The New SEO Problem Isn't AI Content. It's Content Without Evidence

For years, one of the most tempting shortcuts in SEO was also one of the simplest: Ask an AI model to write the article. Copy. Paste. Publish. Rank.

The economics were compelling. What once required a writer, researcher and editor could be produced in minutes. At scale, a single team could create hundreds—or thousands—of pages.

But something fundamental is changing. Not because AI-generated content is suddenly illegal. It isn't. Not because Google has announced a universal penalty for AI-written pages. It hasn't. And not because an AI detector has finally become capable of identifying every sentence written by a machine. It hasn't.

AI-generated content is beginning to acquire provenance. At the same time, search engines are becoming answer engines, and the value of being merely another page about a subject is declining. The result is a new competitive environment in which original information, evidence, experience and provenance may matter considerably more than the ability to produce another 2,000 words on a keyword.

1. The Provenance Era Has Arrived

On 2 August 2026, the transparency obligations in Article 50 of the EU AI Act became applicable.

The legislation requires providers of certain AI systems to ensure that AI-generated or manipulated outputs, including text, can be identified in a machine-readable format, subject to the Act's requirements around technical feasibility and proportionality.

The important point is what the legislation does not say. It does not prohibit AI-generated website content. It does not require Google to penalise AI-written pages. It does not establish a universal AI detector.Instead, it pushes the industry toward something that has historically been difficult to achieve: machine-readable information about content origin. It’s this distinction that really matters.

For years, the internet largely treated a block of text as simply a block of text. Search engines could attempt to infer its quality, authority and originality, but the underlying generation process was generally invisible. But this is beginning to change.

The EU's Code of Practice on Transparency of AI-Generated Content now provides a framework for providers to implement the Article 50 obligations, and major AI companies have committed to the code.

2. A Watermark Isn't a Digital Stamp

The phrase "AI watermark" can create the wrong mental image. It isn't necessarily hidden metadata, a special unicode character or a zero-width space. And it isn't something that disappears when someone pastes the text into Notepad. Modern generative watermarks can exist entirely inside the statistical process used to generate the text.

Consider how an LLM produces a sentence. At every step, the model generates a probability distribution over possible next tokens.

For Example, After:

"The weather today was cold and..."

Several Continuations Might be Perfectly Plausible:

"grey"

"overcast"

"miserable"

The exact choice between them may not matter much to the meaning of the sentence.

That is where watermarking can operate. Instead of changing the words after generation, the watermark changes which statistically plausible token is selected during generation.

The result is ordinary-looking text that contains a statistical pattern. Google DeepMind's SynthID-Text is a particularly important example. The system uses a pseudorandom process based on the preceding context and a secret key. Candidate tokens are then evaluated through a tournament-style sampling mechanism designed to favour choices carrying the watermark signal while preserving the model's normal output quality.

The signal is therefore not "inside" a character. It is inside the sequence of choices. And that is why copying the text from an AI interface into WordPress, Webflow, Shopify or a plain-text file does not inherently remove it. There is nothing special to strip away.

3. SynthID-Text: Watermarking Without Rewriting the Text

SynthID-Text, published by Google DeepMind in Nature, represents one of the most sophisticated production approaches to text watermarking.

At generation step t, the language model produces its normal probability distribution:

p(x_t | x_<t)

A watermarking system introduces a keyed pseudorandom process that depends on the recent context. SynthID-Text's Tournament sampling then draws candidate tokens from the model's distribution and allows them to compete using pseudorandom watermark functions.

The winner becomes the next token.

Repeat the process thousands of times and the individual choices remain effectively invisible to a reader - but collectively they can provide statistical evidence of watermarking.

The important word is statistical. A watermark is not a barcode. It is evidence accumulated over many generation decisions.

Google's research found that its non-distortionary configuration could preserve text quality, including in a large-scale experiment involving almost 20 million Gemini interactions. The research also describes a distortionary configuration that can increase detectability at some cost to quality.

That trade-off is important.

There is no single magical setting called "undetectable but perfect."

Watermarking Involves An Engineering Balance Between:

• Detectability
• Robustness
• Text Quality
• Computational Cost
• Diversity
• Resistance to Editing.

