> ## Content Index > Fetch the complete content index at: https://www.metatalks.ai/llms.txt > Use this file to discover other available public pages before exploring further. # The Internet Just Got a New Reader. No One's Seen Its Face - URL: https://www.metatalks.ai/the-internet-just-got-a-new-reader-no-ones-seen-its-face/ - Published: 2026-08-18T15:01:02.000Z - Updated: 2026-08-18T15:57:17.000Z - Author: Mark Tiangov - Tags: DigitalMarketing, GenerativeAI, AIVisibility Something strange started showing up in the data: traffic patterns were changing, AI platforms were sending highly qualified visitors, and bots were reading pages that barely mattered in traditional search. We started digging, collecting traces and trying to understand who — or what — was consuming this content differently. To help us make sense of the evidence, we brought in[ **Vladislav Pivnev**](https://pl.linkedin.com/in/pivnevvladislav/ru?ref=metatalks.ai), CEO and co-founder of ICODA. Together, we follow the clues, examine the reader's habits and gradually reconstruct a profile of the new presence hiding inside the internet. **MT: What was the first thing you saw that made you think there was someone on the internet reading content in a completely different way from humans?** [**Vladislav Pivnev**](https://x.com/p%5Fvlad?ref=metatalks.ai)**:** The first signal came from analytics rather than rankings. About a year and a half ago, one of our clients started receiving referral traffic from chatgpt.com. The volume was small, just dozens of visits, but the conversion rate was 46%, almost five times higher than traffic coming from Google. That made us look deeper. When we checked the server logs, we found the other half of the story: OpenAI and Anthropic bots were reading pages that generated almost nothing in Google. Someone was consuming our content at scale, sending people who were already close to making a decision, and leaving almost no trace in the traditional user journey. That was the first real clue that we were dealing with a different kind of reader. **MT: If we forget about ChatGPT, Claude and Gemini for a moment, what strange changes in the data simply cannot be explained by traditional SEO?** **Vladislav Pivnev:** There are several patterns that immediately stand out. The first is that rankings remain stable while clicks decline, especially for informational queries that used to generate substantial traffic. Under the old rules, a top position was strongly associated with clicks. Ahrefs measured a 58% decline in CTR for the first organic position when an AI answer appeared above the results. The second is that pages with no meaningful external links can suddenly start generating leads. According to classical SEO logic, those pages should have very limited ability to attract an audience. The third is that branded search and direct traffic can increase even when there has been no corresponding advertising or PR push. Something is recommending the brand inside a closed environment that we cannot directly observe. We only see the consequences. That is where the old analytics model starts becoming blind. **MT: So let's say we know nothing about this new reader. If you had to build a profile from the first clues, what characteristic would you start with?** **Vladislav Pivnev:** I would start with the fact that it does not click. A human reader arrives, scrolls, leaves, comes back later, compares pages and creates a measurable session. This reader can take the text once, carry it somewhere else and never appear on the site again. That breaks a lot of traditional analytics because most of our measurement systems were designed around sessions and user journeys. We are accustomed to asking where someone came from, what they clicked and what they did next. With AI systems, a large part of that journey happens somewhere we cannot see. **MT: What habits can you actually observe? What does this reader do with content that a human almost never does?** **Vladislav Pivnev:** It takes fragments rather than pages. It might extract a paragraph, a table or a list because that particular fragment answers a specific question. The design of the page is largely irrelevant. What matters is whether the information can be understood when it is taken out of context. There are also technical behaviors that make the difference obvious. AI agents can read robots.txt and sitemap.xml, files that a human visitor would almost never open. The same page can also be requested multiple times within a minute by different agents, which is another behavior you would never expect from a normal reader. Once you see those patterns in server logs, the metaphor of a new reader starts becoming much more literal. **MT: If you give this reader two equally strong articles, why does it use one in its answers and seemingly ignore the other?** **Vladislav Pivnev:** The difference often has less to do with overall quality and more to do with whether a fragment can be safely extracted and repeated. Language models construct answers from chunks. That means a strong piece of content needs to remain understandable when a paragraph is separated from the rest of the page. An extractable paragraph names the subject, gives a number, includes a date when relevant and makes it clear who is making the claim. A weaker version might say, “this solution,” “recently” or “as we mentioned above.” Once that paragraph is removed from the page, the meaning disappears. There is another layer that matters just as much: independent confirmation. If a fact exists only on your own website, an AI system has less evidence that it should trust and repeat it. If the same information appears in Reddit discussions, reviews, industry publications or other independent sources, the claim becomes much easier to validate. So the reader is effectively asking two questions: can I extract this information, and can I trust it enough to repeat it? **MT: What traces does this reader leave behind? How can a company tell that AI systems are regularly using its content?