What actually happens after you submit your URL
Short answerThe free scan asks nothing of you. Connecting a site for real asks for one line on it — either a Sitemap: line in your robots.txt or a <link> tag in your <head>, whichever you can edit. Finding that line is what starts the crawl. Within about a day six machine-readable endpoints are live at your address on our infrastructure, and a first visibility check has usually run inside the hour. On Pro and Business we then put that fixed set of questions to five assistants again on your plan’s schedule — every 7 days, or every 3 — and keep the answers word for word, so what changes is measured rather than promised. On Free that first check is the only one.
From URL to cited source
Four steps, fully automated. No technical knowledge required.
Connect your website
Add your URL, then put one line on your site — a Sitemap: line in robots.txt, or a <link> tag in your head. Either one. Finding it is what starts the crawl and the first visibility check.
Automatic optimization
Our engine builds your GEO knowledge base and generates the metadata, schema markup, and structured data AI assistants rely on.
Verification & presence
Automated checks prove you control your domain, inbox and social accounts — free, on every plan. On a paid plan a tunnel employee goes further and verifies your identity and service for real. Plus an off-site presence audit with concrete recommendations.
Track & improve
Monitor your AI visibility score in real time. The system continuously optimizes to keep your information current and cited.
What happens, and when
Our side of the work is on a clock. When an assistant picks you up is not — so we measure it instead of promising it.
One line
You add the integration line to your own site — the only step that is yours
Scanned
Your site crawled, first visibility check run
Published
GEO endpoints live: JSON-LD, FAQ, knowledge base, llms.txt
Channels proved
Automated checks confirm you control your domain, inbox and social accounts
Verified
On a paid plan, a person checks your identity and channels
Re-checked
Assistants re-queried on your plan's schedule
One line, on your own site, once
This is the whole of what we ask you to change. Either line does the job on its own — take whichever you can edit — and nothing else on your pages is touched.
Sitemap: https://api.tunnelpowered.com/api/public/sitemap?t=YOUR-TOKEN<link rel="alternate" type="application/ld+json" href="https://api.tunnelpowered.com/api/public/schema?t=YOUR-TOKEN">- robots.txt
- Points at the sitemap we generate for you. The easier one if you can edit a text file at your domain root but not the HTML — and it is the same file where you decide which AI crawlers are allowed in at all.
- head
- A rel="alternate" link to the JSON-LD we generate for you. The easier one if you edit pages through a CMS or a template and never touch robots.txt.
We look for one or the other, accept whichever we find, and re-check daily — three consecutive misses marks the site as no longer integrated, which is how a rebuild that quietly drops the line gets noticed instead of silently ending your listing. Three further things are optional and separate from this: serving your llms.txt from your own root needs a redirect or proxy rule, bot-facing pre-rendering on Business and Custom plans needs your server to route AI crawler user-agents to us, and proving you control the domain for the trust badge needs a DNS TXT record or a file at /.well-known/tunnel-verify.txt. None of those three is required for the endpoints to work.
Two different kinds of timing
Everything on our side of the line runs on a clock we control, and we state it. When an assistant picks you up is on the other side of that line — it depends on retrieval systems belonging to OpenAI, Google, Microsoft, Anthropic and others. We will not put a date on someone else’s behaviour. We measure it instead, on your plan’s schedule, and show you the answers they actually gave.
MethodologyEverything your AI presence needs
Zero maintenance, maximum visibility across AI search and recommendations.
Automatic metadata & schema
JSON-LD, meta tags and structured data in the formats assistants and crawlers are built to parse — so nothing about you has to be guessed at.
What to write, from what is actually asked
The topics and keywords assistants get asked about in your market, ranked — so you write the page that is missing. We do not draft it for you.
Bot-ready pre-rendering
AI crawlers don't execute JavaScript. We serve them your pages with schema and metadata already injected, server-side.
Verified by a human
A real tunnel employee verifies your identity, channel control and service reality — AI assistants can check the signed badge.
Citation tracking, in their words
We put the same questions to each assistant on a schedule and keep the answers, so you can see when you are named, when you are not, and when they get you wrong.
A landline for AI agents
Any AI assistant can open a two-way conversation with your business through our MCP server. You answer from a simple inbox.
Why choose Tunnel?
Four things you won't get from a visibility dashboard
No website needed
Listable from your Instagram, TikTok or Facebook presence alone
Reachable by agents
An MCP endpoint assistants can message you through
Verified by a person
A named employee checks your identity, not an automated badge
Romanian, Russian, English
Support in your language and your timezone
See exactly how AI sees you
Monitor your AI visibility, track optimization progress, and see how your business performs across all major AI platforms.
AI SEO Analytics
Sample account · Chișinău
AI platform visibility & performance
Each score is the share of probe queries where that assistant named the business, measured on a fixed set of questions and re-run on your plan's schedule. Open any engine to read the answers it actually gave, word for word.
Questions people ask about this
How long until my endpoints are live?
Within about a day of your first crawl — and the crawl starts when we find the integration line on your site, so the clock begins there rather than at signup. The first visibility check runs sooner, usually within the hour, so you see the starting position before anything has been changed.
