AI Can Do the Work.
But Is It Doing the Right Work?:
The Human Strategy Behind Successful Digital Marketing.

You already sit in the chair. You pay for ChatGPT, or Claude, or Gemini. Between jobs you tell it the trade, the towns, and what the page should say. Then you put the draft on the site, in an ad, or on the profile.

That is guiding the tool. A professional still does this job better than that setup. The rest of this page is the record: what ranks, what the model is actually doing when it writes, and how often a false line still gets published by people who do this every day.

Google ranks the finished page. It does not rank the care you put into the prompt. Three large samples, taken with three different detectors, show the same shape. Pages that read as a person’s work hold the top. Pages that read as generated show up more often further down.

Detectors are imperfect. A single page can be mislabeled. The pattern is in the samples, not in one URL.

The top result is usually a page a person wrote

Graphite looked at the first two results pages for 31,493 keywords in June 2025, and kept English articles. In that set, 86% of the ranking articles were human-written and 14% were generated. Only 7% of the articles in position 1 were generated. On keywords where both kinds of page appeared, they paired one of each at random. The human-written page ranked higher. The chance that this was a coin flip was vanishingly small.

Semrush did a separate count in November 2025. Twenty thousand keywords, the top 10 results, blog pages only: 42,000 pages, graded with a different detector. A page in position 1 had an 80.5% chance of being classed human-written and a 10% chance of being classed generated. About eight times. From about position 5 down, the gap gets smaller. “We showed up on the first page” can hide the split that matters, which is the top spot.

Getting the page onto Google is not the job

Ahrefs pulled top-10 pages from 100,000 results in June 2026 and graded the ones with enough text, about 150,000 pages. The finding is not that Google throws out a page because a model helped write it. It does not. The finding is that heavily generated pages were a smaller share of the places people notice.

Pages that were less than half generated made up 82.2% of positions 1 through 3. Pages that were 80% generated or more were 8.4% of position 1 and 11.7% of position 10. The more heavily generated pages showed up more often at the bottom of the first page than at the top.

Ahrefs also found that pages with little or moderate generated text received two to three times the impressions of pages with a high or very high share. More low-AI pages made it into Google’s index, too: 49.28%, compared with 40.35% of the very-high-AI pages.

That is the line between a draft and the work after the draft. A subscription can give you a page to publish. A professional decides whether that page belongs on the site, puts in the facts the model cannot know, checks what Google and the searcher will actually see, and comes back when the page is not getting the right attention. Getting a page indexed is the beginning. The job is earning a place where the right customer sees it.

One top-three page does not make a marketing plan

In that Ahrefs sample, 5.3% of positions 1 through 3 were entirely generated. A draft from the subscription can rank. The usual result is the other one: the top of the page belongs to work that still reads as a person’s.

Believing the page ranks is not the same as seeing it at the top

Semrush also asked 224 people who work in SEO. 72% said AI content ranks at least as well as human-written content. 70% named speed as the benefit. 19% said the tool improves quality. In the same survey, 87% said a person still leads the writing or writes it outright.

The belief and the position-1 count are in the same report. The people closest to the work think the generated page is keeping up. The pages in the top spot say otherwise. An owner steering a prompt is making that same bet.

The point is not to avoid AI. It is to stop treating a page that is drafted and published as a page that has done its job. A professional uses the tool, then does the judgment-heavy work that gives the page a chance to earn attention.

Google’s instruction is to check the page before it goes up

Google’s own page on generative content says the model does not look up facts. It predicts a likely sequence of words. Because of that, the draft can be wrong. Google’s word for the review is “critical,” and the review happens before you publish.

A polished paragraph is not a verified paragraph

A model writes the next likely sentence. It does not have a checked record of your service list, coverage, prices, or business decisions unless you give it one. If a fact is missing from the material behind the prompt, the model cannot tell the difference between the true answer and a plausible one.

That does not make every draft wrong. It means the draft does not prove itself. The review is where someone checks the words against the company, removes what does not belong, and adds the fact the model left out.

The title, the description, the markup, and the image text show in search too

Google names those fields beside the body. The title, the meta description, the structured data, and the text on an image can all appear in results. A pass that only reads the paragraph leaves the lines Google may actually show.

