Answer caching

What gets cached#

• First questions only. The opening question of a conversation is cacheable; follow-ups depend on everything said before them and always go to the model.
• Questions are normalised before lookup. Case and punctuation don’t matter, so “How much is storage?” and “how much is Storage” share one cache entry.
• Cached site-wide. The same question asked on any page of your site hits the same entry. The one exception: questions that point at the visitor’s current page (“is there a limit on this page?”, “what does this product cost?”) are answered against that page and cached per page.
• Cache hits still stream. The answer is replayed word by word: instantly available, but never a jarring wall of text.

Invalidation#

A cached answer is only served while everything it was built from is still true:

• Content changes: every cache entry is stamped with the content hash it was answered from. A re-ingest clears the site’s cache entirely; the next asks rebuild it from the new content.
• Assistant changes: entries are also stamped with a version of the assistant’s instructions, so improvements to how Assist answers roll out without stale replays. Your pinned facts are part of that version: edit a fact in Assist → Overrides and cached answers stop matching immediately.
• Time: a 7-day TTL backstop means nothing lives forever, even if your content never changes.

Need it gone now? Assist → Ingestion → Purge answer cache clears your site’s cache instantly and for free. Use it when you’ve fixed the content upstream but haven’t re-deployed, or when a cached answer needs to go.

Overrides answer first#

Curated Q&As (Assist → Overrides) are checked before the cache and before any model call: a first-turn question that strongly matches one returns your exact words, instantly. Pinned facts steer every model-composed answer and take precedence over page content. Overrides are free on every site.

Cached answers and billing#

Cache pre-warming#

Rather than waiting for visitors to fill the cache, pre-warming reads your ingested content, generates the questions visitors are likely to ask, and answers them into the cache ahead of time, so your most-asked questions are instant from the very first visitor. It lives in your site’s Ingestion tab in Control: pick a scale (up to 100, 250 or 1,000 questions), review and prune the generated list (nothing is spent until you confirm), and optionally turn on auto re-warm, which refills the cache automatically after every content change. Pre-warming is a paid feature; each pre-warmed answer meters at the normal per-question rate, and anything Assist couldn’t answer from your content lands in the question log as a content gap. See pricing.