What does AI actually say about you?
Two answers, side by side: what a named model says about your brand with no context, and what it says after reading only the text an AI crawler can actually get from your page.
Two questions.
One page of evidence.
The shape of a result, not a measurement of anyone. The one real number in it is the readability floor: below 120 words of crawler-extracted text, the check says the page is too thin to ground an answer instead of pretending otherwise.
Example output. Your check reports your own two answers, at their own link.
How to read it honestly
- The memory question always runs. The grounded question runs only when your page clears the 120-word bar, because grounding an answer in text that thin would be theatre, and the verdict says so instead of pretending.
- An open-weight model answers, and the result says so. This free check is not ChatGPT, Claude, Perplexity or any other assistant product; we do not query those here, and the result never claims to speak for them. Those engines are in paid prompt tracking, where Lantad calls each vendor's own API rather than driving their chat app.
- A finished check for the same URL younger than 24 hours is reused rather than asked again, so re-running inside that window shows the same two answers.
- We read the raw HTML once, identifying honestly as LantadBot, and never run JavaScript. A page that builds its text in the browser reads as thin here even though a visitor sees a full page, and that gap is the finding rather than a fault in the check.
The four outcomes
Four states, and the check tells you which one you are in rather than leaving you to interpret two paragraphs of model output. The verdict combines two measurements: how much reliable prior knowledge the model claims about you, and whether your page cleared the 120-word readability bar. A model with only vague, hedged memory of you gets its own wording between these four.
Invisible on both fronts
This model has no prior knowledge of your brand, and your page text is too thin for AI to describe you from a crawl. Invisible on both fronts.
Readable, not yet remembered
This model has no prior knowledge of your brand. Your page is readable, so AI search can still describe and cite you when it fetches the page: that is where to compete.
Remembered, but unreadable
This model recognizes your brand from training data, but your page text is too thin to ground a current, accurate answer. Stale knowledge plus an unreadable page drifts into wrong answers.
Healthy
This model already recognizes your brand, and your page backs it up with readable text. Keep both fresh.
Fetch. Extract.
Ask. Compare.
Four steps, all of them on the record. Nothing in the result is a guess about what a model might say: it is what the model did say, on the run you can open at its own link.
-
Fetch your page as a crawler
One request, identifying honestly as LantadBot, with no browser and no JavaScript. We extract the text an AI crawler can read from the raw HTML.
-
Ask the model from memory
The model answers with no web access and no hints, and is required to say plainly when it knows nothing rather than guess.
-
Hand it your page text
The first 6,000 characters of the extraction go back to the model with one instruction: describe this brand from this text alone.
-
Compose the verdict
Prior knowledge and the 120-word readability bar combine into one banner, plus the next step that matches it.
A result page at its own link
Top to bottom, the result shows four things and states the limits of each.
A verdict banner
One or two sentences stating what was measured: whether the model has prior knowledge of your brand, and whether your page text is thick enough to ground an answer.
The answers, side by side
Before it reads you, and after. Each call walks the open-weight fallback chain on its own, and the result names the engine that actually answered.
What the crawler got
The extracted word count against the 120-word bar, whether a title and meta description were found, and the HTTP status of the fetch.
A next step matched to the verdict
A readable page the model has never heard of points at AI search. A page too thin to describe points at fixing readability first.
Three people this answers a question for
One URL, one minute, and no account. The check is the cheapest way to find out whether the problem is your page or your profile.
Founders
You want to know whether an assistant can describe your company at all before you spend anything on making it describe you better.
Marketers
You have been asked what ChatGPT says about the brand. This is an answer with a link, not an anecdote from one person's chat window.
Writers and SEOs
You are rewriting a page and want to see whether the new copy grounds a clear answer when a model has nothing else to go on.
Honest scope: this check queries one open-weight model live and names whichever model answered. It measures memory and page readability, not the full answer of any specific assistant product. The methodology explains how the scanner side works.
The Lantad brand check asks an open-weight model two questions about your site. The first is answered from memory alone, with no web access. The second is answered only from your page text, extracted from raw HTML the way an AI crawler reads it. You get both answers, the engine that produced each, and the word count your page delivered.
The model is required to say plainly when it knows nothing rather than guess, so a blank first answer is a result, not a failure.
Common questions
Which AI answers, and is it ChatGPT?
It is not ChatGPT. This free check runs on Cloudflare Workers AI, asking a chain of open-weight models and keeping the first answer. Open-weight models are published by their makers for anyone to run, so no consumer assistant is involved and the result never claims to speak for one. Paid prompt tracking is where the assistants people actually use come in: there Lantad calls the vendors' own APIs for ChatGPT, Claude, Gemini, Perplexity, Grok and DeepSeek, which is the same model the vendor serves, reached programmatically rather than by driving the chat app.
The model says it has no reliable knowledge of my brand. Did the check fail?
No, it worked. We require the model to say plainly when it knows nothing rather than guess, and for most businesses outside the household names that is the honest answer. The half you can act on is the second one: AI search fetches pages and answers from what it reads, so a page that grounds a clear answer matters more than being in training data.
What makes a page readable in this check?
We extract the text an AI crawler gets from your raw page HTML and hand the model up to the first 6,000 characters. If the whole page yields fewer than 120 words of readable text, roughly two short paragraphs, the check treats it as too thin to ground an answer and the verdict says so instead of pretending.
Why did my second check finish instantly?
A finished check for the same URL younger than 24 hours is reused instead of asking the model again. That keeps repeat lookups fast and free; to see the effect of a change you shipped, run it again after the window has passed.
How is this different from the full scan?
This check reads the raw HTML once and asks one model about it. The full scan also renders your page in a real browser, diffs what your visitors see against what crawlers read, checks access for all 15 AI crawlers in our registry, and grades the page with ranked fixes. It is free too.
Ask it about your own brand.
One URL, no account, and a result page you can send to somebody else. If the answer is thin, the full scan tells you which part of the page went missing.