TL;DR
- Chat in Hi, Moose can now use tools: it searches an index of your own site's pages and pulls real data instead of answering from general knowledge, on both managed and bring-your-own-key setups.
- Search Console questions are now agentic, so chat plans and runs its own sequence of queries instead of firing one canned report.
- The routing layer that decides what a chat message actually wants was rewritten and tested against a 54-case battery of real phrasings; the local model went 54 for 54.
- New drafts now pick up internal links to related pages on your site; the first version's similarity bar was set so high it never suggested a single link, and that is fixed.
- The library got search, lazy loading, and newest-first ordering, and the marketing site got a free AI search tracking page.
Chat learned to fetch
Until this week, if you asked chat about your own site, it answered from whatever it happened to know. Reasonable-sounding, occasionally right, the way a coworker answers a question about a document they read last month.
Now chat has tools. Ask about your site and it can search an index of your actual pages, find the ones that matter, and answer from what is really there. This works whether you are on a managed plan or bringing your own key. Fetching the real thing beats describing it from memory. I have held this position for years.
Search Console stopped being a vending machine
Search Console questions used to work like a vending machine: your question got mapped to one query, the query ran, you got what fell out. If the answer needed a follow-up, that was your job.
Now it is agentic. Chat looks at your question, decides which queries would answer it, runs them, looks at what came back, and runs more if it needs to. "Which pages lost clicks this month" is not one query, and now it does not have to pretend to be.
The router took a test
Under all of this sits a rewritten routing layer, the part that reads a chat message and decides what it wants: a question, a command, a Search Console lookup, an audit request. Routing mistakes are the ones that make an assistant feel dense, so we built a battery of 54 real phrasings, the kind people type when they are not being careful, and ran the local model against it.
54 for 54. We were suspicious too. We checked. The plan is to keep making the test harder rather than celebrating, but a clean sheet on the first full run was a nice moment in an otherwise ordinary Saturday.
The link feature that never linked
Here is the honest one. New drafts are supposed to get internal links: when Hi, Moose writes a draft, it checks the rest of your site for related pages and links to them where it fits. We shipped the first version, and it suggested nothing. Not bad links. Nothing.
The cause was a similarity threshold. The feature compared the draft against your pages and only offered links scoring above 0.5. On the small embedding model doing the comparing, a genuinely good match scores around 0.33. The bar was not high, it was unreachable, and the feature sat there looking diligent while silently rejecting every candidate it ever saw.
We measured real scores against a live site index and rebuilt the gate around them: a low floor to cut noise, then a band relative to the best match, so quality is judged against what your site actually offers. Now drafts come out with links to the pages that earn them. The lesson, which we will forget and relearn, is that a threshold you never validated is not a standard, it is a wall.
Smaller things
The library can now search. It also loads fifty items at a time as you scroll and orders everything newest first, which matters once a library is big enough to be useful. And the note you see when making an artifact private now says plainly what private does, instead of assuming you would guess right.
Over on the marketing site there is a new free AI search tracking page, for checking how you show up in AI search without purchasing anything.
Good week. The ball is under the desk if anyone needs me.
FAQ
What tools can chat use now?
Chat can search an index of your own site's pages and query your Google Search Console data, so answers about your site come from your site instead of general knowledge. More tools will hang off the same mechanism over time.
Does tool use work if I bring my own key?
Yes. Tool use in chat works for both managed plans and bring-your-own-key setups.
What does the internal-linking feature do?
When Hi, Moose generates a new draft, it checks the rest of your site for genuinely related pages and adds internal links where they fit. The first version's similarity bar was set so high that no real match could clear it, which is now fixed.
What changed in the library?
The library now has search, loads items fifty at a time as you scroll, and orders everything newest first. The note shown when you make an artifact private is also clearer about what private means.
The one who's under the desk for all of it.
