The H1 on the homepage said "Home". That was the whole H1. I replaced it with a line saying what the school is, where it is and what it teaches, and the site went from page 3 to position 1. The typing took 4 seconds.
Nobody pays for the 4 seconds. What gets billed is knowing it was that word and not the 300 other changes you could make on that site, all defensible, all plausible if you ask an LLM to list them.
So when someone tells you, pleased with himself, that his son-who-knows-computers redid his whole SEO with ChatGPT, that the site is live and it looks great... what do you say?
Because anything you say, coming from you, is going to sound like a guy defending his margin. And he would not be wrong to think it.
What That Number Is Not
The site belonged to a small alternative school. 40 kids, a waiting list built entirely on word of mouth, and a homepage untouched since launch. They wanted to come up when parents in the area went looking for that kind of schooling, and they were sitting on page 3.
Before anyone runs off to rewrite an H1, the query did most of the work here. Local intent, low volume, semi-branded, maybe 60 people a month typing it in that region. Not a competitive term. On a competitive term that same edit does nothing at all, and anyone telling you otherwise is selling you something.
What made it work is that on the page meant to rank, the strongest on-page signal about what the page is said nothing except that you had arrived somewhere.
Position 1 landed in Search Console 11 days later, on the next real recrawl. It was not instant and it was not magic, just a page that finally said what it was, on a query with almost no competition.
Nobody Pays for the 4 Seconds
The client saw me type for 4 seconds. He did not see the 2 hours before it, and there is no honest way to make him see them, because what happened in those 2 hours was mostly elimination.
That site had around 300 defensible changes available. The images were unoptimized, the internal linking was thin, half the pages had no meta description, the contact page was a PDF, the blog had been dead since 2019, and each of those is a real finding that a real audit would list and that any model will happily list for you in 8 seconds flat. All of them would have improved something. Exactly 1 of them was the reason the site sat on page 3, and the only way to know which one was to look at what the site was already ranking for, on which queries, at which positions, and to notice that the homepage was pulling impressions on the school's own name and on nothing else. That is not intelligence. That is data the model does not have and cannot guess, and the entire job is standing in front of 300 correct answers and picking the 1 that moves.
What gets billed is the 299 things I did not touch.
Anyone who has raided knows a geared character is not winning because he hits harder. He wins because he knows which mob to pull and which one to leave alone, and pulling the wrong one wipes the group no matter how clean your rotation is.
And this is where I have been stuck for years. Everything I just wrote is true, and it is also exactly what a guy protecting his invoice would write. There is no version of this argument that does not benefit me. The suspicion is structural, it is not unfair, and pretending otherwise would be the most vendor thing I could possibly do.
Which is why I stopped arguing this one and started telling people about a shower.
The Shower With No Slope
A friend of mine had her bathroom redone by her brother-in-law, an engineer. The tiling was clean, the shower niche was square, the grout was perfect. I saw the photos. It was better work than the last 2 quotes she had turned down.
The floor had no slope. A shower base has to fall toward the drain, roughly 1 cm per meter, and that number is free if you know it and invisible if you do not. Nothing about a flat floor looks wrong while the tiles are dry.
8 months later the standing water had gone through the partition wall into the hallway. The repair quote came in above what the tiler she avoided would have charged. He did good work. The problem only announced itself 8 months after the compliments had stopped.
What a ChatGPT-Built Site Actually Looks Like
Small nail salon. Call it Gail's place. The site got redone last spring by her son-who-knows-computers, over 2 evenings and a lot of prompting. It loads fast, the photos are nice, the booking button works. On the screen it is a better site than what she had.
Then open it the way a crawler does.
- 4 pages competing for the same query, so Google picks 1 and mostly ignores the others
- the same title tag and the same meta description on every single route, generated once and pasted everywhere
- copy that describes a nail salon in general and never describes hers, no street, no neighborhood, no name of the woman who has been doing this for 19 years
- an H2 sitting above the H1 on 2 pages, because the template put the section header first and the order never got checked
- 6 pages that have never been indexed, which takes about 40 seconds to verify and which has never been verified
All of them are invisible on the rendered page. The site works on her machine, it works on yours, it works on her son's, and that is precisely the problem. Nothing in the browser is going to tell you that half the site does not exist as far as search is concerned.
