Blog · 2026-10-15 · baryreview editors

A reviewer's checklist for agencies shipping AI-assisted content

Why a checklist and not a read-through

An agency that drafts with AI has a production problem it did not have two years ago: the writing is faster than the reading. A writer who used to produce two posts a day and read both now produces eight and skims them. Agency AI content QA needs to be a fixed procedure, run the same way on every piece, with a record that it happened. A read-through depends on the reader’s energy at 4 p.m. A checklist does not.

This is the checklist we would run before delivery, in the order that catches the expensive problems first. It is written for a reviewer who did not generate the draft; the person who prompted it is the worst-placed to notice what is missing.

The checklist, in eight stages

Stage 1: Brief fit

Before reading a sentence of the body, put the brief next to the draft.

• Does the title promise what the brief asked for, and does the body deliver it?
• Is it aimed at the reader the brief named? A post for facilities managers that explains what HVAC stands for is aimed at the wrong reader.
• Is the length right for the purpose, or was it padded to a word count?
• Did the model answer a nearby, easier question instead of the one asked? The most common brief failure, and the hardest to see from inside the text.

If Stage 1 fails, stop. Regenerate or rewrite. Editing a draft that answers the wrong question is wasted work.

Stage 2: Facts

Read the draft once marking only claims: statistics, dates, prices, names, product features, study findings, quotations, legal or regulatory statements.

• Every number has a source you can open, and the source says that number.
• Every named study, report or survey exists, and found what the draft says.
• Every quotation is real, attributed to the right person, and in context.
• Every date is correct, and nothing is stated in the present tense that stopped being true.

The default on an unverifiable claim is to cut it. A paragraph with no statistic is weaker than one with a true statistic and much stronger than one with an invented one.

Stage 3: Specifics

AI drafts are vague where a human expert would be concrete, and concrete where they have no basis to be.

• Does the piece contain at least one real example, decision, number or opinion that could only have come from someone who knows the subject?
• Are the “examples” generic placeholders (“a small business owner might find that…”) that should be replaced with something the client actually did?
• Is there a point where the draft should take a position and instead lists both sides?
• Where the draft is specific, is the specificity earned (Stage 2) or decorative?

Flag the gaps for the client rather than filling them with more generated text.

Stage 4: Voice

• Opening paragraph: does it start with the subject, or with a warm-up (“In an ever-changing landscape…”)? Cut the warm-up.
• Hedges and announcements: “it’s important to note,” “it’s worth mentioning,” “as we’ve discussed.” Cut.
• Tidy triads: three adjectives, three benefits, three “whether you’re a…” clauses. Break the pattern.
• Empty transitions and summaries at the end of each section. Cut.
• Does it sound like the client? Compare a paragraph against something the client wrote. If the client is blunt and the draft is cushioned, fix the draft.

Stage 5: Structure and logic

• Does each heading describe what is under it?
• Is anything said twice? AI drafts often restate the introduction as the conclusion and repeat a point in two sections under different headings.
• Does the conclusion follow from the body, or does it assert something the body did not establish?
• Is there a contradiction? Read the first and last mentions of each key point together.
• Are there sections that exist only because the outline had a slot for them?

Stage 6: SEO basics

Lightly: a searcher who lands on the page should get what they came for.

• The primary term is in the title, the first paragraph and one heading, and nowhere it sounds forced.
• The meta description is specific to this page and the right length.
• Internal links point to real pages on the client’s site, with sensible anchor text.
• External links open and go where the text says.
• The page has a byline and a date. Google’s guidance on helpful, reliable, people-first content asks whether it is “self-evident to your visitors who authored your content” and whether the use of AI is “self-evident to visitors through disclosures or in other ways.” The client should answer those deliberately, not by default.

Stage 7: Compliance

• Does the piece make a claim that a regulator in the client’s industry would care about? Health benefits, financial outcomes, guarantees, comparative claims against named competitors, “best,” “safest,” “#1.”
• Does it use a trademark, a brand name or a competitor’s product in a way that could be a problem?
• Does a paragraph read like a specific source? Search a sentence of it.
• Does it give advice in a field where advice is regulated, and is there a disclaimer that matches what the page actually says?
• If images or captions are part of the deliverable, do they show what the caption claims?