** **Vladislav Pivnev:** I would separate the evidence into three groups. The first is visible in server logs. Different AI companies have multiple agents with different purposes. OpenAI, for example, has GPTBot for training-related crawling, OAI-SearchBot for search results and ChatGPT-User for requests associated with user interactions. Anthropic has a similar separation between ClaudeBot, Claude-SearchBot and Claude-User. The second group appears in analytics: referral traffic from chatgpt.com, perplexity.ai, claude.ai and similar platforms. It is often small in volume but can convert significantly better than average site traffic. ICODA's own research shows that AI-sourced traffic can have substantially higher conversion rates than conventional organic traffic. The third trace is the most important, and companies often fail to measure it: the answers themselves. Take 100 or 200 commercial queries that describe the problems your customers actually have. Run them regularly across the major AI platforms and track how often your brand appears, where it appears and which sources are cited. The first two groups tell you that AI systems are reading you. The third tells you whether they are actually using you. **MT: Many companies believe they can simply add more keywords, FAQ sections or mentions of ChatGPT. How close is that idea to reality?** **Vladislav Pivnev:** FAQ sections can help, but not because the label “FAQ” has some special power. They help when the question-and-answer format creates a self-contained chunk that an AI system can easily extract. Keyword density is much less relevant. Models do not need you to repeat the same phrase twenty times to understand what a page is about. There is also a lot of attention around llms.txt. The industry has been promoting it as a new standard, but major AI vendors have not confirmed that they use it as a ranking or citation signal. So companies can spend time creating a file that does not actually solve their visibility problem. There are situations where llms.txt can be useful, particularly for coding agents such as Claude Code or Cursor when they work with documentation. But that is very different from treating it as a universal AI search optimization mechanism. **MT: What piece of evidence breaks the old theory most clearly? What have you seen that traditional SEO simply cannot explain?** **Vladislav Pivnev:** The most uncomfortable fact is that Google position has stopped being a reliable predictor of AI citation. Ahrefs analyzed 15,000 queries and found that only 12% of links cited by ChatGPT, Gemini and Copilot were also present in Google's top 10\. That matches what we see in practice: a page can hold the number-one position for a query and never be cited by an AI system, while the same answer can reference a page that does not even appear in Google's top 20. The second problem is that a significant share of citations comes from domains you do not control. Reddit, Wikipedia, YouTube, G2 and industry-specific publications can become important sources for AI answers. That means you can optimize your own website perfectly and still lose visibility because the conversation about your brand is happening somewhere else. The new search environment is distributed. Your website is only one piece of the evidence. **MT: Was there a moment when you were convinced you understood what was happening, and then the data forced you to change your mind?** **Vladislav Pivnev:** Absolutely. That was probably my biggest mistake when we first entered the space. I initially approached it as a new version of SEO. We would improve the website, fix technical issues, structure the content, add the right markup and increase visibility. Then we started analyzing exactly where AI answers were getting their citations. A large portion of those sources were not our clients' websites at all. They were external publications, reviews, discussions and other mentions. That forced us to change the model. We moved a significant part of the effort from the website to the wider information environment around the brand, because that is where AI systems were going to find confirmation. **MT: If it is not really about keywords or “SEO for AI,” what is this new reader actually looking for before it uses a piece of information?** **Vladislav Pivnev:** It is looking for a claim it can repeat without embarrassing itself. An AI system has its own cost of error. If it gives a user an incorrect recommendation, the answer becomes less trustworthy. So it naturally favors information that can be attributed to a specific source, checked and repeated with confidence. In practice, that means the claim should have a clear subject, a number when relevant, a date and a recognizable author or organization behind it. Then there is a fifth element outside your website: independent confirmation. The same fact appearing elsewhere is much stronger than a claim that exists only on your own domain. In that sense, the new reader is managing its own risk rather than judging your copywriting. **MT: Can you give us an example of a structural or editorial change that made content appear more frequently in AI answers? What actually changed?