If you are reading this with a slightly sick feeling because your own site got the same free favor last year, that is the correct reaction. Go look. I will wait, the rest of the article is not going anywhere.
If your site is an app rather than a brochure, the failure surface changes and the invisibility does not. I ran the same audit on 11 vibe-coded apps and the pattern held, builder dashboard reporting a 90 SEO score while the crawler got an empty div.
Graphite, which is an SEO agency and therefore has a direct material interest in the answer landing this way, ran 55,400 Common Crawl URLs through an average of 3 AI detectors instead of the usual 1. The share of mostly AI-generated articles has been flat near 50% for 5 quarters. Their earlier pass found 86% of articles ranking on Google were human-written, and that AI articles which do surface tend to sit lower. Axios covered the same study independently and pointed out the detector's 4.2% false positive rate, plus the fact that paywalled sites block Common Crawl, which would undercount human writing.
There is also an EBU and BBC study where professional journalists graded more than 3,000 assistant answers, and 45% carried at least 1 significant problem. That study covers news questions, on the free consumer versions of ChatGPT, Copilot, Gemini and Perplexity. It says nothing about SEO and I am not going to pretend it does. I bring it up for what it measures, which is how confidently these tools answer questions they have no ground truth for. Your rankings are a question they have no ground truth for.
All of that is aggregate web data. It proves nothing whatsoever about Gail's site, and nothing about yours. It is a base rate, and base rates are what people reach for when they cannot measure the actual case.
The Cousin Who's Good With Cars
Last year I found a Mercedes at a consignment lot. Good car, good price, the kind of listing you screenshot before it disappears. Delivery promised in 3 days.
3 days went by. Nothing.
I called. There was a reason. A week in, there was another reason, and the reason had gotten vaguer. At day 15 I stopped being polite and got the actual answer, which was that the car had never been at the lot. For 15 days that Mercedes had been in 2 places at once, which is a fine property for a particle and a bad one for a car I had already paid a deposit on. It was at the cousin-who's-good-with-cars, who was sorting out a fault on the dashboard display. It would all be fine.
A week after that, I learned the cousin-who's-good-with-cars had ordered the wrong part.
I cancelled.
He Will Never Find Out On His Own
The dealer lost a confirmed sale on a car he had already sold. He never learned that. From where he sat, a buyer went quiet and the car eventually moved to someone else, a few weeks late, for a bit less. There is no line on any report that reads "lost to the cousin", and the cost never landed anywhere he could see it.
That is the shape of the whole problem. The son-who-knows-computers will never find out either, because Gail has 1 site and 1 timeline, and she cannot load an earlier save and run the year again with a real title tag on every page to see what the difference would have been. She sees bookings or she sees no bookings, and either way she puts it down to the season, the neighborhood, her prices. There is no invoice for what did not happen.
The gap between the work and the verdict runs long, the same way it ran 8 months on that shower floor. By the time the effect is legible the cause has stopped being interesting.
I think that delay is the actual mechanism here, more than the quality of the output. Could be I am over-reading 3 stories that happen to line up.
Put the 2 walls together. An argument from me does not clear the first, because I am selling. It does not clear the second either, because there is nothing to point at. Better arguments do not help. I have made better arguments and watched them do nothing.
So stop bringing arguments.
So What Do You Actually Say
You say something that costs him nothing to check and costs you nothing to be wrong about.
"I'm not going to argue with you. Open Search Console and tell me how many of those pages are indexed."
Or, if that lands too hard, either of these:
- "Nice. What does it say under Pages, indexed versus not indexed?"
- "How many impressions did the booking page get last month?"
Notice what those do. They make no claim about the site being bad. They move the conversation onto a screen that neither of you controls, where the numbers came from Google rather than from anyone's opinion. He can turn out to be right. That is the part that makes it sayable out loud.