4. The Critical Distinction: Detection Isn't Authorship

This is where much of the public discussion around AI watermarks goes wrong. Suppose a detector identifies the watermark associated with Claude.

That Does Not Prove:

"Claude wrote this entire article." Anthropic's own explanation is considerably more careful. Its watermark can establish that Claude was likely involved in producing the text. It cannot establish whether Claude wrote the entire passage, substantially edited it, or was involved at another point in the content's creation.

It also cannot tell you that a piece of text was generated by another AI system. A watermark therefore provides model-specific provenance evidence, not universal authorship verification.

There Is a World of Difference Between:

AI Detection

"This looks statistically like AI-generated text."

Watermark Detection

"This text contains statistical evidence consistent with a particular watermarking scheme."

And:

Provenance

"There is verifiable information about where this content came from and how it was produced."

These are three different things. Treating them as interchangeable is one of the easiest ways to misunderstand the next stage of AI-generated content.

5. Why a Z-Score Isn't Proof of AI

The mathematics of statistical watermark detection can be elegant. For some green-list watermarking schemes, a detector can measure how frequently observed tokens fall into a pseudorandomly selected "green" set and calculate a statistic such as:

z = (|T|_G - γT) / sqrt(Tγ(1-γ))

Where:

• T is the number of relevant tokens,
• |T|_G is the number falling into the green set,
• γ is the expected proportion.

A large positive value indicates that the observed sequence is unusually consistent with the watermark hypothesis.

But It Is Incorrect to Say:

"z > 4 means the text is confirmed to be AI." The statistic is evidence under the assumptions of a particular detection method. It isn't a universal authorship certificate.

SynthID-Text, for example, can use Bayesian scoring and evaluates detection performance using measures such as true-positive rates at specified false-positive rates.

The practical lesson is much more important than the equation: watermark detection is probabilistic attribution. The longer the passage, the more generational decisions there are available to accumulate evidence. The shorter the passage, the weaker the statistical signal can become. This is one reason a two-sentence response and a 3,000-word article should not be treated as equivalent detection problems.

6. Where Watermarks Become Weaker

Watermarking also has a fundamental limitation because the model cannot always choose freely.

Imagine a Sentence:

"Isaac Newton's most famous work was called Principia Mathematica."

There isn't much room to manipulate the choice of "Mathematica" without risking an error.

The Same Problem Appears In:

• egal wording
• Mathematical notation
• Factual names
• Exact quotations
• Structured data
• Programming syntax
• Highly constrained instructions.

Anthropic explicitly notes that watermarking is weaker where the model has fewer meaningful choices. The same principle explains why grammar-only editing may provide relatively little new watermark signal. There simply aren't many new generation decisions. This is not a flaw unique to Claude.

It is a fundamental property of statistical watermarking and the more constrained the generation process, the less room there is to encode a signal without affecting the result.

7. Can Editing Remove a Watermark?

Sometimes. But there is no universal "watermark removal trick." so take anything you see on X with a pinch of salt

Changing a comma or replacing a handful of words may leave enough of the original statistical sequence intact for detection to remain possible. A sufficiently substantial rewrite can reduce the watermark signal dramatically because the statistical sequence itself has changed.

Anthropic says that a complete rewrite can remove the watermark. But that creates an interesting conceptual boundary. If an AI produces the original article and a human subsequently restructures every argument, replaces the language, adds original reporting, introduces new evidence and changes the conceptual flow, the final document is no longer simply a copy of the AI output.

That Distinction Matters Much More Than the Simplistic Question

"Was AI used?"

AI-assisted creation and AI-copy publishing are not necessarily the same thing.

8. And this Is Where SEO Gets Interesting

The biggest misconception is that the future of SEO depends on Google building a giant AI detector and deleting anything it identifies. There is currently no public evidence for that.

Google's Published Guidance Says Something Much More Nuanced

AI-generated content is not automatically against Search policies. Google's concern is whether content is created primarily to manipulate Search, including scaled production of unoriginal pages that provide little or no value. Its scaled-content-abuse policy explicitly applies whether content is created by:

• AI
• Humans
• Automation
• Or combinations of all three.

