Know what readers will believe before your ad goes live.
Readback checks marketing drafts against your brand rules and approved facts, then shows the claims readers take away, including the ones the copy only implies. Every finding cites its source and a person makes the final call.
No sign-up. You'll land in the Velura workspace with example drafts already checked.
Vitamin C serum. . 100% natural, . Shop now
5 of 6 readers think results are guaranteed within a week
The clinical study measured radiance at week 4. Suggested fix: “71% saw more radiant skin in 4 weeks*”.
How a draft moves through Readback
- 1
Submit a draft
Paste the copy, add the image, and write one line on what readers should take away.
- 2
Rules run instantly
Banned terms, missing disclaimers, embargoed names, image contrast and alt text, checked by plain code.
- 3
Six readers react
Simulated readers say what they believe after five seconds with the post, including what it only implies.
- 4
Beliefs meet the facts
Each belief is checked against the brand’s approved facts (the Fact Ledger): supported, contradicted, or unsupported.
- 5
A person decides
The creator accepts or disputes each finding; a reviewer approves. Every decision is logged.
Two kinds of check, labelled on every finding
Anything that can be decided exactly is checked by code. Only questions of meaning go to the AI, and an AI finding has to quote the words or point to the part of the image that caused it.
Rule checks [RULE]
Exact, instant, no AI involved.
- Banned and preferred terms, matched to the clause
- Required disclaimers and embargoed product names
- Text-on-image contrast measured from pixels (WCAG 1.4.3)
- Alt text (WCAG 1.1.1) and off-palette colours
AI reader panel [AI]
Six simulated readers, grounded in the draft and the approved facts.
- Implied claims: endorsements, guarantees, “natural”
- Tone against the brand’s voice
- Beliefs that contradict or lack an approved fact
- Translations that make a claim stronger
Beyond the first check
Translations can't quietly promise more
Link a Hindi, Hinglish, Tamil or Marathi version to its English master. Every check compares what readers believe from both, and anything the translation adds or makes stronger, like पक्का (“guaranteed”), becomes a finding a reviewer must clear.
Live posts are re-checked when facts change
When an approved fact changes or expires, every post that cited it joins the Recall queue. Readback opens each live page, looks for the stale wording, and closes the recall itself once it is gone. Expiry alerts arrive before a fact lapses.
Measured against the regulator
ASCI-Bench runs the engine on 100 real Advertising Standards Council of India rulings, three times each, against a single-prompt baseline, with and without the evidence ASCI accepted. Every case and its result is open to inspect.
Try it from any seat
Four demo accounts, one click each. A short guided tour starts automatically when you arrive; reopen it any time from the Tour button in the top bar.
Sign in with an email and password instead
What leaves your machine, and where it goes
Stays on the Readback server
Rule checks: banned and embargoed terms, missing disclaimers, colour contrast (WCAG 1.4.3), alt text (WCAG 1.1.1), palette drift. Plain code, no model involved.
Sent to Google Gemini (gemini-3.8-flash)
Draft text, after emails, Indian phone numbers, PAN, Aadhaar, card numbers and embargoed product names are replaced with tokens such as ⟦EMAIL_1⟧, and the image as uploaded: text, faces or logos inside an image are not masked, so do not upload images that show personal data. Gemini plays six readers, checks beliefs against your Fact Ledger, and proposes span-level rewrites. Text embeddings use gemini-embedding-001 (768 dims), or a newer Gemini embedding model when its daily limit is reached.
Stored in Supabase
Drafts, versions, findings, comments, the Fact Ledger, and an insert-only audit log whose rows are chained by SHA-256. Images sit in a private bucket and are shown through short-lived signed links.
Fetched from the web
Recall checks open the published URL recorded on a post, read its visible text on the server, and keep only whether the stale wording was found (and the matching phrase) in the audit log. Private and local network addresses are never fetched.
Simulated readers are a model’s guess at how people read, not a survey. Readback flags what to look at; a person still approves. With no API key, or with READBACK_MOCK_AI=1, readers are deterministic stand-ins and every review screen says so.