This is already happening: AI content is marked and it is losing.

Everyone is publishing more than ever before. Production on almost every channel we manage has roughly doubled in the last two years, websites, marketing copies and graphic design. But it hasn’t scaled reach, nor have engagement or conversions.

This gap is the story. There is now a lot more content demanding the same limited attention, and platforms and models have started to sort it out. Not for quality, which would be difficult. By origin, which turns out to be easy.

We’ve been saying this was going to happen for a while. We no longer plan for it. The marking has been shipped, the rigs are labeling and our own figures show what it costs.

The platforms are already sorting

We manage marketing for a handful of companies on LinkedIn. When we took over these accounts, the pattern was consistent: posts generated with AI performed worse than posts written by a person. We didn’t want to believe this on other people’s accounts, so we ran the same test on our own profile. Same result.

LinkedIn now allows users to report AI errors directly in the feed via the AI SLop botton. Reddit has updated its community guidelines to take action against primarily AI-generated posts. Facebook labels AI content, and in our experience, a labeled post performs worse than the same post written by a person.

YouTube went further and put it in writing. Creators are required to disclose when AI has been used to generate or significantly modify realistic content. Labels look on the player or in the description. And this is the line worth reading twice: If a creator doesn’t disclose it but YouTube’s systems detect it, the platform automatically applies the label and creators can’t remove it if there’s a high degree of confidence that the content was created with AI.

Worse still, in all of this, it means measurable things. Fewer views. Less dwell time. Less user engement and fewer conversions.

Four platforms moving in the same direction, in the same window.

And now the models are marking it too

This is the part that most people haven’t caught up on.

Anthropic has confirmed on its own support pages that Claude models launched from August 2, 2026 carry machine-readable marking. The text has a built-in watermark that works by subtly skewing the words chosen by the model, so that a pattern becomes detectable on enough content. Files and images contain provenance metadata signed according to the C2PA standard.

This applies globally. There is no way to get rid of it, neither for you nor for the person paying for the subscription. Claude says it doesn’t affect anything, but curiously it refuses to remove it; if you prompt Gemini, it will tell you the same thing; it too watermarks its text.

The driving force is Article 50 of the EU AI Law, whose transparency obligations became applicable on August 2, 2026. The European Commission’s Code of Practice on the Transparency of AI-Generated Content is the framework built to demonstrate compliance with these obligations, and by the end of July 2026, around 190 companies and organizations had signed it.

Read what this code actually covers, because the second half is for people like us. Section 1 imposes obligations on providers, requiring that AI output on audio, image, video and text be marked in a machine-readable format and detectable as artificially generated. Section 2 imposes obligations on deployers, requiring disclosure of deepfakes and AI-generated text posts that inform the public on matters of public interest, unless that post has been subject to human review and falls under editorial responsibility.

Human review and editorial responsibility. This exemption represents a significant amount of work and describes exactly the workflow we have been advocating for the past two years.

Also note what this means for the design side. Provenance metadata travels with the file. A generated hero image arrives at your web designer already with a trace of its origin. This is not a writing problem; it’s a pattern problem, an averages problem. It extends from text to image to the design of each element on the page.

Why are they spending real money on this?

Compliance is the reason given. This is not the only reason.

Each of these signers has a model to train, and each of them has the same problem: the open web fills up with synthetic text, and models trained on thier own output text deteriorate. Model degradation is costly and very difficult to reverse. A reliable signal that says “a machine produced this” is exactly the filter you would create to keep your training data clean.

If you think about this, it makes the direction obvious. People don’t want to consume AI content. The AI does not want to be trained on the AI content. Everyone now has their own AI product to protect. Reddit, LinkedIn, and YouTube no longer just block scrapers; they manage slops within their own platforms, because an average content feed is a slowly dying product.

Google has hinted at where it stands with EEAT. Evidence in our own data suggests it is already a factor in ranking and citatoin rates.

What the top of the SERP actually looks like

We tested the theory at the time, rather than trusting it.

We selected the top ten results for service industry queries in markets where SEO is truly competitive and everyone knows what they’re doing, including New York and London. We ran these top-ranking pages through AI checkers. Seven of the ten were returned as human writing.

That’s not to say AI has never touched them. That’s probably the case. This means that whatever the AI produced was then heavily edited by a person before it went live, to the point where the machine’s fingerprints disappeared.