That is a critical distinction.

The SEO Problem Isn't:

"Google hates AI."

The Problem Is:

AI has made the production of low-value content incredibly cheap. And Google has spent years fighting exactly that economic problem.

9. The Real Threat Isn't AI Content. It's Content Inflation

Before generative AI, producing 10,000 mediocre articles required a large operation.

With modern LLMs, it can be automated. That changes the economics of search.

Imagine 10,000 Websites All Publishing:

"Best project management software in 2026"

Each Article Contains:

• Similar pros and cons
• Recycled statistics
• Generic recommendations
• Slightly different wording.

The internet does not become 10,000 times more informative. It becomes 10,000 times noisier. And this is why Google's current guidance is so important.

Its latest guidance for generative AI Search tells publishers to focus on valuable, unique, non-commodity content, while warning against creating separate pages for every possible query variation simply to manipulate rankings or generative responses.

The Direction Is Not:

"Human writing good. AI writing bad."

It Is:

"Give users something worth finding."

That is a much harder requirement to automate.

10. Generative Search Changes The Objective

Traditional SEO has largely been about winning a position.

Generative search introduces another objective. When an AI system answers a question, it can retrieve information from multiple sources, synthesise it and attach citations. That means there is an emerging difference between ranking for a query and being selected as evidence for an answer.

Google's own guidance for generative Search says its AI features are grounded in its existing Search systems, including retrieval of relevant and up-to-date web pages. That makes the future of SEO less about gaming a separate "GEO algorithm" and more about understanding what makes information useful to both people and retrieval systems.

11. There Won't Be One Universal AI-Search Rulebook

This is another area where marketers need to be careful. It is tempting to imagine that Gemini, ChatGPT and Perplexity all use essentially the same system. They don't necessarily. Their public documentation reveals different emphases.

Google: Relevance, Quality and Retrieval

Google says its generative Search features are built on its existing Search infrastructure.

Its Guidance Emphasises:

• Helpful content
• Unique information
• Non-commodity value
• Relevance
• Existing SEO fundamentals.

It explicitly says that SEO remains relevant to generative search. That means the best GEO strategy is not necessarily a new collection of tricks. It is often excellent SEO combined with information worth retrieving.

Perplexity: Source Classification and Editorial Signals

Perplexity has introduced explicit source labels.

Its Source-Review Process Can Classify Domains As:

• Government
• Academic
• Trusted

The Criteria Include Questions Such As:

• Does the site identify its authors?
• Does it correct mistakes?
• Does it separate news from advertising and opinion?

Importantly, Perplexity says that having no label does not mean a site is low quality.

This is an interesting development because it moves part of the conversation away from: "Is this page optimised?" towards "What kind of source is this?".

That is a fundamentally different signal.

ChatGPT: Meaning, Intent and Source Retrieval

ChatGPT Search also illustrates why "ranking" is becoming less straightforward. Search-enabled ChatGPT can use external search providers and return cited sources. OpenAI's own guidance warns that search results and citations can be incomplete or incorrect and encourages users to inspect the underlying sources.

For publishers, that reinforces a simple point that being retrieved is not the same as being trusted. The system still needs to decide which information is useful enough to incorporate into an answer.

Claude: Provenance At Generation Time

Anthropic's approach adds another dimension. Future Claude models are using a version of SynthID-Text to watermark generated text. That doesn't mean Claude is building a search-ranking system that automatically downgrades watermarked pages. It means that the generation layer itself can now carry provenance information forward into the wider information ecosystem. That could become increasingly important as provenance standards mature.

12. What About AI Systems Favouring Their Own Content?

This is one of the most interesting—and most misunderstood—questions. There is legitimate research showing that LLMs can exhibit self-preference when evaluating generated outputs. Studies have found that models can recognise and sometimes favour their own generations. One proposed explanation involves familiarity: models may assign lower perplexity to language that resembles distributions they are particularly familiar with.

That sounds alarming. But it does not automatically mean that Gemini ranks Gemini-written websites higher. That leap is not supported. In fact, research specifically examining fact-centric retrieval-augmented-generation settings has found no significant self-preference effect under the tested conditions, with factual accuracy exerting a much stronger influence.

