Because a language model writes toward the average of everything it has read. Ask it a question and you get the most expected answer, phrased in the most expected way. That is exactly what makes it useful for a first draft and exactly what makes the result forgettable if you stop there.
You can feel it after about two paragraphs. The confident opener, the balanced middle, the tidy conclusion that restates the opener. Nothing is wrong. Nothing is memorable either, and readers have gotten very good at spotting the pattern.
The fix is not to abandon the tools. It is to give them constraints tight enough that the average is no longer an option. Here is how we think about that.
Three things, mostly. Hedging instead of taking a position, abstraction instead of specifics, and rhythm that never varies. A model hedges because hedging is safest, abstracts because it does not know your particulars, and falls into a steady sentence rhythm because that is the shape of most published prose.
Hedging is the worst of the three. Phrases like "it depends on your needs" and "there are many factors to consider" are technically true and completely useless. They exist because the model has no stake in the outcome, and a reader can tell.
Abstraction is the next problem. A model will write "choose the right tool for your workflow" where a practitioner would write "use a proper applicant tracking system if you hire more than twenty people a year, and a spreadsheet if you do not". The specific version is only possible if somebody supplies the specifics.
Rhythm is subtler but it is the tell most readers react to without naming it. Every sentence lands at a similar length. Nothing is short. Nothing runs long and awkward. Real writing varies because real thinking varies.
No, and this surprises people. Google's guidance on creating helpful content is clear that the method is not the issue. What it objects to is intent, stating that "if you use automation, including AI-generation, to produce content for the primary purpose of manipulating search rankings, that's a violation of our spam policies."
Read that carefully, because the condition is the purpose, not the tool. Content written to help a reader is fine whether a person or a model drafted it. Content written to occupy a keyword is a problem whether a person or a model drafted it.
Google also suggests transparency as a reasonable practice, asking whether "the use of automation, including AI-generation, is self-evident to visitors through disclosures". That is framed as a question to consider rather than a requirement, but we think erring toward openness is the right instinct.
So the real risk is not detection. It is producing something nobody wants to read. That risk existed long before these tools did.
Scaled content abuse is Google's name for mass-producing pages that do not help anyone. Google defines it as "when many pages are generated for the primary purpose of manipulating search rankings and not helping users", and lists "using generative AI tools or other similar tools to generate many pages without adding value for users" as an example.
The word doing the work there is value, not volume. Publishing often is not a violation. Publishing often without adding anything is. Google's policy also names scraping and "stitching or combining content from different web pages without adding value", which is a fair description of what an unguided model produces when you ask it to summarise a topic.
The honest self-test is uncomfortable but quick. Take your last ten posts and ask what a reader could learn from them that they could not learn from the top result on the same query. If the answer for most of them is nothing, the volume is the problem regardless of how the words were produced.
This is the same idea as information gain, which we wrote about in our post on what information gain is and how you add it. A model cannot generate information gain on its own, because by definition it is working from what already exists.
Rules specific enough to fail a draft. "Write in our brand voice" is not a rule, because nothing can violate it. "Never use the word leverage as a verb" is a rule, because you can check it and a draft can break it.
We keep ours in four parts. Banned words and characters, sentence and structure rules, voice rules about person and stance, and evidence rules about what may be stated as fact. Each one is written so a person or a script can check compliance without judgement calls.
The banned list is the most useful and the least glamorous. Ours forbids em dashes and en dashes entirely, because they are the single most reliable signature of an unedited model draft. It also kills the usual vocabulary tells: delve, tapestry, testament, landscape used figuratively, and any sentence beginning "in today's fast-paced world".
Structure rules do more work than people expect. Ours require question-based headings, a direct answer in the first forty to sixty words of each section, and prose paragraphs rather than bullet lists. A model left alone will produce bullet lists for everything, because lists are the average shape of web content.
Make each rule checkable, put the hard constraints near the end of your instructions, and give examples of the wrong version as well as the right one. Models follow negative rules poorly when the rule is abstract and well when it is concrete enough to pattern match against.
Show the failure. "Do not hedge" is weak. "Do not write sentences like 'it depends on your specific needs'. Instead write 'use option A unless you have more than fifty products, then use option B'" gives the model a target rather than a prohibition.
Then verify mechanically rather than by eye. A search for the banned characters takes a second and never gets tired or generous at the end of a long day. Anything you can check with a script should be checked with a script, and the review time you save goes into the parts only a person can judge.
Expect drift over long outputs too. Rules followed carefully in the first section loosen by the eighth, so the end of a long draft deserves more scrutiny than the beginning. We look hardest at the closing sections, which is where generic phrasing creeps back in.
Supply the opinion, the specifics, and the judgement about what to leave out. Those three are where a model has nothing to contribute, because all three depend on having done the work and having something at stake in the answer.
The opinion has to be real. If your article recommends one approach, somebody at your company should actually believe that and be willing to defend it in a meeting. A position generated to sound decisive is worse than a hedge, because now you have to defend something nobody chose.
The specifics have to come from your work. The number of steps in your process, the mistake clients make most often, the thing you tried that did not work. A model will happily invent a plausible version of all three if you let it, and inventing them is the fastest way to lose the trust the article was meant to build.
Knowing what to cut is the last one, and it is the hardest to delegate. A model asked to cover a topic covers all of it. A person who understands the reader knows that three of those sections are obvious and should go. We wrote more about that balance in our post on whether you should let AI write your website copy.
Never let a statistic into a draft that you have not opened the source for. Not a source you remember, not a source that sounds right, but a page you actually retrieved. This is the rule we enforce most strictly, because it is the one with real consequences attached.
Fabricated statistics are the characteristic failure of AI-assisted writing, and they are dangerous because they are so plausible. A made-up figure attributed to a real research firm reads exactly like a real one. The difference only shows up when somebody checks, and the person checking is often a journalist or the company you misquoted.
Our working rule is that deleting an unverifiable claim is always correct. An article with three real numbers is stronger than one with eight where two are invented, because one invented number puts every other number on the page in doubt. We go through the process in our guide to fact-checking AI-generated content before you publish.
Apply the same rule to your own company. Invented experience is still invented, even when it flatters you. If you cannot support a sentence like "we tested this and saw a forty percent lift", write "here is how we would evaluate this" instead. Honest generality beats manufactured specificity every time.
We think it helps in the way that matters now. Search engines and answer engines both have an abundance of correct, generic answers to choose from. What is scarce is a source with something specific to say, and scarcity is what gets cited.
Consider how an answer engine builds a response. It reads several pages, finds the consensus, and writes a summary. Pages that only restate the consensus are interchangeable, so any one of them can be dropped. A page carrying a position or a detail the others lack has to be cited by name, because nothing else supports the claim.
Voice also does something no ranking factor captures. A reader who recognises how you write will click your result over a stranger's next time. That is a slow effect and it does not show up in a rank tracker, but it is the only durable advantage in a market where everybody can produce a competent article in a minute.
Write your banned list this week. Ten words and characters you never want to see in your content, taken from reading your last few posts and noticing what makes you wince. That single page will improve your output more than any change to which model you use.
Then add one evidence rule and one structure rule, and check the next draft against all three. Keep the guide short enough that people actually read it. A two page style guide that gets followed beats a twenty page one that lives in a folder nobody opens.
If you are publishing with AI in the loop and you are not sure whether the output still sounds like you, we are happy to look at it with you. Working out what a company actually sounds like, and then writing that down as rules, is a genuinely interesting problem. Let's talk, and you can find us at phoenix.studio.
Tell us where you want to go. We'll tell you how we'd get you there.