Stage 8: Client-specific claims

Last, and separate from Stage 2 because the source is different: everything the draft says about the client’s own business.

• Products, features, prices, plans, hours, locations, service areas, certifications, staff, history, awards. The model does not know these. It will invent a founding year, a “team of experts,” a “24/7” promise, a “money-back guarantee.”
• Each is checked against the client’s own site or an approved fact sheet; anything not on that list is flagged, not guessed.
• Promises are the priority. “We guarantee,” “always,” “every,” “any” are words the client’s lawyer would want to see before the client’s customers do.

Running it in about fifteen minutes

Fifteen minutes is enough for a 1,000 to 1,500 word piece when the stages are run in order. The budget:

• Minutes 0 to 2: Stage 1. Brief beside draft, title and headings only. Decide whether to continue.
• Minutes 2 to 7: Stage 2 and Stage 8 together. One pass marking every claim, with a search engine and the client’s fact sheet open. Verify what takes seconds; mark the rest for removal or for the client.
• Minutes 7 to 11: Stages 3, 4 and 5. A second pass, reading for the reader. Cut filler as you go, note where a specific is missing, check each heading against its section, read the first and last paragraphs together.
• Minutes 11 to 13: Stage 6. Title, meta, links, byline.
• Minutes 13 to 15: Stage 7. Scan for regulated claims, superlatives, trademarks, and any paragraph that reads like a source.

Two rules make the budget hold. First, the reviewer records rather than fixes. Write the correction as an instruction, quoting the passage, and let the writer or the AI apply it. Second, if a claim takes more than a minute or two to verify, it goes on the client list; it does not eat the rest of the review.

A piece with more than about ten flags has failed the review and goes back to the writer or up to a full review. That is a result, not a failure of the process.

Documenting it for clients

A client who has never seen the review will assume it did not happen, and the first time something slips through, that assumption becomes the conversation. One page per piece, kept with the deliverable:

• Header: URL or title, date, reviewer’s name, draft version, which AI tool produced the draft (if the client wants to know).
• Stage results: eight lines, pass or flagged, with the count of flags per stage.
• The flags: a numbered list, each quoting the passage and stating the correction or the question for the client. A correction that quotes its passage can be applied by a writer, an AI or the client without a meeting.
• Claims removed: which unverifiable statements were cut and why. A client who sees “removed: ‘73% of buyers prefer…’ because no source could be found” learns something about their drafts and trusts the agency more, not less.
• Open items for the client: what they need to supply or confirm.
• Sign-off: who approved delivery and when.

Keep the template boring and identical across clients. After a month, the record shows which problems the drafts produce most, and the agency can fix the prompts or briefs upstream.

When to escalate to a full human review

The fifteen-minute checklist is a filter. Some pieces should not go through a filter; they should be read in full by an editor who is not under a timer. Escalate when:

• The topic is legal, medical or financial, or any area where a wrong statement can harm the reader. Every claim needs a primary source and a professional sign-off; a fifteen-minute pass is not that.
• The piece makes claims about the client’s product that the agency cannot verify, and the client has not supplied a fact sheet.
• The fifteen-minute review produced more than about ten flags. A checklist samples that many problems; it does not catch them.
• The piece is high-visibility: a homepage, a flagship landing page, anything the client will promote or that carries a named author’s reputation.
• The client has been burned before. If they hired you because the last agency shipped a made-up statistic, the first few pieces get the full read regardless.
• The reviewer is also the writer and nobody else is available. Fifteen minutes is better than nothing, but it is not a second pair of eyes.

A full human review can be a senior editor in-house, if one is free. If not, it can be bought per article: publish the draft to an unlisted URL, order a review, and a professional editor reads the whole page and returns numbered corrections in the same quote-and-instruct format, usually within one business day. How that fits an agency workflow, including preview URLs and batches, is on our agencies page.

The short version

Run the eight stages in order, record flags instead of fixing on the spot, and keep a one-page record with every deliverable. Escalate anything regulated, high-visibility or failing the checklist to a full read by someone who did not write it. The checklist makes AI-assisted volume deliverable; the record makes it defensible.