Forty minutes

I spent forty minutes last week proving that a problem didn’t exist.

Forty minutes. On the clock. Paid for by a client who has better things to buy with my time, hunting a missing H1 heading that was never missing, on a page I had open in a browser window while I looked for it.

The report that sent me hunting was nine items long, numbered, and confident. Four of the nine were flatly false. Three were true and completely irrelevant. Two were worth a real conversation.

Nobody wrote that report. No consultant, no competitor, no disgruntled former developer. A business owner typed six words into a chatbot and hit enter.

“How can I improve this website?”

Twelve seconds later he had a list, a knot in his stomach, and a screenshot in my inbox that said “is this stuff true? should I be worried?”

I’m going to answer that honestly. Then I’m going to tell you exactly what we’re doing about it, because we are getting a LOT of these.

Before anything else, he didn’t do anything wrong

There’s a version of this article that makes business owners feel stupid for being curious. I’m not writing that one.

It was 9:40 on a Tuesday night when he ran that scan. Kids were down. He was on the couch with his phone, thinking about his own business instead of scrolling past somebody else’s.

That’s the client I want. Every time.

We are also not an anti-AI agency. Not even close. I build automations in my office at hours I probably shouldn’t admit to. AI drafts, sorts, digs through data, and finds keyword gaps we would have walked right past. It’s in our process daily.

AI didn’t start the fire in digital marketing. Digital marketing was already burning. AI just poured gasoline on it.

So this isn’t “AI bad.” It’s something much narrower, and much more useful to you:

AI is a great idea guy. It’s a terrible building inspector.

An idea guy hands you fifty options so you can pick three. He’s allowed to be wrong most of the time. That’s the job.

An inspector tells you what’s wrong, what’s up to code, and what has to come out. You write a check based on what he says. He is not allowed to be wrong.

Right now, thousands of business owners are handing a clipboard to the idea guy.

The thing nobody tells you about asking AI what’s wrong

Here’s the mechanic underneath all four of these.

When you type “how can I improve this website,” you already told it something. You told it your website needs improving. It is not going to come back and say “honestly, this looks solid, go to bed.”

It’s going to find nine things.

It would find nine things on Amazon. It would find nine things on a site we spent eight months building that’s currently outranking every fence company in the county.

A tool that can only produce problems will always produce problems.

Everything below is a different flavor of that one root.

The Four AI Lies…Well…Where They Are Confidently “Guessing”

You know when your wife asks you how she looks, but that dress isn’t so flattering? Do you lie, or do you soften the blow with a statement like “You’re always beautiful to me, honey!”

AI has a tendency to communicate like this where you really just need a no-bullshit answer – even if that answer is “I don’t know.” This is especially true in matters where you may not really have the experience to tell the difference.

Here are the four “lies” your AI may be telling you:

Lie #1: AI Fluff

“Add more content to your service pages.” “Improve your meta descriptions.” “Consider starting a blog.”

Fluff is advice that shows up with no weight behind it. It isn’t false. It’s unranked. And unranked advice is the same thing as no advice, because you still have no idea what to do Monday morning.

Picture a guy walking past your job site who stops to tell you the gate looks narrow.

Maybe it does. He doesn’t know the HOA spec. He doesn’t know the customer asked for it that way. He doesn’t know about the slope, or the utility easement, or the hour you already spent arguing about it. He’s just a guy with an opinion and no context, and he’s very comfortable sharing it.

That’s Fluff. Because the AI has no idea:

• which pages actually make you money
• what your Google Business Profile is doing
• who you are actually competing with in your city
• why a page is built the way it’s built
• what we already tried last year that flopped

So it hands you the textbook. The textbook isn’t wrong. The textbook is written for everybody, and “everybody” is not a fence company in your city with your crew size, your lead flow, and your service radius.

Advice without priority isn’t strategy. It’s a to-do list written for a business that doesn’t exist.

Lie #2: AI Yes-Man Behavior

This is the sneaky one. If you forget everything else in this article, keep this one.

These models are trained to be helpful and agreeable. Not accurate. Agreeable. Those two overlap almost all the time, and then, right when it actually matters, they don’t.

It shows up two ways.

Version one: you ask “what’s wrong with my site?” and it manufactures a list, because you asked for a list. Nobody wants to be the guy in the room who says nothing’s wrong.

Version two, and this is the bad one: you push back. You say “I don’t think that’s right, our developer told me that’s fine.” And it folds. Instantly. “You’re absolutely right, I apologize for the confusion…” and then it builds a whole new answer on top of your correction and defends that one just as confidently.

