AI features
Four places Cascade uses a model, all of them optional. Summaries and screening are available on the Growth and Enterprise plans, instant answers are an add-on we switch on for you, and the assistant is available to everyone.
None of them can change your data on their own. Summaries and answers are shown and can be turned off, screening only flags, and the assistant asks you to approve every change before it makes it.
Review summaries
Once a product has enough published reviews, the reviews widget shows a short summary of what reviewers say above the list, headed "What reviewers say" and marked as generated. Under it are up to five themes, each with roughly how many reviews mention it and whether reviewers raise it approvingly, as a complaint, or both.
It is generated from published reviews only, in whichever language most of them are written in, and it tells shoppers it was not written by you or by a reviewer. It is kept out of the structured data on the page, because it is not a review anyone wrote, and it never sees the original text of a review you have redacted.
When it refreshes
- Never below five published reviews.
- At most once a day per product.
- After roughly 10 more reviews, or 20 percent growth, whichever comes first.
That policy is about cost per product rather than account size, so it is the same on every plan.
Controlling it
On a product in the dashboard you can regenerate the summary, hide it from your storefront, or go back to the previous version after a regeneration you did not like. To turn the feature off for every product, use Summarize what reviewers say under Settings → Reviews: turning it off hides every summary at once.
Instant answers
When a shopper asks a question in the reviews widget, they get a short answer generated from that product's published reviews straight away, with the reviews it drew on shown underneath. If the reviews do not address the question, the shopper is told so rather than given a guess.
This is the "while you wait" half of Questions and answers: the same ask also goes to your team, and the answer you write is the one that gets published. Instant answers work on their own if you would rather not run a public Q&A, and a public Q&A works on its own if you would rather not run instant answers.
Private to the asker
The generated answer is shown only to the shopper who asked. It is never published on your storefront, never added to the page's structured data, and other shoppers asking the same question each see it privately. Only the answer your team writes is published. Answers draw on published reviews only, and a review you unpublish disappears from answers that cited it.
What is stored, and for how long
Each distinct question is stored once per product, in a normalized form, together with its generated answer. That record powers your Instant answers page and keeps a repeated question from being generated twice. No record ties a question to a shopper. The only per-shopper data the feature keeps is an anonymous rate-limiting token hash, and that is deleted after 14 days.
The Instant answers page
Every generated answer appears under Reviews → Instant answers in the dashboard: the question, the product, how many times it has been asked, and the answer shoppers see. You can hide any answer; shoppers who ask that question are then told the reviews do not say. Showing it again is one click. The questions themselves, and the answers your team writes, are under Reviews → Questions.
Availability and allowance
Instant answers are an add-on: ask us to switch them on for your organization, and until then the switch on the settings page is disabled. Once they are on, your own switch is Show an instant answer while they wait under Settings → Reviews. The generated answer never replaces the one your team writes. The box only appears on products with enough published reviews (the same floor as summaries, five by default), and generation is capped at 500 newly generated answers per month; repeated questions are served from the stored answer and do not count. When the cap is reached, the ask box simply disappears until the month resets. Shoppers never see an error for it.
Review screening
An optional classifier that reads each arriving review and says whether a person should look at it. It catches what a word list cannot: off-topic text, abuse, a review that is really about a different product.
It only flags. It never rejects, unpublishes, edits or deletes anything. A flagged review is held for you with the reason shown in the moderation queue, exactly as the built-in checks (links, profanity, your blocked words) already do.
The one behavior worth understanding: when screening is on, the publish decision waits for it. A review that would have auto-published is inserted unpublished and published once the classifier has read it. That is deliberate. The alternative, publishing and then retracting, means spam is live under your brand for a few seconds, which is the exact window you turned screening on to close.
If the classifier cannot run, reviews simply wait in the moderation queue for a person, which is what happens with the feature off anyway. Nothing is lost and nothing wrong reaches your storefront.
