Editing AI output: what actually matters, in order

Everyone agrees you should edit AI output. Almost nobody says what to edit for, so in practice people do a style pass - smooth the clunky sentences, cut a few adjectives, publish. That is the least consequential edit available, and it feels productive, which is the problem.

Here is the order I actually work in. It is deliberately ordered by consequence: the first pass can save you from a regulator or a refund, and the last one is taste. If you only have twenty minutes, do the first two and stop.

Pass one: claims that carry consequences

This is the only pass that is genuinely non-negotiable, and it takes the longest on the pages that matter most.

A model asked to write about your business does not leave gaps where the facts should be. It fills them with the most plausible-sounding version of a business like yours, confidently, because generating fluent text and stating facts are the same operation to it. The output does not look uncertain. That is exactly what makes it dangerous.

What to read for, hardest first:

• Anything a regulator could read. Outcomes, success rates, safety standards, qualifications, registrations. For a clinic or anyone in a licensed trade this is not a copy problem, it is a compliance problem, and "the AI wrote it" is not a defence anybody has to accept.
• Numbers. Years in business, team size, jobs completed, response times, prices. A wrong number reads exactly as smoothly as a right one, which is why they survive editing that catches everything else.
• Guarantees and warranties. A model will invent a warranty period because most businesses like yours have one.
• Service areas. Suburbs and towns you do not actually cover, which generate enquiries you have to decline and reviews you did not earn.
• Accreditations and memberships. The single most common invention I see, because they are so plausible.

The way to make this pass shorter is to stop it being needed: give the model a context document with the real facts and an instruction to flag gaps rather than fill them. That does not eliminate invention, so you still read every claim, but it changes the default. I have written that up in what AI needs to know about your business before it writes a word and how to stop AI inventing things about your business.

Read the numbers hardest. Everything else you might notice while skimming. A wrong number you will not.

Pass two: does this page own the right search

The second-most expensive failure is structural and completely invisible if you read the page on its own. Two of your pages targeting the same search term compete, Google picks between them inconsistently, and both rank worse than one strong page would have.

The check is quick if you have a keyword map and nearly impossible if you do not, which is the argument for building one first - plan the site first, generate second. With a map, the pass is:

1. What single search term does this page own? If you cannot say it in one phrase, the page is not finished.
2. Does any existing page already own that term? If yes, you do not have a new page, you have an expansion of an old one.
3. Does the page cover things another page owns? Cut them and link instead.
4. Is the term in the title, the first heading and the first paragraph, in the words a customer would actually use rather than the industry version?

The failure mode this prevents is the expensive one, and it compounds: the cannibalisation trap in an AI buildout is what it looks like at scale.

Pass three: the links

A model writing one page has no idea what else exists on your site, so it either writes no internal links at all or invents links to URLs that sound like they should exist. The second is more common than people expect, because the invented URLs look so reasonable nobody checks them.

• Click every internal link. Every one. This takes ninety seconds and catches the invented ones.
• Decide which existing pages will link TO this new page, and add those links yourself. A new page with nothing pointing at it is an orphan, and orphans do not rank.
• Check the anchor text says what the destination is about rather than "click here" or "learn more".

Why this matters more than it sounds is in internal linking is the thing AI gets most wrong.

Pass four: put specifics back in

Now, and only now, the writing itself. The highest-value edit is not smoothing prose - it is adding back the things only you know.

Go through and replace every general statement with a specific one. "Fast response times" becomes "we aim to be there within two hours in the central suburbs and we will tell you on the phone if we cannot". "Experienced team" becomes the actual number of years and the actual number of people. "We handle all types of X" becomes the three you actually want and the one you decline.

This pass does two jobs at once. It makes the page persuasive, and it makes it un-copyable - a competitor cannot lift a sentence that contains your real policy.

Pass five: the opening two paragraphs, by hand

Rewrite the first two paragraphs yourself, from scratch, without looking at the draft. Not edit - rewrite. This is the highest-return sixty seconds in the entire process, because the opening is where a reader decides whether there is a person behind the business, and it is the part a model writes most generically.

Say the thing you would say if the customer were standing in front of you. It will be less balanced and more useful than what was generated, and that is the point.

Pass six: style, if you have time

Last, and genuinely optional. Vary the paragraph lengths. Cut the sentence that restates the heading. Delete "in today's fast-paced world" and anything adjacent to it. Remove one adjective per paragraph.

I put this last deliberately. A page with perfect prose and an invented accreditation is a worse page than a clunky one with correct facts, and the industry has the priority backwards.

A worked example

Here is a paragraph as generated, for a plumbing business:

With years of experience serving the local community, our team of qualified professionals is committed to providing fast, reliable plumbing solutions. We pride ourselves on our attention to detail and customer satisfaction, ensuring every job is completed to the highest standard.

Nothing in that is wrong. It is also true of every plumber in the country, contains no fact, and would survive a style edit untouched - which is exactly why a style edit is the wrong pass. Here is the same paragraph after passes four and five:

I have been doing this since 2011, mostly in the western suburbs, and about half my work is burst pipes and blocked drains at inconvenient hours. Callout is $95 and it comes off the job. If I cannot get to you within two hours I will say so on the phone rather than let you wait.

Same length. Every sentence contains something a competitor could not honestly copy. Note what changed and what did not: the grammar was never the problem, and nothing was made more eloquent. Facts were added and generalities removed, which is the entire edit.

How long this actually takes

For a standard service page, once you have a context document and a keyword map, this is fifteen to twenty-five minutes. Without those, it is closer to an hour, because passes one and two turn into research rather than checking.

That is the honest trade with generated content. The writing gets much faster. The checking does not, and the checking is where the value was. If a content programme has no budget for these six passes, it is not producing content - it is producing pages, and those are different things. The wider structure is in using AI for your website content without wrecking it.