** **Vladislav Pivnev:** The most useful change can be described in one sentence: we stopped writing articles and started writing answers. We take a long piece of content and divide it into sections where each section addresses one real customer question. The first two sentences under a heading should already provide the answer rather than promising to explain it later. Then we restore the entities. Instead of “the platform,” we use the actual name. Instead of “recently,” we give the month and year. Instead of “grew significantly,” we give the percentage. We also move comparisons into tables where appropriate. Tables can be extracted almost as complete units, while prose is more likely to be reconstructed from separate fragments. And we remove internal references such as “as we mentioned above.” They may work for a human reading the whole article, but they make an isolated paragraph much harder for an AI system to understand. That is the difference between writing a page and writing a collection of answers that can travel independently. **MT: If you compare this new reader with a human role, is it closer to a search engine, an editor, a researcher or a fact-checker?** **Vladislav Pivnev:** A fact-checker who has a deadline. A search engine gives you ten links and leaves the selection to you. An editor improves the material. A researcher may spend a week investigating contradictions and following sources. This reader has a different job: it needs to produce one answer and take responsibility for that answer immediately. That is why it likes numbers, dates, names and independent confirmation. It can use beautiful writing, but beautiful writing without verifiable details is much less useful to it. **MT: Which companies are making the biggest mistakes when trying to appeal to this reader? What are they still doing as if the investigation has not even started?** **Vladislav Pivnev:** The first group consists of companies that continue buying rankings and assume the problem is solved once they reach the top of Google. The second group blocks AI crawlers through Cloudflare or robots.txt and then wonders why ChatGPT cannot find them. The third group builds websites almost entirely in JavaScript, hides important information inside PDFs and puts prices behind forms, then wonders why AI systems describe the company only in generic terms. The fourth group buys an llms.txt file and a GEO checklist and assumes the problem is solved by adding one more technical asset. The fifth group is probably the largest: companies have almost no presence on the third-party platforms where AI systems look for independent confirmation. All of these companies are operating as if the internet still has one reader, and that reader is human. **MT: If we put all the evidence together, who is this new reader? How would you describe the suspect now that we have a complete profile?** **Vladislav Pivnev:** We are looking at a reader that may arrive once, take the exact fragment it needs and never return. It does not care about your visual design. It may not see information rendered through JavaScript. It can read technical files that humans ignore. It takes a paragraph rather than a page, and it prefers a paragraph that can stand alone with a clear subject, number, date and source. It also checks you against other parts of the internet because your own website is only one piece of the evidence. And there is one final detail that matters enormously to businesses: this reader can influence people who are already close to making a decision while leaving very little evidence about how that decision happened. That is why AI visibility is becoming much more important than simply generating another visit. **MT: If someone wants to conduct the same investigation inside their own company, what evidence should they collect first to build their own case file?** **Vladislav Pivnev:** Start with your own data. There are usually more clues there than people expect. First, take your server logs for the last 90 days. Filter them for AI agents and look at what they read, what they ignore and where they encounter errors. Then check your robots.txt and Cloudflare bot rules. Make sure you have not accidentally blocked the systems you want to be visible to. Next, create a separate GA4 segment for referrals from chatgpt.com, perplexity.ai, claude.ai and similar platforms. Look at conversion separately instead of mixing AI traffic into your overall acquisition numbers. Finally, build a list of around 100 queries that describe the problems your customers actually have. Run those queries regularly for a month and record which brands appear, which sources are cited and how frequently your company is mentioned. That gives you the beginnings of a real case file rather than another generic GEO checklist. If you want to take the investigation further, ICODA has compiled the broader methodology, data and case studies in **HackGPT: How to Become №1 in AI Search**. The guide covers AI visibility, citations, third-party sources, content structure and measurement, with real cases and numbers from ICODA's work. [Read the HackGPT AI Marketing Book by ICODA](https://icoda.io/ai-marketing-book/?utm%5Fsource=chatgpt.com)