verifiable signature, and a load ceiling. This is the axis a publisher usually means. |
| Probeable | May GEFT test how the site behaves? | Trying filter combinations to learn a source's grammar generates load and changes state. It requires explicit authorization — always, without exception. |
| Retainable | May GEFT keep what it read? | Reading once and storing a compiled model of your query semantics are separate grants with separate lifetimes. |
| Executable | May GEFT run queries for third parties? | Serving an agent's request against your source on someone else's behalf is the last axis, and the narrowest. |
Probing in particular — trying filter combinations to learn how a source behaves — always requires explicit authorization, and no amount of crawl access substitutes for it.
A named agent with a checkable signature
GEFTBot will not rotate addresses, imitate a browser, or try to look like ordinary traffic. Every request it sends will be signed with a key published at a fixed address, so your origin — or your WAF, without asking us — can decide for itself whether a request is genuinely GEFTBot.
The user agent string is a convenience for your logs. Anyone can copy it. The signature is the identity.
| Field | Value | Note |
|---|---|---|
| Agent name | GEFTBot | One name. GEFT never presents itself as a browser or another product. |
| User agent | GEFTBot/1.0 (+https://geft.ai/bot) |
For readability in your logs. Not a proof of identity — it can be copied. |
| Request signature | Ed25519 · HTTP Message Signatures | The only valid proof that a request came from GEFT. Short-lived, per request. |
| Public key directory | /.well-known/http-message-signatures-directory |
Published on geft.ai so any origin or WAF can verify a signature independently. |
| Crawl scope | One passive discovery crawl | Public links, sitemaps, and structured data. No forms, no logins, no untried parameter combinations, no recrawl. |
| Crawl directives | robots.txt honored, including Crawl-delay | Disallowed paths are not requested. No robots.txt (404) means no restrictions; a server error or timeout means no crawl. |
| Operator | GEFT | A named, contactable operator, not an anonymous crawler pool. |
| Contact | hello@geft.ai | Include the domain and any affected paths in your message. |
Full crawler reference: geft.ai/bot.
Claim your site in four steps
- Invitation — GEFT reads once, then invites you. GEFTBot crawls your site one time, passively, within your robots.txt, and does not come back to recrawl it. GEFT then contacts the site with an invitation to claim it.
- Verify — Prove the domain is yours. Claiming is optional. If you want your site AI-ready on your own terms, you verify ownership of the domain. No account on your side, no code change, no integration.
- Review — See what GEFT discovered. The facet set, the query grammar, and the evidence behind each finding, as GEFT learned them from what your site already shows publicly.
- Set terms — Decide what GEFT may do. Your settings compile into a machine-checkable grant the runtime checks before every request. Default is deny. Restrictions always apply; expanding access, such as active probing, needs your explicit authorization.
To start a claim, email hello@geft.ai from an address at the domain concerned.
The seven controls you set
- What — Which entities, capabilities, and data fields are in scope — and which are not, even where they are publicly visible.
- Who — Which agent identities and which declared purposes are allowed to reach your source through GEFT.
- How — Which access path is used: your official API, a published feed, declared structured data, or ordinary page reads.
- Freshness — How often GEFT may re-read, and how long a compiled answer may be served before it must be re-verified.
- Load — Requests per second, burst, concurrency, and a daily ceiling. Your stated limit is an upper bound, never a starting point.
- Economics — Whether access is free, metered, or contractual. Payment never buys ranking or prominence in a result set.
- Indexing — Whether what GEFT learned may be retained, for how long, and whether it may be used to answer without returning to your source.
A setting is not a request we take under advisement; it becomes a machine-checkable permission that the runtime checks before every request. Anything you have not explicitly allowed is denied.
Principles GEFT holds to
- AI discovers meaning. It does not manufacture truth. AI proposes; probes test; evidence promotes; deterministic execution serves.
- No AI at serving time. Once a site's grammar is learned and authorized, every request runs through deterministic, compiled execution — not a model.
- Public access is not permission. Crawl, probe, and serve are separate grants, each authorized on its own terms. Default is deny.
- Unknown is reported, never guessed. When a result cannot be supported, GEFT says so. It does not fill the gap with a plausible answer.
- A verified result is contextual, not absolute. It means the source supports this answer, in this context, at this time — not that it is universally true.
Contact
Questions about your site, the crawl, or claiming: hello@geft.ai.