How Do ChatGPT and Other Language Models Search for Websites? GEO and AEO Do Exist After All, Although You Need to Be Wary of Courses and Offers to Promote Websites for ChatGPT.
First of all, I must point out that a key part of positioning yourself in language models is classic SEO. ChatGPT relies on what ranks highly on Google (and Bing).
Google Search itself has been based on artificial intelligence for many years, quite similar to that used in language models. Promoting a website solely for language models and “rejecting SEO” does not make much sense, yet it is easy to come across such offers.
At the same time, ChatGPT searches websites somewhat differently. People use text models differently than Google (they describe the problem more precisely), and the models themselves use longer key phrases.
Additionally, ChatGPT has certain instructions for searching through content; it works differently when you are looking for bicycles in the area and when you ask about health-related topics. Other models also have such rules or will introduce them.
A simple experiment: search for something on Google, and then on ChatGPT. You’ll notice that the results differ. The most important takeaway from this article is as follows: Ask ChatGPT why it recommends those particular websites.
Table of contents
-
It is worth exercising caution with offers to promote a website on ChatGPT.
-
How can you find out how ChatGPT searches the internet? The simplest way is simply to ask it.
-
Additionally, this translates into the way an SEO support network is built.
-
List of specific activities worth doing when promoting a website for GEO and AEO
-
What is not gaming the system in SEO and GEO, even though it may seem that way?
It is worth exercising caution with offers to promote a website on ChatGPT.
There are many courses and offers for promoting businesses in language models that are created without due consideration, full of errors, and typically produced in pursuit of a trend. Therefore, the claim that SEO/Google is a thing of the past and that positioning in ChatGPT is what matters now makes no sense whatsoever. It is impossible to promote a business for language models without promoting it in search engines.
The Plague of Three-Letter Acronyms.
One particular plague in the world of website promotion is the creation of new three-letter abbreviations. Marketing agencies hope that this will make them look more modern; additionally, if a given abbreviation catches on, they will have an advantage in positioning themselves.
For this reason, terms such as GEO and others were often criticized among specialists. There is no point in making clients (people ordering website promotion) learn a dozen different abbreviations.
GEO is quite a significant part of SEO; a website is promoted differently when language models are not considered at all.
However, the more I read about it, the more I come to the conclusion that GEO is such a significant part of SEO that it can already be considered a separate field. Especially since the methods models use to search the internet have developed considerably since these features were first made available. At the same time, not every company that promotes itself online needs specialized promotion aimed at language models.
What I Base This Text On:
I have some minimal experience promoting websites in language models. I did not do this intentionally; it just happened that some of the companies whose websites I developed were recommended by ChatGPT, which allows me to compare theory with practice. It is not extensive experience, but it is something.
The rest of the text is based on guidelines from OpenAI and Google, as well as our own tests.
GEO and SEO: Differences and Similarities.
The foundation of visibility in language models is visibility on the internet. And for many years now, visibility on Google has not been about simply adding keywords and links between pages. Search engine systems have been fighting spam and quick tricks for promoting websites from the very beginning.
As early as 2019, approximately three years before the launch of ChatGPT 3.5, Google introduced the machine-learning-based BERT algorithm, which, like language models, analyzed the text on a page and attempted to understand it. For a long time, a problem in website promotion was the so-called “SEO salad”—the forced insertion of keywords in order to achieve the appropriate density. For many years now, Google has understood not only synonyms but also the description of a given topic using different words.
So, if you write a text about “bicycle repair,” Google understands that passages about “pumping up tires, adjusting brakes, unscrewing bolts,” etc. are related to the title, and that the text delivers what the heading promises.
However, Google’s algorithm still does not work perfectly, and pages filled with keywords that offer little value are sometimes still displayed high in search results.
As language models become increasingly widespread, Google has had to take stronger measures and learn to distinguish texts written by people knowledgeable about a given subject from hastily produced articles generated with a single prompt in ChatGPT.
