How to review AI-generated content before publishing
In October 2026, Google’s page on content made with generative AI began saying plainly that you need to fact-check by hand and review all AI-generated content before publishing. The guidance also covers titles, meta descriptions, structured data and image alt text, since they all show up in search results.
The hard part is volume. With AI, a small team produces in one afternoon what used to take a week, and review becomes the bottleneck. That opens the door to two bad shortcuts: asking the same AI to review what it made, or reviewing by taste, with a “not feeling it” that nobody can turn into a fix.
This is the fourth piece in our series on AI agents applied to brand work. It covers how to set up a review that catches errors of fact, brand and writing, and how we do it at Headcore.
Why AI shouldn’t review its own work alone
Using a model to evaluate texts has a name in research: LLM-as-a-judge. The study that popularized the technique, by Zheng and colleagues, published in 2023, showed that a strong model agrees with human evaluators in more than 80% of comparisons. The same study also described this judge’s biases.
The first is position bias. Swapping the order of two answers often changed the verdict, and only the best model tested stayed consistent in more than 60% of cases. The second is verbosity bias: some models preferred the longer answer even when it only repeated what the other one said.
The third is a preference for what sounds familiar. Wataoka and colleagues showed in 2024 that models give higher scores to texts they find more predictable, which tends to favor their own style. A model reviewing a text it wrote itself, in the same conversation, stacks all three problems: it already knows what it meant to say and reads what it expected to read.
None of this rules out AI in review. What changes is the design: the reviewer works in another context, without seeing how the text was made, and with criteria set before reading the piece. In Claude Code, that is what subagents are for, since they run in a context window separate from the main conversation.
The three layers of review
A review that works separates three questions, because each one calls for a different kind of knowledge.
| Layer | Question | Who handles it best |
|---|---|---|
| Fact | Is what’s written true, and where did it come from? | Someone who knows the subject, with the primary source open |
| Brand | Would the brand say this, in this way? | Someone who knows the brand’s rules, with the brain or brand book at hand |
| Text | Does it read well and sound like a person wrote it? | An editor, human or agent, working from written rules |
In the fact layer, the rule is to open the original source, because a summary of it often loses the detail that changes the meaning. While checking a recent draft, we found 5.7% growth attributed to one specific mall. In the original source, the figure referred to sales across all of the group’s malls. The sentence looked right, cited the right source and was wrong.
In the brand layer, the question is whether the text respects the positioning, the voice and what the brand never says. This is where a written record of the brand is sorely missed, because without one each reviewer applies their own idea of the brand.
The text layer covers rhythm, clarity and the patterns that give away AI writing, such as a punchline closing every paragraph or a list of three adjectives. We collected these patterns in our map of AI writing patterns.
Set the yardstick before you read the piece
Reviewing by taste fails because the criterion shows up after the piece has been read. The reviewer senses that something doesn’t work and goes looking for a justification. What we did was replace the question: instead of “did I like it?”, the review asks “what does this fail against?”
Every rejection must point to a yardstick that predates the piece and applies beyond it: a communication principle with a source, a written brand rule, a platform specification. If no one can point to the yardstick, the comment is opinion and doesn’t fail the piece. It can become a suggestion, and sometimes it turns into a question for the brand to answer.
How our review panel works
At Headcore, every piece goes through mesa-criativa, one of the plugins we use in Claude Code. The panel has ten lenses, each looking at the piece from a different angle. Each lens reads alone, without seeing the others’ views, and its first assessment is frozen before discussion. Then the lenses debate the problems they found, and a log shows that nobody rewrote their initial opinion.
The panel doesn’t average scores. One fatal flaw fails the piece, even if the other scores are good. Averaged out, a piece with three comfortable scores and one serious error would scrape through; at the panel, it goes back to production with the reason in writing.
What each lens can hold a piece to follows a three-tier hierarchy:
| Tier | What it is | Weight |
|---|---|---|
| Canonical | Communication principle with a named source | Sets the score |
| Brand | The brand’s Brand Brain, audited and versioned | Sets the acceptable minimum |
| Empirical | Pattern observed on a platform, with a date | Informs only, never fails a piece |
The most useful rule applies to the brand tier: no rejection on brand grounds without citing the Brand Brain section that supports it. If the brain doesn’t cover the point, the panel logs an open question for the brand to answer. A piece is never failed over a rule that doesn’t exist.
The output is a verdict with the required rework, sent straight to whoever produces the piece: the agent for pieces, for pages or for video. The brand memory behind this review is explained in our piece on CLAUDE.md and the Brand Brain.
One piece or a batch
Not everything needs the panel. When one concept generates three hundred ad variations, reviewing them one by one doesn’t fit the deadline, and it isn’t necessary either. Variations of the same concept tend to go wrong in the same way.
So we separate what is scarce from what is repetitive. The concept, two to four per campaign, is judged piece by piece at the panel. The batch goes through automated spec checks, such as text length, format and banned terms, plus sampling. The first piece of each family becomes the reference the others follow.
What stays with a person
In our operation, AI prepares the review and a person decides. No piece goes out without human approval. A person also handles the fact-checking that calls for judgment: deciding whether a source is reliable, whether a 2025 number still holds in 2026, whether a claim about a client can be published.
When a review finds an error that could happen again, it becomes a written rule in the Brand Brain. That is how today’s review reduces the work of the next one.
Checklist to get started
- Write down, before production starts, what fails a piece.
- Keep the fact, brand and text checks separate, even if the same person does all three.
- Open the primary source for every number, and note its date.
- If you use AI to review, use a different context, without the production history, and hand over the criteria along with the piece.
- Don’t average: one serious error fails the piece.
- Review titles, meta descriptions, alt text and structured data too.
- Turn every error that could come back into a rule.
The full series
This piece is part of a five-part series on AI agents applied to brand work:
- Subagents and agent teams in Claude Code
- Skills and plugins in Claude Code
- CLAUDE.md and Brand Brain: brand memory for AI
- How to review AI-generated content (this piece)
- How we use AI in Headcore’s operations
Frequently asked questions
Can you use AI to review AI-generated content?
Yes, as long as the setup is designed with care. Studies on LLM-as-a-judge show position bias, a preference for longer answers and a tendency to favor texts that sound familiar. Review works better in a separate context, with no access to production, with criteria written beforehand and final approval from a person.
Does Google penalize content made with AI?
Google’s guidance, updated in October 2026, asks that all AI-generated content be fact-checked and reviewed by hand before publishing, including titles, meta descriptions, structured data and alt text. The page focuses on the accuracy and quality of what gets published.
What is the difference between brand review and text review?
Brand review asks whether the brand would say it, according to its positioning, voice and written rules. Text review looks at clarity, rhythm, mistakes and the patterns that give away AI writing. A text can be flawless and still be off-brand.
Does an AI detector take care of review?
No. A detector tries to tell whether a text seems machine-written, and that says nothing about whether the facts are right or whether the brand would say it. The three layers of review are still needed, with or without a detector.
How do you review a lot of content without stalling production?
Separate the concept from the batch. The concept is judged piece by piece; the variations go through automated spec checks and sampling, with the first piece of each family as the reference. Errors that repeat become rules, which cuts down review in later rounds.
What is a blind review?
It is when each reviewer reads the piece without seeing anyone else’s opinion and records their assessment before any discussion. This keeps the first comment from swaying the rest. On our review panel, each lens’s initial assessment is frozen before the debate.
Sources: Google Search Central, guidance on generative AI content (updated on 01/10/2026) · Zheng et al., Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena, 2023 · Wataoka et al., Self-Preference Bias in LLM-as-a-Judge, 2024.
