What a line-by-line pass does
A line-by-line pass takes a finished AI draft and checks every sentence against four questions: does it say anything, is what it says specific, does it sound like the person publishing it, and can that person stand behind it if a reader or a regulator asks. It is not a proofread. Spelling and grammar are rarely where AI copy goes wrong; the faults sit inside sentences that are grammatically clean and still empty, still vague, or still making a claim nobody checked.
The pass works sentence by sentence, because a paragraph can read smoothly while every sentence in it fails one of the four questions. Read the draft once for sense, then go back to the first line and do not move to the second until the first has earned its place.
Four faults to catch
Padding: the sentence that repeats the one before it
AI models are built to elaborate, so a draft will often make a claim once as a plain statement and then again, a sentence later, dressed in different words: "This makes the process straightforward. It also ensures a smooth and seamless experience for the customer." The second sentence adds nothing; it just proves the first one was true by saying it twice. Delete the restatement and check what is left still reads as a complete thought. If it does, the deleted sentence was padding.
Specifics the model could not have known
An AI tool does not know your product, your delivery times, or your customer numbers, so where a draft needs a specific detail it will often invent a plausible one instead: a percentage, a turnaround time, a claim to be "the UK's leading" something. None of these were supplied, so none of them are facts, and a reader has no way to tell a real figure from a confident guess. The pass finds every specific claim and asks where it came from; if the answer is nowhere, either replace it with a real figure or drop the claim and let the sentence stand on what is actually true.
Tone that has drifted from house voice
Left unguided, an AI draft settles into a house style nobody chose: safely formal, mildly enthusiastic, and interchangeable with the draft it would produce for a different brand on the same brief. The pass reads each sentence and asks whether it sounds like this business talking to this reader, or like a summary of what a business in this category tends to say. Where the second is true, the fix is usually to cut a hedge, name the actual thing, or drop an adjective that is doing the job of a fact.
Claims the copy cannot back up
AI drafts are confident by default, which means they will happily produce a claim about being fastest, safest, or most trusted without any evidence behind it. In UK advertising these claims sit under the CAP Code, which requires that anything stated as fact can be substantiated; the plain-English guide to ASA and CAP rules covers what that means in practice. The pass flags every superlative and every unqualified claim of fact, and either finds the evidence for it or rewrites the sentence so it stops claiming what nobody can prove.
The order to work in
A pass is easier to do in a fixed order than by jumping around the page on instinct.
- Read the whole draft once, without editing, to get a sense of what it is trying to do.
- Go sentence by sentence from the top, checking each one against the four questions above.
- Mark every specific claim (numbers, superlatives, named outcomes) and confirm each one against something you actually know.
- Read the marked-up draft aloud; a sentence that no one would actually say is the one to rewrite next.
- Check the result against the guidance on what AI is good at and bad at in commercial copy, since the same weaknesses tend to recur across drafts.
A checklist version of this same pass, usable against your own draft, sits on the copy review checklists page.
Once the edit is done, it is worth testing the result rather than trusting your own ear alone. The readability and clarity checker runs entirely in your browser, so the draft you paste in never leaves the page, which matters here because the text is your own unpublished copy. It returns a Flesch-Kincaid reading age, the average sentence length, a count of passive-voice sentences, a list of flagged jargon terms, and an overall verdict of Clear, Review or Revise, with the sentences behind that verdict marked. It cannot tell you whether a claim is true or whether a sentence still sounds like the AI draft it came from, since those are exactly the judgements the pass above is built for, and the verdict itself can be wrong on any given sentence. But a page that still comes back Revise after a full line-by-line pass is worth reading again before it goes live.