The AI doesn’t lie. It writes what
you actually felt.

Every Revue review draft is generated by AI — specifically Gemini 2.5 Flash, a large language model from Google, with Anthropic Claude Haiku as an automatic fallback if Gemini is unavailable. We use AI for one reason: to make writing a review take 30 seconds instead of 3 minutes.

Stars don’t rank you. Words do. AI lowers the effort of writing words.

What the AI does

• Reads the rating you tapped (1–5 stars).
• Reads the highlight chips you selected (e.g. “Friendly staff,” “Cozy atmosphere”).
• Reads the dishes you say you ordered, if you told us.
• Reads any free-text personal note you added.
• Writes a first draft in a natural, varied voice — never identical between two diners, even with the same inputs.

What the AI does NOT do

• Make up dishes, ingredients, or details you did not mention. The model is constrained by a strict banned-vocabulary system and a facts-ledger discipline; it must stay within what you and the restaurant’s public Google profile actually say.
• Pretend to be a different person. Drafts are first-person, neutral, and meant for you to make your own.
• Submit the review on your behalf. You copy the draft and paste it into Google. Your finger taps Submit on Google’s page. We never auto-post.
• See or store your Google account. Google authenticates you, not us. We do not receive a Google identity token.

You are always in control

The draft appears in an editable text box. You can rewrite any word, delete the whole thing and write your own, or tap “Try another version” for a different first draft. You always edit before posting. Your finger taps the submit button on Google. We do not see the final review unless you choose to share it with us via the private feedback channel.

1–3 star reviews: no AI at all

For 1, 2, or 3 star ratings, we do not generate any AI draft and we do not offer AI “polish.” The text box starts empty. You write your own words. Period.

Why: critical feedback in someone else’s words is a worse review and a worse signal to the restaurant. It is also the kind of fabrication risk we refuse to take. If you had a bad experience, your voice is the only voice that should describe it.

How we test the AI

We run an automated diversity and hallucination check on every meaningful prompt change. The most recent measurements:

• Self-BLEU diversity: 0.002 (target ≤ 0.35; we are ~175× better than target — meaning two diners with identical inputs receive two clearly different drafts).
• Hallucination rate: 0% on the canonical test set (no invented dishes, no invented dates, no invented ingredients).
• Banned-vocabulary leakage: 0% on day-of-week, clock-time, and meal-segment terms (we strip these from the input and forbid them in the output to remove AI tells and time-of-visit hallucinations).

If you ever see a draft that mentions something that did not happen — an ingredient that is not in your dish, a restaurant style that does not match, a Monday-night visit when you came on a Saturday afternoon — please email support@revue.today. We treat that as a bug, not a feature.

Why we disclose this

The FTC’s 2024 Endorsement Guides update made clear that AI-generated review text should be transparent to consumers. We agree, and we go further than the rule requires: every diner sees an inline “AI-assisted draft” note before they edit, with a link back to this page.

See our FTC Review Policy for our broader anti-gating and anti-incentive commitments.