A decent model will write 2,000 words about search in about 40 seconds. It will use clean headings, balanced paragraphs, and a conclusion that repeats the introduction. It will also say almost nothing a reader could check.
That is the tell. Not the vocabulary, not the cadence, not some hidden watermark. The draft reads as machine-made because it carries no evidence that cost anyone anything. No dated source. No number from a real system. No position a smart person might argue with.
I run AI through most of my publishing system. The output does not read like AI, and the reason is structural. The evidence layer is mine, and it sits upstream of every prompt. That layer is the job of content engineering, and it separates a page that gets quoted from one that gets scrolled.
TL;DR
- The real problem: a draft sounds like AI when its paragraphs contain no checkable fact, so fix the input first.
- Google’s position: quality matters, production method does not, and commodity content is explicitly the loser in generative search.
- The four tells: no numbers, flat rhythm, symmetrical hedging, and an ending that restates the opening.
- The workflow: evidence folder, one-sentence claim, assembly, voice pass, and a gate that rejects real drafts every week.
Sounding like AI is an evidence problem
Most teams treat this as a tone problem. They send the draft to a rewriter, ask for a warmer voice, and publish on Thursday.
The draft still has no source in it. An editor can fix rhythm in an afternoon. Nobody can edit a fact into a paragraph that never had one.
That is why the same paragraphs keep showing up across a category. Ten vendors describe the same five trends, cite the same three blog posts, and land on the same advice. A buyer reads two of them and stops.
There is a second cost, and it compounds. Google’s systems and the answer engines now read your page as evidence, not as brand copy. A paragraph with no named number, no dated source, and no first-hand observation gives a retrieval system nothing to quote. It gives a human reader nothing to remember. The page can rank for a while on domain strength and still lose the citation to a smaller site with one specific claim.
The blunt test is simpler. Pick a sentence at random and ask what it came from. If the honest answer is “the model,” the sentence gets rewritten or cut. Tone never enters the conversation.
If a sentence has nothing in it that could be wrong, it has nothing in it worth citing. Change the input before you touch the tone.
What Google actually says
Google published its position on AI content on February 8, 2023. The wording has held. Its stated focus is “the quality of content, rather than how content is produced.” The same post says automation used mainly to manipulate rankings violates the spam policies.
Google went further in its 2026 guide to AI features in Search. That guide makes a narrower claim. Unique, non-commodity content “will likely influence your website’s presence in generative AI search in the long run more than any of the other suggestions.”
Read the guide’s example closely. A first-hand review carries a perspective built on personal experience. A summary of existing content restates what is already available elsewhere. That is the exact failure mode of a draft built from a keyword list and a competitor scan.
The guidance is generous about tools and strict about inputs. You can use the model for the whole assembly. You cannot let it be the source.
Arguing about AI detectors wastes a week, and no search engine uses one as a ranking input. The only detector that matters is a reader who has already read forty versions of your paragraph.
The four tells I gate on
None of these tells are about word choice, and all four survive a thesaurus.
| Tell | What it looks like | What I do |
|---|---|---|
| No checkable number | Claims about growth, cost, or speed with no figure, date, or source | Attach one sourced number or delete the paragraph |
| Flat rhythm | Five sentences per paragraph, every sentence 18 to 22 words | Rebuild as one-sentence and six-sentence blocks |
| Symmetrical hedging | Every claim paired with its opposite, so nothing is decided | Pick the position and defend it in the next line |
| Recap ending | Last paragraph restates the introduction in softer words | End on the next action, a number, or a failed test |
The first tell does the most damage. A page full of unverifiable claims cannot be quoted, because a model summarizing it has nothing stable to lift. The paragraph reads fine and cites to nobody.
The second tell is easier to spot in a screenshot than in a doc. Scroll the page and look at the shape of the text, because readers feel nine matching paragraph heights before they can name them.
Hedging is the one that costs you money. A vendor that will not say which buyer should not buy has no point of view. Buyers now run that comparison through a chatbot in about nine seconds.
The workflow I actually run
Five stages, in this order. The model touches two of them.
