Chapter 12: Content That Can’t Be Copied

Last updated: November 1, 2026.

AI is very good at summarizing what’s already been written online. Ask “How do I choose a filling machine?” and it pulls together what hundreds of pages have in common in a few paragraphs. If your page is one of those hundreds, your contribution to the answer melts away. Even if you’re cited, someone else could just as easily have been cited instead.

The way out is the things you know and others don’t: original data, real tests, case studies and expert opinion. In all of them, you are the source of the information.

I think this is the longest-lasting investment you can make in AI search. You make your technical fixes once, and your competitors make them too. They can’t copy your data, your tests or your customers’ experience.

1. Common knowledge is losing value

In its guide to AI search, Google splits content into two groups (Google Search Central).

The first is commodity content. Google’s example is “7 Tips for First-Time Homebuyers.” According to the guide, this kind of content is usually based on common knowledge, could come from anyone, and doesn’t add much that’s new for the reader.

The second is non-commodity content. Here the example is a homebuyer’s own story: “Why We Waived the Inspection & Saved Money: A Look Inside the Sewer Line.” The guide describes this kind of content as going beyond common knowledge and drawing on expertise or experience.

In the same guide, Google says content people find original, compelling and useful will, over the long run, affect a site’s presence in AI search more than any of the guide’s other recommendations. It also gives a warning: “Don’t just recycle what others on the internet have already said, or could easily be produced by a generative AI model.”

This isn’t the first time Google has said it. In a May 2025 blog post, John Mueller recommended “unique, non-commodity content” for doing well in AI search, and pointed out that people ask longer, more specific questions in these experiences (Google Search Central Blog).

The second half of that warning describes something new. Commodity content used to at least bring traffic: people searched on Google, clicked your page and read it. Now AI gives the same information inside the answer. Publishing content an AI could write on its own means telling it something it already knows.

Take filling machines. An article titled “Types of filling machines: piston, peristaltic, overflow” is common knowledge. Every manufacturer’s site has something like it, and the model can describe it from memory. An article titled “We filled foaming liquid soap at different nozzle and speed settings: at what speed does it overflow?” is different. Nobody who hasn’t run that measurement can write it.

2. Information gain: what are you saying that’s new?

In SEO this topic usually goes by the name information gain. The idea comes from Google’s “Contextual estimation of link information gain” patent, filed in 2018 and granted in June 2024 (Search Engine Journal).

The patent describes estimating how much additional information a new document offers a user, based on the documents they’ve already seen. If someone has read three pages on a topic, does the fourth tell them the same things, or something new?

A patent doesn’t prove Google uses the method. Search Engine Journal’s analysis notes that the text mostly talks about automated assistants and contains no clear statement that it’s used in AI Overviews. It also says the popular reading, “the page with more information than competitors ranks higher,” isn’t supported by the patent.

So it’s healthier to use information gain as an editorial question than as a ranking signal. Before you publish a page, ask: what will someone who has read the top 10 pages on this topic learn from ours? If the answer is “nothing,” the page won’t add anything to an AI answer either.

Query fan-out makes this question matter even more. As I explained in Chapter 3, the model runs several searches for a single question and can read many pages. If those pages all say the same thing, the model picks one and the rest add nothing. A page with different data, a different measurement or a different experience adds a piece to the answer that can only come from there.

3. The web is filling up with AI text

Using about 43,000 English-language articles sampled at random from Common Crawl, Graphite calculated that in November 2024 the number of new AI-generated articles overtook the number written by people (Graphite). Graphite later reran the study with three different AI detectors and noted the share came out 3.3 points lower on average. The exact share is debatable, but the direction is clear.

The same team’s second study is closer to our subject (Graphite). When they looked at Google results for 31,493 keywords in June 2025, 86% of ranking articles were written by people. Among articles cited by ChatGPT and Perplexity, the figure was 82%. Only 7% of articles in the top position were AI-generated.

The study has limits. AI detectors make mistakes; Graphite reports a false positive rate of 4.2% for the detector it used. AI-assisted content that people edited heavily wasn’t measured either. So this data isn’t saying “don’t use AI.” What it shows is that while much of the web is turning into AI text, both Google and AI assistants still pick their sources largely from content written by people.

Google’s guidelines for its quality raters point the same way. According to the September 11, 2025 version, a page should get the lowest rating if all or almost all of its main content is copied, paraphrased, auto-generated or AI-generated with “little to no effort, little to no originality, and little to no added value” (Google Search Quality Rater Guidelines). The same section adds that using AI tools doesn’t by itself determine the level of effort or quality, and that high-quality content can be made with them.