Sit with what that means for a second.

Whichever way you lean, it leans with you.

Every trade association has this guy. Ask him about vinyl, he loves vinyl. Ask him about ornamental aluminum, always been an aluminum guy. Bring up a supplier you’re mad at and he’s been saying the same thing for years. He’s pleasant. He’s easy to talk to. I’d buy him a drink.

I would not let him bid a job.

A tool that agrees with you isn’t confirming you’re right. It’s confirming you asked.

And it never tells you it’s doing this. There’s no confidence score. No footnote saying “by the way, I just reversed my entire position because you sounded skeptical.”

Lie #3: AI Pre-Render Fatigue

Now the technical one. Stay with me, because this is the flavor that produces the scariest looking reports, and it’s the one that wastes the most of everybody’s time.

There is not one version of your web page. There are three.

1. The raw code. What the server hands over before anything runs. This is lumber on the trailer.
2. The rendered page. What exists after the browser runs the JavaScript, loads the fonts, pulls in the review widget, drops in the schema. This is the finished fence.
3. What Google actually indexed. Google’s own copy, on Google’s own schedule, which can lag behind both of the above.

Most AI tools only ever see number one. They grab the raw code and never run the JavaScript. So on a modern website where a lot of the page gets assembled in the browser, the AI is standing in front of a trailer full of lumber, writing a report about your fence.

That report comes back and says:

• “This page has almost no content.”
• “Missing meta description.”
• “No structured data or schema markup found.”
• “Your H1 heading is missing.”

Then you pull the page up in your browser. It’s all there. Content, headings, reviews, the whole thing.

Here’s the frustrating truth: you’re both right. You are looking at two different versions of the same page.

Some tools do the exact opposite, by the way. They render everything beautifully in a headless browser and hand you a clean bill of health on things Google’s first pass never sees. Same disease, opposite symptom.

The report isn’t lying to you. It’s describing a version of your page that no customer will ever see.

I call it fatigue on purpose. The false positives are annoying, but the fatigue is what actually costs you. Chase four ghosts and you stop trusting the alarm. Then a real problem walks in, and it looks exactly like the last four.

Lie #4: AI Imposter

I saved the worst one for last. This is the one I think about.

A lot of general chat AI cannot actually visit your website. Depending on the tool, the mode, your firewall, your security settings, and honestly just the day, the fetch fails. Or gets blocked. Or times out. Or comes back with a cookie banner and nothing else.

Here’s the problem. It very often does not tell you that happened.

It writes the audit anyway.

From your web address. From the words in your page slug. From what it knows about fence companies in general. From a cached copy of your site from two redesigns ago.

And it reads exactly like a real audit. Same tone. Same numbering. Same certainty.

My dad was a commercial fisherman. He’d bring in a catch, and the buyer would weigh it on the buyer’s scale, write a confident number on a clipboard, and hand him a check.

Dad didn’t have a scale of his own on that dock.

So the number on the clipboard was the number. Because it was written down. Because the man holding the clipboard sounded certain. And because there was no way to check it from where my dad was standing.

That is exactly what an AI Imposter report is. Somebody else’s scale. Somebody else’s clipboard. A number that sounds certain, handed to a man with no way to verify it.

I’m not going to pretend that’s just a technical quirk to me. Watching what confident numbers can do to an honest man who has no tools to argue with them is a big part of why this company exists at all.

So no, I’m not going to wave off a bad audit and tell you not to worry about it.

I’m going to hand you the scale.

A FAKE audit looks identical to a REAL one. That’s what makes it the dangerous one.

What this actually costs you

I’m not being picky about accuracy for sport. Ghost problems cost you money in three separate ways.

1. Budget burned on nothing. Every hour we spend disproving a hallucination is an hour not spent on something that would have put a job on your calendar. You paid for both hours. Only one of them was worth anything.

2. Good work torn out. This is the one that stings. A page that was built deliberately, ranking well, converting well, gets rewritten because a report called it thin. Then it stops ranking. I have watched this happen to businesses that were doing fine.

3. Trust drift. Nine “problems” a month, every month, and eventually you’re lying awake wondering if anybody’s minding the store. You shouldn’t have to live like that. I would much rather build a process than let a good relationship quietly rot from the inside.

So here’s our standard, in writing

We are not going to tell you to stop running AI on your website. I’d be a hypocrite, and frankly I’d rather you stay curious about your own business.