Two external sources with dates, one internal data pull, and one customer conversation. If the folder is thin, the article does not start. This step takes longer than the writing and decides the outcome.
Name the argument, the evidence behind it, and the reader who would disagree. If you cannot do it in 20 words, the model will produce a topic instead of a piece.
Every number in the brief carries a source line. The model orders, compresses, and transitions. It does not add facts, and it does not add examples. One invented statistic ruins a page that took a week to research.
Not “make it sound human.” Match the rhythm of your last ten published posts. I check average sentence length, short-sentence share, first-person rate, and a list of phrases I never use.
A gate that never rejects a draft is decoration. Mine catches missing sources, banned phrasing, and paragraphs with no claim. Rejection is the cheapest quality control available to a one-person team.
What this looks like on my own system
I publish across three properties with one operator and a nightly system. KSB sits at 253 posts as of November 2026. Most of those drafts touch a model at some point, and the ones that hold position share a trait.
They carry a primary data point and a named source. That is it. The prose can be plain. The evidence cannot be borrowed.
The LinkedIn study I ran makes the same point from a different angle. My team analyzed 2,279 posts from 246 creators. Accounts posting three to five times a week had a median engagement of 195. Daily posters had 92. Volume was not the variable that separated them. Specificity was.
My AI use is heavy and public, and I am not going to pretend otherwise. I use it for outline expansion, first drafts, compression, and the boring middle of an edit. I do not use it to decide what is true.
The material only I have goes in first. Three properties worth of search data. The calls where a CMO explains why the last agency got fired. The tests that failed and the two that worked. A model cannot generate that, and no prompt makes it up.
This is the part most “AI SEO” advice skips. The people getting cited are not better prompt writers. They are operators with proprietary inputs, and the model just helps them ship faster. If you want the role described properly, I wrote about the content engineer separately.
What to measure in 90 days
Advanced Web Ranking counted AI Overviews on 83.5% of the US searches it tracks in November 2026, a record. The answer layer is not an experiment anymore. It is the first screen most buyers see.
Surfer’s analysis of 405,576 AI Overviews found that 52% of cited sources also rank in the top 10. Which means roughly half of the cited pages sit outside it. Being quotable is a separate contest from being ranked.
Four things get tracked per page, and none of them are word count. Impressions and position for the target query. Whether the page earns a click. Whether it shows up in an AI answer for its main question. And whether a sales call mentions the argument.
The fourth signal is the slowest and the most honest. A page that changes how a buyer frames the problem is doing more work than a page with better dwell time. The mechanics of how those citations get chosen sit in what actually gets cited in AI search. The structural half is in internal linking as a growth lever.
At day 90, keep pages in the top 20 and expand them. Pages stuck between 20 and 50 get a new title and a sharper opening. Pages below 50 get folded into a stronger URL instead of fed another rewrite. Adding pages to a cluster that is not ranking is how teams end up with 400 posts and no traffic.
Use the model, own the evidence
Using AI for SEO is fine. Using it as a substitute for evidence is what produces the prose everyone recognizes and nobody quotes. The tool is not the risk, and an empty source folder is.
Start with one page this week. Rebuild the evidence folder, write the claim in a sentence, and let the model assemble only what you can defend. Then run the page through a gate that will actually reject it. If you want the full build, the AI content accelerator installs the evidence layer, the voice pass, and the rejection gate in four weeks.
Readers spot AI prose because it has no cost in it. Sources, numbers, and a position are the cost. Put them in before the prompt, and the output stops reading like a machine wrote it.
Sources: Google Search Central, Google Search’s guidance about AI-generated content (February 8, 2023, Danny Sullivan and Chris Nelson); Google Search Central, Optimizing your website for generative AI features on Google Search (2026); Advanced Web Ranking, Google AI Overviews stats (November 10, 2026); Pew Research Center, Americans and AI 2026 (June 17, 2026, survey fielded February 17 to 23, 2026); G2, 2026 Buyer Behavior Report (July 22, 2026, n=1,000+ B2B software buyers); Surfer, Google AI Overviews study of 405,576 searches (November 5, 2026).