So There Are Two Separate Observations:

Observation 1:

LLMs can exhibit self-preference in some evaluation environments.

Observation 2:

That does not establish that AI search engines deliberately favour websites generated by their own models. The distinction matters. The more defensible prediction is that different AI-search ecosystems will develop different combinations of:

• Retrieval
• Relevance
• Freshness
• Authority
• Source Quality
• Safety
• Provenance
• Citation Confidence.

That is enough to create meaningful differences between platforms without requiring a secret "our model wins" rule.

13. The Evidence Premium

This leads to what may become the most important shift in content strategy. Call it the Evidence Premium. When AI can produce a competent article about almost anything in seconds, competent prose becomes cheap.

What remains scarce? Information that the model cannot simply invent. So, let’s compare 2 pages: 

Page A

The 10 Best CRM Platforms for Small Businesses

It contains:

• Feature lists
• Generic advantages
• Disadvantages
• Pricing
• Conventional recommendations.

An LLM can produce it almost instantly.

Page B

We analysed 46 CRM implementations across UK SMEs over 18 months, comparing onboarding time, employee adoption, support-ticket volume and three-month retention.

That page contains something fundamentally different. An underlying reality. The second publisher has produced information. The first has mostly produced language. That difference could become enormously valuable.

14. What Does Genuinely Defensible Content Look Like?

Consider a few examples like travel.

Commodity:

The best things to do in Barcelona.

Evidence-led:

We visited 42 attractions over four months and recorded average queue times, ticket prices, opening-hour reliability and accessibility.

Now think about software.

Commodity:

The best CRM software for startups.

Evidence-led:

We migrated six startup sales teams between HubSpot, Salesforce and Pipedrive and measured setup time, user adoption and pipeline completion.

What about finance?

Commodity:

How to save money on your energy bill.

Evidence-led:

We analysed 12 months of energy consumption across 200 households before and after implementing five specific interventions.

Not to mention Product reviews

Commodity:

The best washing machines in 2026.

Evidence-led:

We tested 17 machines for cycle duration, noise, energy consumption, water usage and stain removal under identical conditions.

The distinction is obvious right? AI can generate the first version. AI cannot legitimately manufacture the underlying experiment.

15. This Changes What E-E-A-T Means In Practice

E-E-A-T is Often Reduced to a Checklist:

• Author bio
• Credentials
• Citations
• Company information.

But the deeper principle is more useful. Experience creates information. If someone actually…

• Tested the product
• Interviewed the customers
• Performed the experiment
• Visited the location
• Analysed the dataset
• Ran the campaign
• Implemented the software
• Repaired the equipment

…then they possess information that wasn't available to the model simply because the model knows how to write. This is why first-party information becomes strategically powerful. The advantage isn't merely that humans wrote it. The advantage is that someone experienced something.

16. From "Who Wrote This?" to "Where Did This Come From?"

This may ultimately be the biggest conceptual change.

The question typically is "Was this written by AI?". But, this is becoming less useful.

The Better Questions Are:

• Who created the underlying information?
• What evidence supports the claim?
• Was it independently observed?
• When was it measured?
• Who reviewed it?
• What methodology was used?
• Can another source verify it?
• Is the source authoritative in this subject?

That is provenance in the broader sense. And it goes beyond text watermarking. A watermark can tell us something about how content was generated. Editorial metadata, citations, datasets, authorship and methodology can tell us something about where information came from. The future will likely need both.

17. The New SEO Stack

The content strategy of the next few years should therefore look less like a writing factory and more like an information operation.

01: Use AI for Architecture

LLMs are excellent at:

• Keyword Clustering
• Research Organisation
• Outlines
• Information Architecture
• Internal Linking
• Schema Generation
• Content Briefs
• Summarisation
• Repurposing
• Analysis.

Use them aggressively where they create leverage. But don't confuse production efficiency with authority.

02: Put Humans Where Reality Enters The System

The human contribution should be strongest where actual knowledge is created:

• Interviews
• Testing
• Expert Judgement
• Original Research
• Experimentation
• Customer Feedback
• Analysis
• Interpretation.