But we’re also not going to keep triaging forwarded screenshots one at a time, because that’s how good agencies bleed out slowly and stop doing the work that actually grows your company.

Three parts. This is now how every AI finding gets handled.

Part 1: One door, three attachments

Any AI audit, scan, or finding comes to us in one place. Your account channel, not a text message to whoever you talked to last week.

And it comes with three things:

• The tool. Which AI, and which mode or model if you know it.
• The exact prompt. Copied and pasted, word for word. Not summarized.
• The URL you handed it.

That’s the whole requirement. Three things.

Now, why the prompt? Because it looks like busywork and it absolutely is not.

The prompt IS the diagnosis.

“How can I improve this website” and “audit this page and tell me whether it meets Google’s helpful content guidelines” will produce completely different reports off the exact same page. Without seeing the prompt, I cannot tell you whether the tool found something real or whether you accidentally asked it to invent something.

Nine times out of ten, I can tell you which of these four issues you’re looking at from the prompt and the tool name alone. Before I ever open your site.

No prompt, no review. Not because we’re gatekeeping. Because a review without the prompt is a guess, and I’m not going to bill you for a guess.

Part 2: Every finding gets one of three labels, and we show our work

Verified. We’re on it. Real, reproducible, and it matters. It goes into the queue with a priority and a timeline. And when this happens, we’re going to thank you out loud, because sometimes you are going to catch something we missed and I want you to keep looking.

Valid, but not now. Technically true, low impact. We’ll tell you what it’s competing with for hours so you can overrule us if you want to. It’s your money and your business. You just get to see the trade before you make it.

Not accurate, and here’s the receipt. We show you what the tool actually saw. Usually that’s the raw source next to the rendered page, side by side, or a fetch log showing the thing never reached your website at all.

That third one is the entire reason I wrote this article.

You should never be stuck choosing between trusting the robot and trusting your agency. That’s a garbage position to put a client in. You should get to look at the evidence yourself and decide.

Part 3: We’re getting in front of it

Here’s the honest part. A lot of these late-night audits happen because you wanted to know what was going on with your website, and it was faster to ask a chatbot than to ask us.

That’s on us. Not on you.

So on your reporting cadence, you’re going to get a plain-language note covering what we checked, what we fixed, and what we deliberately chose NOT to do, with the reason attached.

That last section is the one that matters. Most AI findings that get sent to us are things we already looked at and passed on for a specific reason. You just never got to see the reason. That’s a communication failure on our end, and it’s a fixable one.

How to actually get something useful out of AI on your own website

I’m not going to hand you a list of restrictions and call it a day. So let’s talk about doing this well, because there is real value sitting here.

Stop asking AI to be the inspector. Start using it as the idea guy. It’s genuinely excellent at that.

Prompts that will actually help you:

• “Here’s the text from my gate repair page. Pretend you’re a homeowner with a sagging gate and a $600 budget. What questions do you still have after reading this?”
• “Give me 20 questions a first-time fence buyer asks before they call anybody.”
• “Read this paragraph and tell me where it gets confusing.”
• “What in here would make you trust this company more, and what would make you close the tab?”

Notice the pattern. You’re asking it to be a reader.

Not a judge.

Two rules that will save you most of the grief:

Rule one: paste the actual words in. Don’t hand it a link and hope for the best. Copy the text off your own page and paste it into the chat. That single habit kills AI Imposter outright and takes out most of Pre-Render Fatigue with it, because now the tool is definitely looking at the same thing you’re looking at.

Rule two: ask open, not loaded. “What’s wrong with this page?” gets you problems whether or not any exist. “What is this page doing well, what is it doing poorly, and is any of it actually worth changing?” gives it room to tell you it’s fine.

And one bonus test. Push back on it once, on purpose, about something you know it got right. If it flips and apologizes, you just learned it wasn’t sure the first time either. That’s a five-second check that tells you how much weight the whole conversation deserves.

One more thing

Most agencies would read this article and see a problem. Clients getting smarter. Clients asking harder questions. Clients showing up with reports and wanting answers.

I see it the other way around.

I would rather have a client pasting his own website into a chatbot at 9:40 on a Tuesday night than a client who hasn’t thought about his website since the day we launched it. One of those businesses is going somewhere.

The goal was never for us to be the only ones in the room who understand your website. The goal is a foundation solid enough that you can poke it, question it, and run it through whatever new tool shows up next year, and it holds.

Relationships first. Then the foundation. Then growth. In that order, every time.

So keep sending them. Send me every single one.

Just send the prompt with it.

I’ll bring the scale.