It should be noted that AI detectors are not reliable instruments. They produce false positives and false negatives and we would not make a business decision on just one result. The signal here is consistency of trend across ten competitive results, not the accuracy of a single score.

The pattern aligns with everything else on this page. At the top of tough markets, humans are still in charge.

E-E-A-T is reason, and AI can’t fake it

Google has made it clear for years that it wants experience, expertise, authority and trust. We’ve written about this many times, because it’s the framework that explains almost everything else.

A language model alone cannot produce true E-E-A-T. It produces the average of what he has already seen. Experience means having actually done the thing. Expertise means knowing which of the plausible answers is the right one in this specific case. Neither is available for a system whose function is to predict the most likely next word.

Give it some real industry experince and it will become a useful writing tool. Ask it to provide the experience himself and you’ll get exactly what everyone else in your industry is publishing, worded slightly differently.

That’s the whole mechanism. This is not a conspiracy against AI. Just an average, applied to a ranking system explicitly designed to reward the opposite.

What our own numbers say

We stopped guessing and measured it. Four separate tests, all at sites we control or manage.

Citation rate here means how often content is retrieved and cited by AI search tools, measured with Semrush’s AI Visibility tool, which shows which prompts triggers AI citations.

A warning about the 103 pre-2023 positions. These pages are older, and the older pages contain backlinks and authority that the newer pages haven’t had time to earn. Part of this 72% is linked to age and not to paternity. This is exactly why the second test is more important: 57 recently written articles, competing on an equal footing with AI content, still came out with a 63% lead.

A/B pairs are the ones that should worry anyone running a content program, because dwell time, bounces, and conversion actions per session have nothing to do with domain authority. Same site. Same traffic sources. Same three month window. The only variable was how it was written.

The indexing problem that no one warns you about

Our hosting team noticed something distinct, and it matches.

When customers generate a full website with AI and put it online, some of these sites are not indexed by Google at all.

Google has never published an official indexing schedule, there are many factors that go into this. What we know from doing this for a decade is that content that Google likes tends to get indexed in about two weeks, sometimes even faster. On the AI-generated sites we host and where indexing took place, the average was closer to two months.

Not all. But it’s enough for us to raise it now before a customer goes down this path.

If you’re wondering why the site you built in an afternoon isn’t producing queries, indexing is the first place to look. The rest of the reasons are based on our AI entity and digital authority optimization page .

We asked ten people why they stopped reading

The numbers tell you what happened. They don’t tell you why. So we interviewed ten people and deliberately chose a split: five of our own clients, two members of our internal team, and three members of the public with no technical background. Developers and non-developers, customers and business people.

Ten is a small sample and we are not going to present it as research. It was the consistency of the responses that was worth noting.

• Seven in ten people said they don’t avoid content because it’s AI. They avoid it because it’s generic and unoriginal.
• Three of them said that as soon as they spotted a tell, a phrase like “in conclusion” or an obviously generated image, they wondered if they could trust what they were reading and left.
• Six said it made them question the brand behind the content, not just the article.

This first conclusion runs counter to the popular version of this argument. People don’t carry out ideological control. They are bored. The rejection is not about technology. It’s average.

This is a much harder problem to solve, because averaging is what the tool is designed to do.

Seven customers. Seven industries. A website.

Here’s the part that makes the average argument impossible to dismiss, because you can see it rather than measure it.

Lately, clients have started sending us AI-generated mockups of what they want their site to look like. No briefs. Finished drawings, made in an afternoon, arriving in image form. We now have seven, from seven different companies, in seven unrelated sectors.

They look alike.

Not the same in the sense that two good sites in the same sector are similar. The same. The same dark background with an accent color running through everything. Same gradient on an industrial stock photo. Same title centered in capitals with a small decorative ruler on each side. Same row of four icons with two words below each. Same three-column card grid, then another three-column card grid, then a stat strip, then another card grid.

An agricultural produce retailer and a LEGO collectibles store came to the same layout. This should not be possible. This happens because the tool does exactly what it was designed to do: produce the average of every website it has ever seen, and the average of every website is a dark hero, a row of icons, and three cards.

Now look at them as a designer would, or for that matter as a client would.

There is no hierarchy. Each strip on the page has the same visual weight, meaning nothing is emphasized and the eye has nowhere to rest. There are six or seven calls to action of equal importance, so there is no main action. The body text is way too small compared to a huge section filler. Each section is symmetrical, so nothing stands out by contrast. And the entire page tries to say everything the company does, in one fell swoop, above and below the fold, without any editorial decisions being made about what actually matters.

