A link is source material, not a search
Paste a link into the conversation and ANVL reads that page and answers from it. It does not go looking on the web — and the reason is the same one that decides what goes into your speaker notes.
Paste a web address into the conversation — what does it say on https://…/about-us? — and ANVL fetches that page, reads the text and answers from it. The page's title and domain appear as a chip on your message, the same way an attached document does. Ask for a summary, a deck or a post built on the page, and the proposal lists the link among what you have given. Build builds from it the way it builds from an uploaded document, and the fact check counts the page's text as a source.
That is the feature. The interesting part is what it does not do.
ANVL does not google
It would have been easy to let the conversation search the web. Plenty of tools do. We chose not to, and the reason fits in a sentence: a search hit is not your fact, but a link you give is one you stand behind.
Everything ANVL makes is meant to go out under your name. A number in a deck has to come from somewhere you would defend if a customer asked. When you paste the link to your own case study, or to the tender you are answering, you have chosen the source. When a model picks the third result for a query it wrote itself, nobody has.
So what is fetched is the one page you pointed at — not the rest of the site, and not whatever else the web has to say on the subject. What is not on that page, the conversation does not know, and it says so rather than filling the gap.
When the page cannot be read
A link fails more often than you would think. The page is gone, the address is one ANVL is not allowed to read, the link is to a PDF, or the site takes too long to answer. In each case the reply starts with Could not read the link and the reason — before anything else, so you are never handed an answer that quietly rests on nothing.
A PDF link points you to the document route instead: upload the file, and the conversation reads it from there.
The page is material, not instructions
A page you paste is something to work from. Two consequences you might not think to ask about:
- Every address ANVL fetches on your behalf passes a guard that refuses private networks. While this feature was being reviewed, the guard turned out to miss one unusual way of writing an internal address. That was closed before anything shipped.
- Text from a fetched page can never write to your brand memory unless you tick it yourself. A page cannot teach ANVL something about you behind your back.
The same rule, inside the deck
The same week gave the editor a way to write speaker notes. The Notes tab has two buttons: Write notes for this slide, and one for every slide that is still missing them.
Notes are what you say, not what the slide shows: the bridge from the previous slide, the point, the figure to stress, the example. They are written in your brand's tone, and only from what is in the deck and the order behind it. Whatever only the presenter can know is left as a gap marked fill in, not a guess. A day later the notes got three lengths, from a few keywords for someone who knows the subject to a full script for someone who does not.
That is the link rule again, one level down. ANVL works from what you gave it. Where that runs out, it tells you, instead of finding something to put there.
And it stops nudging
One smaller change from the same week, because you will notice it. In a stress test, seven of ten answers ended with an offer to carry on with something — even after a no, a typo or a question about the weather. Now the conversation answers what you asked and stops. And a text short enough to fit in a reply — a caption, a tagline, five headlines — is written straight into the reply, in your tone, instead of becoming a proposal card with questions.
If you are evaluating ANVL: this is how it treats sources. What you hand it, it reads closely and works from. What you did not hand it, it does not go and find.