As AI text gets cheaper, the thing that gains value is information only you can produce. If everyone can write the same article with the same tool, the difference comes from what you can’t hand to the tool.

4. The data you already have

When people hear “original data,” they tend to picture a big survey or an industry report. Those are valuable, but they aren’t the only option. Most companies are already sitting on data nobody has published:

• Sales and order data: which product sells more in which city, in which season, to which type of customer? What’s the average order size?
• The questions that reach your sales team, call center and WhatsApp line most often, and how they change through the year.
• Service and return records: the most common fault, the most common reason for returns.
• The price history of a product or service over the past three years.
• Your own measurements, such as delivery times, installation times, energy use and durability.

This book has an example. In Chapter 1, I wrote that among İmza’s clients, traffic from AI platforms usually stays under 10% of total traffic, but those visitors convert at around 6 to 7%. With general e-commerce conversion rates around 1%, that’s a big gap. You won’t find that number anywhere else; it comes from our own client data. An AI can only learn it from where we’ve written it down.

When you publish original data, state the method: how many customers, what date range, how it was measured. A number without a method looks weak to readers and to models alike. Give a date, too. A phrase like “in the first nine months of 2026” tells people when the figure applies and lowers the risk of AI passing off an old number as current. And don’t hide the limits: if the sample is small or the industry is narrow, say so. An inflated number may get attention in the short term, but it hurts your brand when it gets repeated wrong.

You don’t have to publish data you need to keep private. Sharing a ratio, a trend or a range instead of raw figures is often enough. A sentence like “More than half of our orders come from outside major cities” says something without revealing revenue.

5. Real tests: measure before you write

For any site that sells products or writes about them, tests are the strongest kind of content that can’t be copied. Google’s guide to writing reviews sums up how to do it well (Google Search Central). Some of its points:

• Show that you know the product; provide photos, audio, video or other evidence that supports your experience.
• Share quantitative measurements of how the product performs in different categories.
• Explain what sets the product apart from its competitors.
• Describe its pros and cons based on your own research.
• If you recommend a product as “the best,” back it up with first-hand evidence of why.

The guide recommends focusing on quality and originality in reviews, not length.

Picture a filling machine manufacturer. For each machine, its site lists the catalog specs: capacity, number of nozzles, motor power. Competitors’ sites list the same things. This manufacturer could take the five product types customers ask about most, say liquid soap, shampoo, thick cream, vinegar and cooking oil, and run them through its own machine. For each product it measures bottles per minute, fill accuracy and the speed at which foam starts to overflow, films the test and puts the results in a table.

A page like that gives the user who says “I need to fill 3,000 bottles of a foaming product a day” a concrete answer no other page can give. It also gives AI a reason to recommend this manufacturer. In Chapter 3 we covered the difference between being mentioned, being recommended and being cited; to recommend you, the model needs an answer to “why this brand?” A test result can be that answer.

Don’t publish only the good results. Write down which products the machine struggles with, too. That builds trust and filters out customers you aren’t a fit for. The pros and cons point in Google’s guide asks for exactly this.

6. Case studies

A case study shows how a product or service worked for a specific customer. In B2B and services, I think it’s the most valuable type of content, because the decision maker’s question is usually “Will this work in my situation?”

A good case study has a simple structure:

1. Starting point: who the customer is, what the problem was, which numbers they started from.
2. What was done, why that solution was chosen, how long it took.
3. Result: which number changed, by how much and when.
4. What didn’t go as expected.

A case study needs numbers. “Our client was very happy” tells neither readers nor AI anything. A number sticks, and it can be quoted.

An example from our own field: for an e-commerce client selling volleyball and basketball shoes, traffic fell 10% while revenue rose 180%. That one sentence is enough to make you question the idea that SEO success means traffic growth. A case study about a result like that should show step by step what changed, which traffic was lost and why, and where the revenue came from. Only the team that did the work can write it. An AI can’t produce that result from memory.

There are two practical problems. One is client permission. Many clients don’t want their names used; in that case you can share the industry, the scale and the numbers and leave the name out. An anonymous case study doesn’t carry the weight of a named one, but if the numbers are accurate it’s still valuable.

The other is selectivity. A case study page that only shows the most impressive results starts to read like a sales brochure after a while. Including average results and the times things went wrong builds more trust.