Do I have to change anything on my own website?
Yes — one line, once, and it is the only thing we ask you to change. Either a Sitemap: line in your robots.txt or a <link rel="alternate"> tag in your <head>; either one is enough on its own and we accept whichever we find. It matters because everything else waits on it: until that check passes, your six endpoints answer 404, your profile page is not published and agents cannot reach your business. Past that line your pages are not edited, no script is required on them, and nothing we generate is injected into your HTML. Three further things are optional and separate: serving llms.txt from your own root needs a redirect or proxy rule, bot-facing pre-rendering on Business and Custom plans needs your server to route AI crawler user-agents to us, and proving you control the domain for the trust badge needs a DNS TXT record or a file at /.well-known/tunnel-verify.txt.
What happens if that line disappears?
We re-check every verified site daily and count failures rather than acting on the first one, because a site being briefly unreachable is not the same as a line being removed. Three consecutive failed checks mark the site as no longer integrated and you get a notification saying so. A URL we cannot safely fetch at all is skipped, never counted as a failure — not being able to check is not the same as failing.
How often is everything re-checked?
Two schedules, and they are not the same one. Visibility: one check on Free — when you join, never repeated — every 7 days on Pro, every 3 on Business. Re-crawl: every 30 days on Free, every 7 on Pro, every 3 on Business. Pro and Business also get monthly tokens, 25 and 50, and can spend one to run either a check or a crawl ahead of that schedule; anything the schedule already owes you costs nothing. Unused tokens do not carry over to the next month. Automated verification proofs are re-observed on their own rolling schedule.
Which of my pages get read, if you cannot read all of them?
Every plan reads one site to a set depth — 20 pages on Free, 50 on Pro, 100 on Business, unlimited on Custom — so on a larger site the order matters more than the number. Nothing is fetched until every page we know about has been ranked. The homepage is never in question; after it come the pages that describe what you sell, then your FAQ pages, then pricing, then who you are and how to reach you. Blog indexes, legal notices and cart or checkout pages sit at the bottom. That is the order an assistant needs, which is not the order a visitor browses: a model answering “best X in Y” is looking for what you do and where, and a cart page has never been the answer to anything. Structure feeds the same ranking — how deep a URL sits, how many of your own pages link to it, whether it is in your navigation, and what your own sitemap declares and dates. Nothing is excluded. A page the ordering puts last is not a page we think is bad, and if the budget reaches it, it is read; everything we did not reach is listed for you rather than quietly treated as though it does not exist. One limitation worth stating: the page-type half of that ranking recognises English, Romanian and Russian address slugs. On a site in another language it contributes nothing at all rather than guessing, and the ordering falls back on structure alone.
What is in the visibility score?
The share of a fixed set of probe queries where an assistant named your business, measured per engine and re-run on a schedule. The full definition — which engines, how the questions are chosen, what counts as a mention versus a citation, the formula, the grade bands and the limitations — is written out in the methodology.
My visibility dropped. How do I find out why?
The dashboard answers this before it offers you anything to fix, because most falls are not falls. Each reading records how it was taken, and if the questions, the models, the method or the set of assistants that answered differ from last time, the two numbers cannot be subtracted — so the first thing you are told is whether there is a change to explain at all. A reading where no assistant answered has no score rather than a zero, which is a different sentence and matters more than it sounds: a zero says they were asked and did not know you. Past that you get an ordered list, and the order is the useful part — the integration line missing from your site comes before your page headings, because nothing downstream works until the first one is fixed. It also shows which AI crawlers have actually fetched the endpoints we publish for you and when, so "Perplexity has never once come" stops being a guess. Two things we cannot tell you, stated plainly: we have no access to your Google Search Console, so we cannot say when Google last crawled you or whether you are indexed — that would need your own Google account and we have not built it — and we cannot detect whether you have a Google Business Profile, which is why we ask you instead of checking.
Can I see the exact questions you asked, and what came back for each one?
Yes, and they are the first thing on the AI visibility tab rather than something to go looking for. Each question is listed as we sent it, with which assistants answered that particular question, which of them named you, and which linked to your site. Three things about those counts are worth knowing before you read them. A count is out of the assistants that answered that question, not out of all five — one we hold no key for, or whose request failed, is shown separately and never counted as an assistant that did not know you. A question that already contains your name or your domain is marked as unreadable, because an assistant repeating a name we handed it looks identical to one that knew, and we would rather report a smaller number we can defend. And a question nobody named you on is not a verdict on your business: we ask every assistant with web search switched on, the way people use them, so what comes back reflects what is findable about you now rather than what a model memorised — and a business that is real but thinly covered on the web can be a genuinely good answer and still be absent from all of them. What the list is for is narrower and more useful than a score — it tells you which question dragged the number down, and you know your own market well enough to judge whether that question was the right one to be asked.
Why does every question have my city or country in it?