Then the test. 5 checks, binary verdicts, free, all inside the Search Console property that should already exist and which is itself check number 0 if it does not.
- Pages crawled. Is the count in the same neighborhood as the number of pages on the site.
- Pages indexed. Same question, and this is where most of these sites fall over, because the sitemap was never submitted or the template shipped a noindex.
- At least 1 impression in the last 28 days, per page. A page with 0 impressions is not competing badly, it is absent.
- Several pages surfacing on the same query. That is cannibalization and the Queries view shows it in about a minute.
- Anything sitting at position 4 to 15. That band is where a change actually moves money, and it is the only band worth touching at all.
There is no boss fight here. It is a panel with numbers on it, and it either says the site is legible or it says it is not.
The page indexing report has looked the same since roughly 2018 and I have never once gotten a clean export out of it that survives a paste into anything else. You click the number, you get a list, you scroll the list. I gave up on that fight around the time I gave up on the URL inspection quota, which is 10 a day and has been 10 a day forever.
Some people will run the test, find 4 red lights, and still not hand $1,000 to $1,500 to an agency for an initial engagement. Fair enough.
What the model is missing is not intelligence, it is your site's data. Plug that in and most of the guessing disappears. ChatSEO connects an agent to Search Console, Analytics and the CMS, and ships an MCP so you can drive the whole thing from Claude instead of a dashboard. I went into why a model without your data invents keyword volumes already, which is the technical half of this argument and the reason I am not re-explaining it here.
And now the awkward part, which I would rather say myself than have you catch. That link is an affiliate link. I have spent this entire article establishing that an argument from an interested party does not clear the wall, and I just became the interested party. So do not take the recommendation. Take the test above, which involves no link, no signup and no money, and if the 5 checks come back clean, close this tab and go do something better with your evening.
This only closes the mechanical half anyway.
What This Still Doesn't Replace
Connecting the data kills the guessing. It does not kill the judgment call, and the judgment call was always the part worth money.
The test tells you whether the site is legible. It does not tell you which of the 5 correct opportunities in front of you deserves a quarter of your attention, or that the 47 impressions on a page are noise while the 12 on another are a market opening, or that the page sitting at position 3 should be left completely alone even though every tool in existence will suggest 6 improvements for it. Doing nothing on purpose does not come out of a dashboard.
So the original question closes, in a direction I did not expect. Diagnosis got cheap. The part I used to charge the most for, the sitting down and looking, now takes an evening and a connection. What is left is the arbitration, and I still have no idea how to put a price on that in front of somebody who just watched the rest happen for free.
So next time someone tells you his son did the SEO in an evening with ChatGPT, promise me you will not tell him about the H1, or the shower, or the Mercedes. None of it lands. He hears a guy defending his margin before you finish the sentence.
Just ask him to open Search Console with you. 5 minutes, his laptop, his numbers. Either the pages are in there or they are not, and it stops being about you.
And if it turns out you like doing that, there is a whole street of small businesses around you sitting on sites that have never been crawled once 🤓
Sources
- Graphite, AI Content in Search & LLMs: https://graphite.io/five-percent/ai-content-in-search-and-llms
- Graphite, AI Now Writes as Many Online Articles as Humans: https://graphite.io/five-percent/research/ai-now-writes-as-many-online-articles-as-humans-do
- Axios, AI-written web pages haven't overwhelmed human-authored content: https://www.axios.com/2025/10/14/ai-generated-writing-humans
- EBU / BBC, AI assistants misrepresent news content 45% of the time: https://www.ebu.ch/news/2025/10/ai-s-systemic-distortion-of-news-is-consistent-across-languages-and-territories-international-study-by-public-service-broadcaste
- CBC / Reuters, Top AI assistants misrepresent news content: https://www.cbc.ca/news/world/ai-assistants-news-misrepresented-study-9.6947735
This post may contain affiliate links. If you click them, I might earn a small commission (costs you nothing, and helps me keep shipping quality articles every day for your reading pleasure).
Top comments (0)