7. Expert opinion: content with a name on it

Expert opinion is the view of someone who has done the work themselves for years. An engineer saying “Don’t fill this product on this machine,” or a dentist saying “I don’t recommend this treatment in this situation.” These statements come from experience, which is what separates them from common knowledge.

Google’s questions for evaluating content are clear on this (Google Search Central). Was the content written or reviewed by an expert or enthusiast who demonstrably knows the topic well? Does it show first-hand expertise from actually using a product or visiting a place? If it draws on other sources, does it add substantial value instead of copying or rewriting them? This is the “experience” part of E-E-A-T we discussed in Chapter 2.

There’s a similar pattern on the AI side. The “GEO: Generative Engine Optimization” study, published in 2023 by a research team that included Princeton University and presented at KDD 2024, tested this (arXiv). The three methods that worked best were adding quotes from credible sources, adding statistics and citing sources. They increased a page’s visibility in generative engine answers by roughly 30 to 40% in relative terms. Classic keyword stuffing didn’t help. According to the researchers, these methods especially helped sites ranked lower in search results.

The study was run in a test environment the researchers built themselves, and it doesn’t measure today’s ChatGPT or AI Mode. But it points in the same direction as Google: concrete numbers, a named expert and verifiable sources make content more useful to AI.

In practice, that means:

• Put a real name on the content. The name, title and experience of the author or the expert who reviewed it should appear on the page. The quality rater guidelines list fake author profiles built with AI-generated images and made-up bios as an example of deceptive practice. Don’t invent an “expert” who doesn’t exist.
• Use the expert’s own words. Do a 20-minute interview with your engineer, sales manager or service lead and quote their clearest sentences.
• Don’t shy away from opinions. Sentences like “I think,” “we don’t recommend” and “in this case, choose this” don’t appear in common knowledge. That’s where expertise shows most.

That’s also why this book is written in the first person. A sentence that starts with “I think” can only belong to one person.

8. What an agency or brand knows firsthand

Under data, tests, case studies and expert opinion lies the same thing: the knowledge a company builds up while doing its work. You could call it institutional memory.

Three examples from our own institutional memory as an SEO agency:

• Noticing that a news site’s sitemap was cached for about 10 minutes, and cutting that down to seconds.
• At Filmler.com, opening only the pages with content to indexing, out of about 450,000 cast and crew records.
• Starting uptime monitoring for every client after one client’s site was down for about an hour.

None of these come from general advice like “keep your sitemap up to date.” We learned all of them on the job. An AI can answer “How do I build a sitemap?” from memory, but it can only learn these examples from where we’ve written them down.

Every industry has its own version. For a filling machine manufacturer, institutional memory is the answers to questions like these: Which nozzle causes trouble with which product type? What’s the most common mistake during installation? What do customers ask about most in the first month after buying a machine? Those answers live in the heads of the sales team, the service technicians and customer support, and most of the time they’re never written down.

One of the content team’s main jobs is getting this knowledge out of people’s heads and onto the page. It doesn’t take a big process. A monthly “What did customers ask this month?” meeting with sales, a quarterly look at the five most common issues in service and return records, and a short, concrete “note from the field” from the technical team each month are enough.

In B2B this matters even more. One of our B2B clients gets 30 to 40 visits a day and sells a €20,000 machine. For a site like that, I don’t think writing hundreds of blog posts makes sense. What it needs is a few pages that answer the questions of the handful of people making the decision better than anyone else can. When visitor numbers are low, each page’s value comes from how original its information is.

9. The customer’s voice

Some content that can’t be copied is written by your customers: product page reviews, Q&A sections, photos customers take themselves.

A customer review covers what the catalog doesn’t: whether the fit runs narrow or wide, how many days shipping took, what the product looks like after six months. That information comes from real use and isn’t written anywhere else. The questions users ask AI are often of the same kind: “Will this shoe fit a wide foot?” “Can this machine fill thick cream?”

Show reviews in the page’s text. If they only appear in a pop-up that opens on click or in a JavaScript component that loads late, bots are less likely to see them; the JavaScript rules from Chapter 2 apply here too. Have an expert answer the Q&A section. A clear answer from someone on the sales or technical team is valuable both to the person who asked and to anyone else who asks AI the same question.

Don’t delete negative reviews, either. A page with only five-star reviews doesn’t inspire trust. A calm, concrete reply to a negative review is one of the best pieces of content for showing how a brand works.

Buying or faking reviews is the opposite of everything in this chapter: not content that can’t be copied, but fake content. Reviews outside your site, such as Google reviews, complaint sites and marketplaces, are a separate topic covered in Chapter 14.

