What it does
Review Themes takes your shop's recent customer reviews and groups the recurring feedback into themes — a handful of things customers keep praising, and a handful of things they keep complaining about — instead of leaving you to read reviews one at a time. Each theme comes with a short label, a mention count, the average star rating tied to that theme, and one representative quote pulled straight from a review. The idea is to help you fix root causes (a recurring complaint about a specific service, say) rather than just replying to reviews individually as they come in.
This is a read-only tool: it analyzes reviews you already have and doesn't post replies, edit reviews, or write anything back to your account.
Where to find it
Review Themes lives at `/app/ai/review-themes`. The intended path is through the AI Tools hub (`/app/ai`), where it's one of the cards under the "Owner intelligence" section — open the hub and click in.
There's no dedicated sidebar entry for it, and no direct sidebar entry for the AI Tools hub either — the hub isn't in the sidebar rail, so you won't find a labeled "AI Tools" menu item to click through. Reach it by opening Growth & Marketing from the sidebar (under "Run the shop"), then clicking its AI Assistant tab, which loads the hub inline (that tab also has an "Open standalone" link straight to `/app/ai`), or by navigating straight to the URL above.
The command palette (⌘⇧K / Ctrl+Shift+K — not plain ⌘K, which is reserved for the AI Help bar in the top toolbar) can jump you to some AI tools by name, but its current registry doesn't include Review Themes yet, so searching for it there won't surface a result. Use the Growth & Marketing path or the direct URL instead.
Live vs. stub mode
At the top of the page you'll see a badge: green "● Live" or amber "● Stub mode." Stub mode means no ANTHROPIC_API_KEY is configured in the environment, and you'll also see an amber notice explaining that the UI still works for testing but isn't producing real model output. Once a key is set, the badge flips to Live and results come from the actual model. Every result you generate also states which mode produced it — after a run, a small line under the Run button reads "N reviews analyzed · last X days · live model" or "· stub output" — so you can tell at a glance whether what you're looking at is real analysis or placeholder data, even if you're not sure what the top badge currently says.
Running an analysis
The "Analyze reviews" card has two number fields: Lookback (days), defaulting to 90 (allowed range 1–365), and Max reviews, defaulting to 100 (allowed range 1–200). Nothing runs automatically when the page loads or when you change these fields — you have to click Run. While a run is in progress, the button shows a spinner and reads "Analyzing…" and is disabled to prevent double-submits.
One thing worth knowing: the results panel and the "N reviews analyzed · last X days" caption are always tied to the exact Lookback/Max reviews values of your last completed run — they don't linger on screen showing mismatched old numbers. If you change either field to a combination you haven't run yet, the panel clears back to "Run to generate review themes." (and the caption disappears) until you click Run again.
Reading the results
Before you run anything, the Themes card just says "Run to generate review themes." Once a result comes back, the card header updates to show a count like "3 praise · 2 complaint." If the shop simply has no reviews in the selected window (or nothing worth clustering), you won't get an error — you'll get a plain, neutral summary sentence from the tool itself explaining that, centered in the card.
When there are themes, you get a short owner-facing takeaway sentence at the top, then two labeled groups: "Customers love" (praise themes, marked with a thumbs-up) and "Fix at the root" (complaint themes, marked with a thumbs-down). Each theme row shows its label, the average star rating for reviews tied to that theme, how many mentions it had, and one italicized example quote pulled from an actual review. Use the mention count and rating together — a complaint theme with a high mention count and a low average rating is the one to prioritize; a low-mention theme is more likely noise from a single reviewer.