The same page says a stack of generated pages that add nothing for the reader may break the spam rule on scaled content. A page for every town you might drive to, spun from one prompt in an afternoon, is that stack in a service business. One city page, or a page for the service in that city is the longer version of which pages earn a place.

Your prices, your towns, and the jobs you turn down are the facts it will fill in

Ask a general model for a page and it will complete the page. If you did not hand it the price, it will still write a price. If you did not hand it the towns the trucks actually cover, it will still name towns. If you did not say which jobs you refuse, it will offer those jobs in the same confident voice as the ones you want.

A tune-up written up as free when you charge for it. A city on the page the crew does not run. A job you stopped taking, still offered as if the truck will be there today. The prompt can be specific and the missing fact still gets filled in.

A test of 600 prompts, graded by people, found the same weak spot. Niche questions tripped every model in the set. A trade is that kind of niche. The model has seen a thousand companies that look like yours. It has not stood in your shop.

Human strategy begins before the prompt. It decides which facts matter, which jobs the company wants, what the page needs to say no to, and what has to be checked before it goes live. Those are business decisions, not writing tasks.

Marketers who do this every day still publish the false line

NP Digital surveyed 565 marketers in the United States who already use these tools, and graded 600 prompts across ChatGPT, Claude, Gemini, and three other models.

They run into a bad answer several times a week

47.1% said they hit an inaccuracy several times a week. More than 70% spend one to five hours a week checking what the tool wrote. Only 23% were comfortable publishing with no review.

More than a third have shipped one

36.5% said a hallucinated or wrong line had gone live. Another 39.8% had a close call. When they named where the errors show up day to day, full-page writing was 42.7%, and the markup side, HTML and schema, was 46.2%.

These are people whose job is the marketing. They still miss. An owner doing the same check between calls has less of that week, not more.

The strongest model in the test was fully right on about six prompts in ten

ChatGPT led the test, and it was still wrong often

ChatGPT had the highest share of fully correct answers, 59.7%. Its misses were often a misread of the question rather than a wild invention. A misread on your page is still a wrong page. Six fully right answers in ten means four that were not.

Claude and Gemini landed lower

Claude was fully correct 55.1% of the time. When it missed, it more often left something out than made something up. Gemini was fully correct 51.3% of the time, and omission was its common miss. A page that leaves out the towns you cover, or the jobs you actually do, reads fine and sends the wrong work.

None of the three is a hire. The best score in the test still needs a person who knows the company to say what is missing and what is false.

A professional is paid to make that correction

You know the price, the towns, and the jobs you will not take. That knowledge is yours. It does not move onto the site because you prompted well once. Someone has to put it in, check the lines that will show in search, and change the page when the phone or the rankings say the first version was wrong.

That is the week a professional is paid for. Your week is the jobs. The subscription can draft. It does not come back next month and own the miss.

The fact check, approval, and performance review belong in the marketing process. They should not become a last chore for whoever opened the chat window.

Keep ChatGPT, Claude, or Gemini. Use them to move a first pass along. The finished page, the facts that are true only for your company, and the correction when that page is wrong belong with a person who does this as the work.

An owner guiding the tool is still the amateur on the account. The samples above are why.

If you want someone to take on the strategy, the checking, and the work after the draft, see how we work or schedule a call.

Figures are theirs. Detectors disagree with each other, so the three ranking studies are read as the same shape, not as one combined score.

1. Jose Luis Paredes, Gregory Druck, and Ethan Smith, Graphite, “How Does AI-Generated Content Perform in Search and Answer Engines?” October 14, 2025. graphite.io
2. Margarita Loktionova, Semrush, “Does AI content rank well in search?” April 1, 2026. semrush.com
3. Ryan Law, Ahrefs, “Google Doesn’t Punish AI Content; It Punishes Bad Content,” July 27, 2026. ahrefs.com
4. Google Search Central, “Google Search’s guidance on using generative AI content on your website.” developers.google.com
5. Neil Patel, NP Digital, “AI Hallucinations, Errors, and Accuracy: What the Data Shows,” February 3, 2026, updated April 25, 2026. neilpatel.com