Readers can tell. Here is what they do about it

The usual version of this argument is about detection. Can people tell? Can Google tell? Is there a tool that tells? I think that framing misses what actually happens, because a reader is not running a test. They are deciding whether to keep reading, and they make that call fast and mostly without articulating why.

So the useful question is not "can they tell". It is what they do in the four or five seconds after something feels off, and where that shows up in your numbers.

Nobody thinks the words "this is AI"

When I sit with someone while they compare three businesses, they never say a page is generated. What they say is vaguer and more damning. It sounds like an ad. It does not really say anything. It could be anyone. Sometimes they just scroll past without comment, which is worse, because there is no objection to answer.

That is the actual mechanism. Generic copy does not trigger a judgement about how it was produced. It triggers a judgement about whether there is a real business with real opinions behind it. And that judgement gets made before anybody has consciously decided anything.

Which is why the detection framing leads people astray. You can pass every AI detector on the internet and still lose the reader, because the reader is not testing for provenance. They are testing for evidence of a person.

They are not thinking "a machine wrote this". They are thinking "there is nobody in here", and then they are gone.

What they actually do

In rough order of how often I see it:

• They skim to the bottom looking for a price, a name, or a photo of an actual human, and if none of those are there they leave.
• They open a competitor in another tab and compare. This is the expensive one, because you have now paid to introduce someone to the alternative.
• They go to your Google reviews instead, because reviews are the one thing on the internet a business cannot write about itself. If the site says nothing, the reviews become the whole decision.
• They ring you anyway, but with less trust than they would have had, which shows up as more price pressure and more questions you thought your website had answered.

Only the last of those is visible in your analytics as anything other than a bounce, and none of them look like a ranking problem. That is why this cost is so easy to miss: the traffic is fine, the positions are fine, and the phone is quieter than it should be.

Why the damage lands on enquiries, not rankings

I want to be careful here, because there is a lot of confident nonsense about AI content and rankings. As far as I can tell from my own client work, well-structured generated content does not get penalised for being generated. Pages that go backwards go backwards for structural reasons - two pages fighting over the same search, invented claims, no internal links - and those are covered in the honest guide to using AI for your website content.

The generic-voice problem is different, and it is not a Google problem at all. It is a conversion problem. You are not penalised. You are skipped. And being skipped does not appear in any rank tracker, which is precisely why it goes unaddressed for months.

There is a version of this in my own study data that I find clarifying. Across 480 grids of Google Maps results I measured whether star rating correlates with where a business appears, and it is a clean null - rating does not move position. But nobody sensible concludes ratings do not matter. They matter enormously for whether the person who sees you rings you. Voice works the same way: invisible to the ranking, decisive at the point of choosing.

The specific tells, from the reader's side

The tells worth knowing are the ones a customer notices, not the ones a linguist would list. I have written up the pattern itself in why ChatGPT content sounds like ChatGPT; this is what it feels like to a person who is not thinking about AI at all.

• Every paragraph is the same length. Real writing has a one-line paragraph in it somewhere, because the writer got emphatic.
• Nothing is refused. A real business says "we do not do that" about something. Generated copy is relentlessly accommodating, and accommodating reads as unspecialised.
• No numbers with edges. "Years of experience" instead of "since 2011". "Competitive pricing" instead of "callout is $95 and comes off the job".
• No opinion anybody could disagree with. If nothing on the page could annoy a competitor, it probably says nothing.
• The order is encyclopaedic rather than persuasive - "what it is", "benefits", "why choose us" - rather than starting with the thing the customer is actually worried about.

A test you can run this afternoon

Detection tools are not the way to check this, partly because they are unreliable and mostly because they answer the wrong question. Here are three checks that answer the right one.

1. The competitor swap. Take your main service page, replace your business name with a competitor's, and read it. If it is still true, it was never about you. This is the single most useful five minutes you can spend on your website.
2. The stranger test. Send a page to someone who does not work in your industry and ask one question: what does this business do differently from the others? If they cannot answer, neither can your customer.
3. The specifics count. Print the page and circle every fact a competitor could not honestly copy - a real number, a named suburb, a policy, a job you did. On a good page there are ten. On a generated page there are usually none.

The third one is the most diagnostic because it is countable. It also tells you exactly what to fix, which the other two do not.

The fix is smaller than people expect

You do not have to rewrite the site by hand. Almost all the recoverable value is in the first two paragraphs of your most important pages, because that is where the decision gets made, and in the specifics scattered through the rest.

What works is giving the model material only you have - the way you explain the tricky thing, the objection you answer on every job, your actual position on the thing your industry argues about - and then rewriting the opening yourself. The mechanics of that are in how to actually give AI your brand voice, and the wider structure is in using AI for your website content without wrecking it.

How strong is the evidence, honestly

I should be straight about what kind of claim this is, because there is a lot of assertion in this area dressed up as fact.

"Readers can reliably identify generated text" is not something I have measured and I would not state it. The studies that exist are mostly on short samples in a lab, where somebody is being asked to look for it - which is the opposite of a customer comparing three plumbers at nine at night.

What I have is weaker and, I think, more relevant. I have watched people compare businesses, I have read a lot of pages on sites that went quiet, and the pattern is consistent: nobody names it, and the vaguer the copy, the faster they move on. That is an observation across a few dozen cases, not a measurement, and you should weigh it as such.

The reason I am comfortable acting on it anyway is that the fix costs almost nothing and helps regardless. Rewriting your opening two paragraphs to sound like you is worth doing whether or not anyone could have identified the alternative as generated. There is no version of this where being more specific about your own business makes you worse off, which is a rare and useful property in marketing advice.

I would rather you published ten pages that sound like a person than fifty that sound like everybody. Not because the fifty get penalised. Because nobody finishes reading them.