How to review AI content before it’s published

Start with the work that will appear

An image can have the right colours and the wrong product name. An email can follow the voice guide but lose its font when rendered. Reviewing the prompt or the editable source alone will miss those problems.

Start with the finished file, page or message in its intended placement. Keep the task beside it: who it is for, what it should say and any required assets. Then open the brand guidance the work was supposed to follow.

Those references answer different questions. The task might call for a new product announcement; the brand guide establishes how that announcement should look and sound. A new campaign idea does not automatically change the brand’s standing rules.

This article explains the brand check and what to do with its findings. Fit it into the publishing process you already use. An in-house email, an agency’s client delivery and content produced inside a customer-facing product can have different authorization and approval requirements.

Compare specific elements with specific rules

“Make it feel more like us” gives the next editor almost nothing to act on. Name the element, the rule it differs from and the change that would address the difference.

Some checks are straightforward comparisons, such as exact spelling. Others need interpretation: a photograph may follow the palette and still use a composition the brand avoids.

If reviewers cannot point to a rule or a useful reference, the guidance may need work. Writing a brand voice guide explains how to turn broad adjectives into instructions. For visual work, keep the original logo, fonts and references accessible alongside the rules.

Request a brand Review

Rainbrand maintains brand guidance and assets that agents can use while making and checking work. Its Review compares work against that brand and returns findings for the requested checks, with reasons and evidence pictures where available.

For a practical request, prepare a clear visual capture of the output. Render a page, email or slide and submit its screenshot; provide a finished image directly. Include the exact wording expected in the image when asking for a copy check. Use the file upload guide if the connected agent cannot access your file.

The current request takes one primary image and lets you select colour, typography, logo and copy checks. Review a multi-slide piece through separate requests for the slides you need checked. The API reference documents input formats and fields for code calling the same capability.

Review the attached Atlas Pro launch image against Acme. Check colour, typography, logo and copy. The headline should read “Atlas Pro is here.” and the call to action “Explore Atlas Pro”. Wait for the Review, then show me each verdict, its reason and any evidence pictures. Identify anything the supplied image or references did not let you check.

A screenshot gives the reviewer the visible result. It does not establish that links work, that a page can be used with a keyboard or that an email will render correctly in every client. Check those in the environment where they matter.

Read the finding before acting on the verdict

A Review returns a verdict for each requested check. Read it with its reason and evidence, rather than treating the collection as a single publishing score.

An uncertain typeface might be easier to resolve by inspecting the font used in the source file than by generating another image. A logo check marked not applicable deserves attention if the task required a logo. The reason determines the useful action.

An operation that failed to complete has not produced a brand verdict. Resolve the error before treating the work as reviewed.

A pass applies to the submitted input, references and requested checks.

Keep the evidence’s scope visible

In our Eager example, Claude made a blackcurrant-flavour campaign concept using Rainbrand’s brand guidance and files. Eager is a juice brand.

The saved final carousel Review is specifically for slide 5. It records passes for colour, typography, logo and copy. This is one of its saved copy crops:

These findings cover the submitted slide, not the whole carousel. The colour finding also records that product-packaging colours were excluded as requested.

The guide to keeping AI content on brand shows the demonstration’s separate landing-page correction. Here, the useful habit is to keep each result attached to the exact output and instructions it examined.

Revise the work and review the result

Review supplies findings. Your agent or editor makes corrections with its own tools, then requests a new check of the revised output. Change the editable source where possible so later exports retain the fix.

A bounded request keeps an automated revision loop useful:

Use the Review findings to revise the Atlas Pro image with your editing tools, then request another Review. Make up to two editing rounds. Stop when the applicable checks pass, an operation fails or a finding needs a decision you cannot resolve from the task and brand guidance. Show the final file, the changes and anything still open.

The two-round limit is an example, not a product requirement. Choose a limit that fits the work and its cost. Repeated attempts are unlikely to resolve a missing font file or an unanswered product question.

Inspect the revised composition as well as the named correction. Changing a line break can crowd the logo; replacing a photograph can make the text hard to read. A result saved for the earlier file does not describe the changed one.

For recurring work, request a Brief before creation. A Brief supplies the brand guidance and assets selected for the task, reducing the need to reconstruct them during review. It records the brand version it used. The current Review request names a brand rather than accepting that saved version as an input, so notice if the guidance changed between creation and review.

Keep a short record beside the work. A document, issue or existing publishing tool can hold it:

Finished file or rendering:
Intended placement and task:
Brand reference or Brief and its version:
Checks performed:
Findings and evidence:
Corrections and revised file:
Unresolved decisions:
Other publication checks, where relevant:

If your process requires client sign-off or publication approval, keep that decision with the same version too. A brand verdict does not grant an agent permission to publish.

When a finding recurs, check whether the cause is outdated guidance, missing files or failure to follow an existing rule. Correct the cause you actually found. In Rainbrand, prepare a brand update, review it and apply it when the brand context needs changing; request a new Brief afterward. Teams can use this through their agents, and products can embed the review loop in their own workflow.

Check what the brand review cannot establish

Finished work still needs accurate facts and usable presentation. Keep these checks alongside the brand findings, at the depth your task requires.

Facts and claims. Ask the drafting agent for an evidence list, then open the sources:

List every factual claim in this finished draft, including features, numbers, prices, dates, comparisons and quotations. For each, identify the source and quote the supporting passage from the supplied material. Mark unsupported claims explicitly. Do not rewrite the draft.

This follows Anthropic’s guidance on making claims auditable. Verify the quoted passage and what it supports. A correctly spelled claim can still be false. For advertising, the FTC’s business guidance also emphasizes that advertisers need evidence for express and implied claims. Regulated or unfamiliar claims may need specialist review appropriate to the market.

Rights and required disclosures. Check the intended use of photographs, fonts, logos and quotations against the permissions you hold. Resolve any applicable disclosure requirement for the channel and content. Visual similarity to the brand cannot establish either.

Accessibility. Check the actual placement, including these common issues:

• Give informative images text alternatives that serve their purpose. Use the W3C’s alt decision tree to distinguish informative, functional and decorative images.
• Check text contrast, generally at least 4.5:1 for normal text and 3:1 for large text.
• Use real text instead of images of text where the technology can achieve the presentation, subject to the criterion’s exceptions.
• For video carrying prerecorded audio, check captions, including meaningful sounds as well as speech.

You can start without a dedicated proofing platform. WebAIM’s Contrast Checker measures colour pairs, WAVE helps identify web accessibility issues, and Vale can apply written style rules to prose. Their findings answer different questions; none supplies missing evidence for a product promise.

For the next finished piece, gather the final output, the task and its brand reference. Make the comparisons, record the findings and give every unresolved item a concrete next action. Carry any required publication decision through the process your team or product already uses.