Reviews that look AI-written
Screening also runs longer reviews past a specialist AI-writing detector, and flags the ones it is confident about with "AI screening flagged this as possibly AI-generated", followed by the score the text was given.
Treat it as a prompt to read the review, not a verdict:
- It only flags. Nothing is rejected, and shoppers are never told anything about it.
- Short reviews are skipped, because no detector is reliable on two lines of text.
- Only high confidence is flagged, because these detectors report false positives more often on people writing English as a second language. A flag is a reason to look, and a well-written review from a real customer is a perfectly ordinary thing to find.
It comes with Screen reviews with AI and has no separate switch.
Turn screening on with Screen reviews with AI under Settings → Reviews.
The dashboard assistant
A chat panel that answers questions about your own store and makes changes you approve. It is on its own page under Assistant, with your past chats beside it, and behind a launcher in the bottom right of every other screen.
Ask it the things that would otherwise be a filter and a spreadsheet:
- Which products got 1-star reviews this week?
- How many loyalty members are close to the next tier?
- What is my average rating doing month over month?
Charts, in the chat and in your library
When the answer is a count, an average or a trend, the assistant draws it. The chart appears in the conversation above the tool row, in the same colors and with the same table view as the charts on Analytics, so a number you asked for in passing looks the same as one you built by hand.
Ask it to keep one and it will offer to save it: a name, the resource, the chart type and the query behind it. It checks the chart type against the shape of the query first, so it does not ask you to approve a chart that cannot be drawn. Saved charts land on the dashboard you pick, and the assistant can arrange them into a dashboard for you as well. It can also read what you already have, so you can ask what is in your library before adding to it.
Both of those are changes, so both wait for your approval like any other. Once a chart is saved, the assistant draws it under the approval card so you can see what you just kept.
What it can see
Exactly what you can see, and nothing else. Every action it takes goes through the same API the dashboard uses, carrying your own session, so it cannot reach data you could not open yourself, and it cannot reach another organization.
Every change waits for you
The assistant can read freely. To change anything it has to ask, and it has to say in plain language what the change does before it is allowed to: you get a card with that sentence, and the request underneath it, with approve and reject. Nothing is written until you approve.
A change you approve is recorded as yours, because you approved it.
Attaching a screenshot or a PDF
Use the paperclip in the composer to attach an image or a PDF to a message: a screenshot of something that looks wrong, a supplier's spec sheet, a report you want the numbers read out of. PNG, JPEG, GIF, WebP and PDF are the formats it reads.
One message may carry up to 5 MB of attachments. A larger file is refused, with a note saying so, whether you pick it in the composer or send it another way. If you need more room on your account, ask us to raise it: it is one setting per organization, like the rest of your plan limits.
Your chats
Conversations belong to you rather than to your organization, so another member cannot read them.
Each one gets a short name once your first question has an answer, and the chat list shows that name with the question that started it underneath. Until it has been named, the list shows the question on its own.
Everything you can do to a conversation is in the ... menu at the top of the chat you are reading. Delete one and the conversation goes; anything it actually changed stays, because that is a change to your store rather than a chat message.
Telling us what it should do differently
Beside the send button is a feedback button. Use it any time. When the assistant runs into something it has no way to do, like editing your storefront theme, the button grows a Send feedback? label, because that is the moment you have something to tell us.
What you write goes to the Cascade team by email, with your name, your organization and the conversation's ID so we can ask you about it. The conversation itself is not sent.
Turning AI off entirely
The review features are switches under Settings → Reviews. With no model provider configured for your account, none of these features do anything and everything else works exactly as it does now: reviews still arrive, still screen against your word lists, and still wait in the queue.
What is different from the MCP server
The assistant is Cascade's own chat panel. The MCP server is how you point your own assistant (Claude, ChatGPT, or anything else that speaks MCP) at your store. They share a tool surface and the same rule about only reaching what you can reach, but the MCP connection is authorized separately, per client, and you choose what it can see on the consent screen.