This looks like a poster you made in sixth grade. You had one sheet, you had to fit the whole project on it, so you jammed everything in and drew borders around the spaces.

It’s not a style problem that you solve with a better color palette. This is what happens when no one has made a decision. Design is a series of decisions about what to leave out and what the visitor should do next. A template with on average thousands of layouts has no idea what your particular client should do next because it doesn’t know your client. So it includes everything, weighs everything equally, and hands you a page that converts no one.

The point that matters for this article: it is the same failure as the copy. Text, images, layout, it all comes from the same averaging process, it all comes to the same place. When we say that the problem of averaging extends beyond writing to design, that’s what we mean, and we have seven pictures of it.

Same skeleton, better manners

Not all generated designs arrive out of order. Some seem perfectly professional, and these are more dangerous, because there is nothing obvious to oppose.

We were recently sent to a plumbing job which was tidy. Sensible order, clear title, one primary and one secondary button, appropriate emphasis break in the middle. Scroll through it quickly and you would call it a decent local service website.

Then place it next to the others and count the sections. Hero with a split image. Band of four icons. About the block with three ticks. Card grid. Division into two columns. Group call to action. Grid why choose in six elements. Three-step process. Area section. Two-column problem list. Three testimonies. Accordion FAQ. Three blog cards. Call to action at the bottom of the page.

It’s the same skeleton as the commodities trader and the LEGO store. Different color, different industry, more padding, better manners. Identical bones.

Which is the most useful example, because you can’t consider this one as poor execution. The model survives by being well executed. This is what the average looks like when it’s sharp.

Then open it on a phone

Each of these decisions is a desktop decision, and no one generating them is checking the other view, because they are looking at the image of a desktop page.

Stack it reactively and the strip of four icons becomes four rows of a small icon and two words, most of a screen saying nothing. The eight-card service grid becomes eight full-width cards in a column. The grid of six elements why choose stacks into six others. The three-step process, three more. The problem list has two columns, six additional rows. Then three testimonials, then the FAQ, then three blog posts. The visitor scrolls through the better part of twenty nearly identical blue and white blocks to reach the end of a page they never asked to read.

The images aggravate the situation rather than prolong it. This page contains a generated hero, a generated team photo, several generated photos of a plumber at work, and a decorative card that means nothing at 390 pixels wide. Everything still needs to be loaded. On a phone using mobile data, it’s your biggest content painting that’s gone, on the exact device the customer is holding.

And think about who that customer is. Someone with a burst pipe, standing in the water, on the phone, just wanting one thing. The page offers them a call button at the header, a hero button, an emergency button, a strip call button at mid-height and two others at the footer. Six ways to do the one thing they came for, none of them dominant, spread out on a scroll that takes a minute to go through.

A designer makes this page shorter on mobile, not longer. Reduces the service grid to the four that actually sell, drops the decorative card, keeps a sticky call button, and kills the other five. These are decisions about a specific customer at a specific time. The model has no view of all this, because it does not know who is in the water.

And then there is the part that is simply made up

The testimonials on this page are invented. Three quotes, three plausible names, three job descriptions, no actual clients. The star rating cites a number of satisfied customers that no one has counted. The team photograph is generated, so the trust signal is an image of people who don’t exist. The phone number in the emergency banner and footer is a placeholder that would have been posted online. One of the service area buttons says “Your Area.”

That’s the E-E-A-T argument in one page. The model knows that a converting local service site has testimonials, a rating, a team photo, and details, so it produces the shape of all four. He cannot produce the substance of any of them, because the substance requires having done the work. So it works.

It’s no longer just a conversion problem. Posting fabricated reviews and a rating that no one got is a consumer protection issue, and it is on the customer’s domain under the customer’s name.

Average is a slow death

We live in the age of personalization. Everything a consumer touches, from their banking app to their streaming service to the ads in their feed, is tailored to them. And in the midst of all this, an entire industry has begun to integrate graphic design, web design, copywriting, and campaign strategy into a machine whose defining characteristic is that it produces the average of everything it’s ever seen.

Any brand that has been around long enough knows what the average does. It doesn’t kill you on a Tuesday. It erodes you. No one chooses you, no one remembers you, and the investigations drag on for eighteen months without ever appearing as a single identifiable cause.