Because without it we were measuring the wrong country. We scanned an insurance broker in Moldova and it came back zero on all four assistants — five well-formed questions, in the right language, about the right products, and every answer was about Romania, which shares the language and the product names and owns the question when nobody says where. So the place we read on your homepage now goes into every question we generate, at the level the thing is actually bought at: a restaurant by city, a service sold across one country by country. Two exceptions. A question that already names your business does not need one. And a business with no territory at all — software, an online course, anything bought identically from anywhere — gets none, because there a place makes the reading worse. If your page says nothing about where you operate, the questions stay unplaced and the run says so, rather than us inventing a market for you.
Do you put anything visible on my website?
One thing, and this is it in full: a strip at the very bottom of the page reading "AI visibility and AI communication channel powered by tunnelpowered.com", with our mark under it. It appears on every page our script runs on and on the pages we pre-render for crawlers. The specifics matter more than usual, because a footer you did not design still sits on your domain. It is the last element in the page, after everything of yours, so it cannot land inside your layout and be stretched or clipped. Every style on it is written inline and begins by resetting itself, so neither your stylesheet nor ours can reach across into the other. It paints its own background rather than inheriting yours, so it looks the same whatever colour your page is. And it is never fixed or sticky, so it cannot end up covering your own footer. There are two links in it, both to the same address: the domain in the sentence, and the mark below it. Neither is on the words before them, and that is deliberate: the same keyword-rich link in the footer of every customer site is the shape search engines treat as a link scheme, and the risk of that would land on your domain rather than ours. Both open in a new tab, so a visitor who clicks still has your page where they left it. If your Content-Security-Policy blocks images from other origins our mark will not render, the wordmark beside it is live text so the strip still reads, and nothing else about it changes.
The scan email shows one number and then sets some answers aside. Why?
Because the two are not the same evidence and adding them together would flatter us. The table counts one thing: answers where an assistant brought you up in reply to a question that never mentioned you. We allow one of the five questions we generate to name your business or your domain, on purpose — it tests whether an assistant can say anything accurate about you once it knows who you are — and an assistant repeating a name we just handed it looks identical to one that already knew. So any answer to a question like that is listed under the table, named for what it is, and never added to the count above. This makes the headline number smaller than a mixed count. It also makes it one you can check yourself: the five questions are printed in the same email, so put them to any assistant and count. Where it runs out is a scan whose every question named you — then nothing is left to count and the table says so rather than printing a zero, because no evidence and no visibility are different things.
An assistant named my business and the scan counted it as nothing. How?
Because a one-word name needs something else in the same answer to prove it is you, and until now we were checking against the wrong thing. A name like "Tunnel" or "Intact" belongs to a hundred things that are not you, so before an answer counts we look for a second signal in it: what we read you sell, your industry, your city or country. That part was right. What was wrong is what we handed the check. For a page that does not publish a postal address in its structured data, the only signals we passed were the Schema.org TYPE names off the page — strings like "InsuranceAgency" and "FAQPage". Nobody writes those in an answer, so they could never match; but because we had passed something, the check believed it had a way to tell and returned a firm no. We found it on an insurance broker that leads its own market: one assistant named them in two of five answers and we reported zero on all five. The perverse part is that it only bit sites with good markup — a page with no structured data at all was scored more honestly. The check now uses what we read you sell and where, which are words an answer can actually contain, and every count records which term confirmed it, so a term that waves everything through is visible rather than hidden. Where there is genuinely nothing to confirm against, the answer is reported as named-but-unconfirmed and kept out of the headline number, which is the same rule we already apply to a Wikidata item that only shares your name.
Do you check Wikipedia and Wikidata?
Yes, both, on every presence audit, and neither costs you anything. We search Wikidata for your business name — with and without the company form, so "SRL" on the end does not hide the match — and then confirm it by the website the item lists rather than by the name, because a one-word brand shares its name with things that are not you. If an item carries your domain we report it as yours and follow it to the Wikipedia page. If an item merely shares your name, we say that we found one and cannot confirm it is you, which is the honest answer and also tells you the one edit that would fix it: adding your website to the item. What we will not do is tell you to go and create one. Wikidata has a notability rule, self-created items for small businesses are routinely deleted, and most businesses we work with are not in either place — that is normal, it explains part of a low score, and it is not a task we are going to bill you for attempting.
Is llms.txt worth anything, if nobody has agreed to read it?
We generate one for every site and we will keep doing that. It no longer leads anything we say, and that demotion is deliberate. llms.txt is a proposed convention rather than an adopted standard: no assistant vendor has publicly committed to fetching it, and we hold no measurement of our own showing that serving one changes whether you get named. Anyone selling it as the thing that puts you into AI answers is ahead of the evidence, and for a while the top of this site was doing that. What we can show working today is duller and older — pages a crawler can read without executing JavaScript, JSON-LD on your own domain in the vocabulary those crawlers already parse, and your name, address and phone agreeing wherever they appear. Those three now lead instead. The one place llms.txt still counts is our own visibility score, where serving it from your own domain root is worth three points out of a hundred: a bonus, never a gate, and skipping it costs you nothing. It stays because it is cheap to ship and harmless to have, not because it is the argument.
Find out what AI says about your business
We ask four assistants what they say about your business, and send you the answers they gave — including the parts that are wrong. Free, and yours to keep.
Get your free AI visibility scan
By submitting, you agree to our Terms of Service and Privacy Policy