The goal isn't to keep AI away. The goal is to make the final information more than AI could have produced from existing information alone.

03: Build Proprietary Evidence

Create things competitors cannot simply rewrite:

• Original Datasets
• Surveys
• Benchmarks
• Case studies
• Experiments
• Photographs
• Interviews
• Proprietary Research
• Customer Statistics
• First-party analytics.
  

This is the content equivalent of building a moat.

04: Make The Source Obvious

Tell both humans and machines:

• Who created the information
• Who reviewed it
• When it was published
• When it was updated
• How it was researched
• What evidence supports it.

Don't make the provenance of important claims difficult to discover.

05: Optimise for Retrieval

Pages should be easy for search and AI systems to understand. Use:

• Descriptive headings
• Clear answers
• Concise factual statements
• Supporting evidence
• Authoritative references
• Crawlable HTML
• Appropriate structured data
• Visible authorship
• Publication and update dates where relevant.

But avoid turning this into another keyword-stuffing exercise.

18. The New Content Workflow

The workflow we'd back looks like this. AI does the research and pulls the topic together. A person then goes and gets the evidence: the test, the interview, the numbers from the account. AI structures that evidence into a draft. Someone who knows the subject checks it and adds their reading of what it means. An editor makes it clear. The publisher puts a name, a date and a method on it. Search retrieves it, and the AI answering the question cites it. AI is in that chain four times. What's changed is that the material it's working from is richer than anything it could have written on its own.

The important part is that AI isn't removed. It becomes infrastructure. The difference is that the source material becomes richer than the model's generic output.

19. What The Next Generation of SEO Teams Should Measure

Traditional content teams often obsess over:

• Articles published
• Keywords targeted
• Word count
• Rankings
• organic sessions.

Those metrics won't disappear and remain as important as ever.

But they should be joined by another layer:

Information Metrics

Original data points published

How many claims on the site originate from proprietary research?

First-party Evidence

How much content is based on direct customer or operational information?

Citation Acquisition

How often are other sites and AI systems citing your original findings?

Source Visibility

Which pages are appearing as cited sources in generative search?

Evidence Density

How much of a page consists of verifiable information rather than generic explanation?

Update Velocity

How quickly can proprietary information be refreshed when the underlying reality changes?

This moves the content team away from measuring how much they publish toward measuring how much information they create.

20. The Uncomfortable Conclusion

The internet is about to experience something strange. AI will not necessarily make content less valuable. It may make generic content less valuable. When anyone can generate a competent article about a topic in seconds, competence becomes a commodity. 

The scarce asset becomes everything AI cannot legitimately manufacture:

• Experience
• Evidence
• Original research
• First-party data
• Expert judgement
• Provenance
• Trust.

This also changes the role of SEO.

The old question was “how do we get this page to rank?”. The new question is “why would a search engine or an AI answering a user need this page?”.

If the answer is "because we wrote a slightly better version of what everyone else has already said." then the competitive advantage is disappearing.

If the answer is "Because we generated information that didn't previously exist." the equation changes completely.

At the end of the day: If AI can write the page. Can it become the source?

The future of search probably won't be divided neatly into human content versus AI content.

That's too simplistic. AI will be involved in the production of an enormous proportion of the web.

The more important distinction will be between information that is merely generated and information that has an underlying source of truth. Watermarking may help machines understand how a piece of text was generated.

Provenance standards may help establish its origin. Search systems may become better at assessing authority, relevance and reliability. But none of those technologies changes the fundamental economics of information.

The organisations that win the next phase of SEO will therefore not necessarily be the ones with the biggest content factories. They will be the ones with the best information factories. They will use AI to scale what machines are good at. They will use humans where judgement and experience matter.

And, most importantly, they will create evidence that did not previously exist.

Because The Defining Question of The Next Era of Search May Not Be:

"Was this written by AI?"

It Might Just Be:

"Why should anyone, or any AI believe you?"

And in a web increasingly flooded with synthetic language, the answer may be the most valuable SEO asset of all…

proof.

‍