Gen Z wants expertise and personality. Generation X wants it. Baby boomers like me want it. This has been at the heart of good marketing for at least a decade and nothing in the current toolset changes it. On the contrary, tools make true personality rarer and therefore worth more.

We are not anti-AI

We use it every day. We build with it. The argument is not that it should be avoided.

The discussion is about who is driving.

Giving AI to an untrained person is like giving them an F1 car. Lots of people could rock one. Most would drive it slower than a normal car and much more dangerously, and none would achieve what a professional would do with the same machine. The machine is not the variable. The driver is.

Used by someone who knows what a page should do, AI is a real accelerator. Used as a replacement for knowledge, it produces exactly the average result that platforms tag, models score, and readers skip.

Where does it go next

The tipping point is no longer a prediction. The infrastructure is live.

Suppliers mark the release at the generation. Platforms label and, in the case of YouTube, implement labels automatically, whether the creator discloses it or not. Regulators have a code of good practice which has 190 signatories. And each of these companies has its own model for protecting itself from downgrade, meaning none of them have any incentive to turn back.

What follows from this is simple. As the volume of content generated increases, audiences spot it better, not worse. Consumers are already learning what’s happening and telling us in interviews. Businesses will follow the numbers, as they always do. Hype will build on anything that produces results at a defensible cost.

When that happens, writers, designers, and strategists who can produce something that isn’t average become more valuable than they were before all this started. No less. This is not a prediction we expect. This is currently shown by our own data, platform policies and the top of the SERP.

We built the agency around that. Our work is still done by people who have been doing it for years, with AI in the toolbox and away from the steering wheel.

If you want the version that applies to construction sites, we have written it under web design in East London . The visibility side of search is on our AI entity and digital authority optimization page .

Sources

• How Claude Brands AI-Generated Content, Anthropic
• Code of Practice on Transparency of AI-Generated Content, European Commission
• The slop button of LinkedIn’s AI, Gary Monro
• How YouTube Works: AI

Frequently asked questions

Is AI content ranked on Google? It can. Google’s stated position is that it rewards useful content, regardless of how it was produced. Our own data suggests that the practical results are different, particularly when it comes to indexing speed and the frequency with which content is cited by AI search tools. When we analyzed the top ten results for competing service requests through AI checkers, seven came back as human writing.

Is AI content penalized? Not in the sense of a formal sanction. What we see is softer and more damaging: slower indexing, lower citation rates, shorter wait times, and fewer conversion actions. The effect is the same no matter what you call it.

Can AI content be detected? More and more, and not by conjecture. Text in newer models has an embedded watermark, and generated images and files have signed provenance metadata. Detection tools that infer based on writing style are unreliable. Model-embedded watermarking applied at the time of generation is a completely different matter.

Does the watermark disappear if I edit the text? Anthropic’s advice is that heavy editing, paraphrasing, or translation may leave no detectable mark, and short passages may not. A light reread won’t remove it. A real human rewrite is another matter.

Does a detected watermark mean the content is written by AI? No, and this is where it becomes unfair. The mark appears on text produced by a model, including text in which a person provided the substance and the model only edited or translated it. This is a weak positive signal, not proof of paternity.

Should we disclose the use of AI in our marketing content? Under the EU Code of Practice, deployers must disclose AI-generated text posts that inform the public on matters of public interest, with an exemption where the post has undergone human review and falls under editorial responsibility. Seek advice on your own obligations, but the direction is clear: human scrutiny is what keeps you on the right side.

Can I simply generate my website design with AI and send it to a developer? You can, and customers do. What happens is almost always the same layout, regardless of industry: dark hero, row of icons, three cards, repeated. It seems finished and there is no hierarchy, no main action and no decisions behind it. It’s useful as a mood board. This is not a design.

My AI-generated design looks good. Why wouldn’t that work? Because you are looking at an image of a desktop page. Stacked on a phone, these grids of eight cards and strips of four icons become twenty almost identical blocks for the visitor to scroll through, with generated images loading onto mobile data along the way. Mobile design is a set of decisions about what elements to remove. Generation adds.

So, should we stop using AI completely? No. Use it where it’s good: research, structure, first drafts to react to, repetitive production work. Don’t use it as the final voice of your brand on a page where you hope to drive inquiries.