10. Making original information quotable

Having original information isn’t enough; AI also has to read it correctly and pass it on correctly. In Chapter 3 we saw that the model uses a page paragraph by paragraph rather than as a whole. That matters even more for original data.

• Put the number and its source in the same sentence. Instead of “Most of our customers…,” write something like “According to X company’s 2026 analysis of 1,200 orders…” When the model carries that paragraph into another answer, your brand goes with the sentence.
• Put the method next to the number. Keep the sample, the date range and how it was measured in a short paragraph right below the figure.
• Put test results and comparisons in tables. Tables are easier for readers and models to follow; we covered the rules in Chapter 10.
• Back up video with text. A test video is strong evidence, but write the results on the page as text too. Stating a number only inside a video hides it somewhere hard to read.
• Publish at a single, permanent URL. Don’t bury the data in a PDF report or a slide deck. Put it on a page that stays at the same address when it’s updated.

You might worry that if you publish your data, competitors and AI will use it. They might. But I think the bigger risk is that nobody knows your data exists. If a competitor quotes your number, they usually link to you or mention your name. If AI uses it, you’re the source. Original data is also one of the strongest reasons for other sites to cite you; we’ll get to that in the Digital PR part of Chapter 14.

If you want to keep your content out of model training, we covered the robots.txt options in Chapter 9. Think about that decision together with your visibility in live search; mixing the two up can take you out of answers without meaning to.

11. Is writing with AI off-limits?

No. Google looks at the result, not the tool. Its scaled content abuse policy, announced in March 2024, targets producing large numbers of unoriginal pages to manipulate rankings, whether the content is made by automation or by people (Google Search Central Blog). The AI optimization guide also asks that, if you use generative AI tools, your work follows Google’s essentials and spam policies.

The right split, in my view: the knowledge comes from you, and AI can help with structure and wording. You gather the test results, the customer questions and the expert’s notes; AI can help you organize them, suggest headings and turn them into tables.

If you tell an AI “Write a 1,500-word article on choosing a filling machine,” what comes out is common knowledge by definition, because the model doesn’t know what you know; it knows what everyone has written. It can’t produce a test result or customer story you haven’t given it. If it does, it’s making it up.

We’ll cover the risks of producing content with AI, and quality control, in Chapter 19.

12. Starting small

All of this can look like a big job. It gets easier once you see where in the company the raw material sits:

Rather than tackling everything at once, you can work in this order:

1. Pick your ten pages closest to a sale: product, service or category pages.
2. For each one, ask “What does this page have that no other site does?” and write the answer in one sentence. If you can’t, the page is common knowledge.
3. Add one original element to each page: a measurement, a customer question and its answer, an expert note or a case summary. Not all of them at once, just whichever is easiest to get.
4. Sit down with your sales and service teams for an hour and write down the ten questions they hear most, with the answers. That list becomes raw material for the content you’ll produce over the coming months.
5. Check again in three months: are these pages mentioned, recommended or cited more often in AI answers? I explained how to measure that in Chapter 7.

What AI doesn’t know

AI already knows common knowledge and can summarize it faster than you. What you can add is something it doesn’t know: your own numbers, your own tests, what happened to your customers, what your engineers have learned over the years. Most of that knowledge already exists inside the company. What’s missing is the habit of getting it onto the page on a regular basis.

That wraps up the content side. The next chapter looks at the brand itself: does AI know who you are, and if so, does it get you right?

Checklist

Sources

• Google Search Central: Optimizing your website for generative AI features on Google Search (last updated July 10, 2026)
• Google Search Central Blog: Top ways to ensure your content performs well in Google’s AI experiences on Search (May 21, 2025)
• Google Search Central: Creating helpful, reliable, people-first content (last updated October 5, 2026)
• Google Search Central: Write high quality reviews (last updated December 10, 2025)
• Google Search Central Blog: What web creators should know about our March 2024 core update and new spam policies (March 5, 2024)
• Google Search Quality Rater Guidelines (September 11, 2025 version)
• Search Engine Journal: Google’s Information Gain Patent (August 15, 2024)
• Graphite: More Articles Are Now Created by AI Than Humans (October 14, 2025)
• Graphite: How Does AI-Generated Content Perform in Search and Answer Engines? (October 14, 2025)
• arXiv: GEO: Generative Engine Optimization (Aggarwal et al.) (November 2023, KDD 2024)

Previous: Chapter 11, Content in the Age of AI · Table of contents · Next: Chapter 13, Brand and Entity