Blog / AI & search /
Using AI to compare content drafts without outsourcing judgment
Two drafts land in your inbox. Both read well. You ask an AI tool which is better, and it confidently picks one.
Hold on. Is it the one that answers your reader’s question, or just the one with nicer adjectives?
AI can help you compare structure, clarity and coverage. It shouldn’t make the editorial decision by declaring a winner.
Define what “better” means
Start with the purpose of the page. Who is it for, what question brought them there, and what should they be able to do afterwards? Pick a few criteria before comparing: accuracy, usefulness, completeness, clarity, tone and fit with the page’s intent. Some criteria are non-negotiable. A draft with an unsupported claim should not win because it reads smoothly.
Give the comparison tool both drafts and the same brief. Ask it to identify differences and cite the specific passages behind each observation. Avoid vague prompts like “Which is best?” I don’t trust a ranking with no standard behind it. They invite confident rankings without a standard. Ask instead where each version answers the question, where it creates friction and what both are missing.
Inspect the comparison yourself
Treat the output as an index to review, not a verdict. I’d rather open every cited passage myself than take the summary’s word. Open the cited passages and decide whether the model’s description is fair. It may call a caveat “unnecessary” when the caveat protects accuracy. It may praise detail as “comprehensive” when it distracts the reader. It may prefer familiar phrasing over a specific explanation that better matches the audience.
Check claims against sources separately. Draft comparison cannot establish whether facts are true, current or properly qualified. If one version includes an example, confirm that it is real and authorised. If both versions use the same unsupported number, the comparison will not make it safe. Keep fact checking as its own review step.
Make a reasoned choice
Sometimes one draft is clearly stronger for the stated goal. Sometimes the answer is a mix: the direct opening from one, the explanation from the other, a rewritten transition. Record the reason for the choice in a sentence. That gives the next revision a basis, and stops anyone treating AI’s preference as an objective score.
If the drafts are both weak, write a better brief or ask the subject expert for the missing detail. Don’t keep generating variants just to find one that sounds convincing. More alternatives can create more review work without adding evidence or insight.
Keep ownership with the editor
The person responsible for publication owns the final decision. When that’s me, I make sure I can explain the choice. They should be able to explain why the selected draft serves the reader and what was verified. AI can point to differences quickly, but it cannot take responsibility for the accuracy, tone or business fit of the page.
Compare structure with the reader’s task
A useful comparison can reveal sequence problems. Does one draft give the answer before background? Does it explain a prerequisite before asking the reader to act? Does it separate the basic path from exceptions? Ask AI to point to where each draft supports a step in the reader’s task, then verify the map yourself. A longer draft is not automatically more complete if the extra material is unrelated.
Check whether both versions are working from the same brief and source set. If one had more information, the comparison is not fair. Do not let the model fill missing knowledge by guessing which draft is correct. Give both versions the same facts, or compare on something narrower like clarity and leave the facts to the editor.
Preserve useful differences. A slightly more conversational passage may fit the author’s voice; a more structured explanation may be easier to scan. The decision can combine both rather than choosing a whole draft. Keep a note of the principle behind the edit, such as “state the caveat before the recommendation”, so the next revision remains coherent.
You can also compare the drafts against a short list of required points. Ask whether each one appears, where it appears and whether it is supported. This helps reveal omissions without turning the model into a quality judge. The editor still decides whether a point belongs in the article. This helps most when the drafts are structured differently and a side-by-side read hides what’s missing.
If you are using the comparison for a headline or introduction, judge it against the content that follows. A strong opening that promises an answer the article never delivers is not the better version. Check the transition from the initial claim to the supporting explanation. Also notice whether a draft assumes the reader already understands a term. The choice should improve the whole reading experience, not merely make the first paragraph punchier.
Before making a choice, read both drafts as a real visitor would rather than as separate documents scored against a list. Notice the order of ideas, the effort required to understand the page and whether the next step is obvious. A model may identify that one version is shorter, but only the editor can decide whether the removed detail was necessary. If readers need a caveat to act safely, keep it even when it makes the section less compact.
This approach works for introductions, outlines or revised sections, especially when reviewers are short on time. It’s less useful when the question depends on confidential context or specialist judgement the tool doesn’t have.
Related reading
- AI and SEO: using AI to audit a content catalogue
- Using AI to research competitors’ content (carefully)
- Why I still write my own drafts
What criteria would you set before comparing the next pair of drafts?