Thus, when it comes to Google ranking, writing meaningful content that addresses users’ needs (i.e. in accordance with Google’s recommendations) is becoming increasingly important, while trying to manipulate rankings by filling the site with content and building links on a massive scale is becoming less and less important.
It is also becoming increasingly important to build a coherent ecosystem beyond the website itself. Reviews on Google Maps and social media are gaining importance (company profiles as well as mentions among other users). The premise is simple: it is relatively easy to automate content creation on the website and link building; creating hundreds or thousands of fake social media accounts is more difficult.
SEO is not dead.
Language models have taken over part of the internet traffic. That’s a fact. In the past, when someone wanted to learn something, they searched for terms on Google (as well as on YouTube, TikTok, and elsewhere). After finding the right guides, some users went to the sales section.
Nowadays, when someone is looking for a solution to a particular problem, they increasingly start by talking to ChatGPT.
Still, when someone is looking for a specific product, they most often start by searching for its name on Google. Shopping assistants are still in their infancy. Dedicated programs are less popular, and simple queries in a regular language model are often not useful because the challenge is properly collecting data from the website (with a high risk of hallucinations). Admittedly, some people ask a language model to create a list of products, but this applies to specific cases. As a rule, the more expensive and complex the purchase, the more often people use the assistance of a language model.
However, the difference between searching with a language model and searching on Google is quite significant. Longer queries and two levels of difference.
First of all, people use a language model differently than they use Google. Queries submitted to a language model are longer, and the model also has a broader context. Here, we can distinguish between cases where the user asks it to search the web and cases where the model decides “on its own” to search the web in order to provide a better result.
Additionally, the language model uses longer, more precise phrases and initially selects which results interest it. In short, ChatGPT filters out SEO specialists’ tricks more effectively than Google does.
If the topic concerns issues related to science, ChatGPT assigns low priority to blogs and focuses on scientific research, institutional websites, etc.
How can you find out how ChatGPT searches the internet? The simplest way is simply to ask it.
This of course does not provide one hundred percent certainty, but it allows you to get a preliminary understanding. It is best to first ask it to search the web (or ask a question that will trigger a search), and then ask why it chose precisely those results and how it searched the internet.
Besides that, general queries such as “ChatGPT, how do you search the web?” are also useful.
I asked the model to provide a list of 10 articles about the simple repairs every driver should know how to perform (such as changing a tire, checking the oil level, etc.), and then I asked why it had chosen those particular articles.
Additionally, clicking the “sources” button displays a simplified reasoning process.
What did I discover this way?
The model first carried out internal reasoning, used the knowledge it had, initially searched the web, and divided the searches into several categories:
“1. General Basic Car Maintenance”
This was about texts that gather the basics in one place: fluids, tires, dashboard warning lights, the battery, inspections, and car equipment.
2. Emergency Items on the Road
That is, situations where a mechanic is not readily available: a flat tire, a breakdown on the road, a dead battery, or the need to mark the stopped vehicle.
3. Changing the Wheel and Tire
It’s one of the most important practical skills, so I looked for separate guides focused solely on it.
4. Operating Fluids
Here I separated the things a driver should be able to check or top up from the things that are generally handled by a service center.
- Car fluids: a guide
5. Tires, pressure, and tread
This is routine maintenance that the driver can perform themselves, without a mechanic.
6. Lights, Bulbs, Windshield Wipers
These are small things, but important for safety. I looked for them as separate topics because they are often overlooked in general guides.
7. Official and More Reliable Sources
I didn’t want to rely solely on shop or service blogs. That’s why, when researching safety and mandatory equipment, I also looked for websites of the police, institutions, or major technical brands.
“
Then I asked for specific logs (queries). They cannot be trusted 100%, but they are certainly not entirely fabricated. If the model used a given tool, it has relatively good access to what it did with it, although these are not complete logs. For some queries, it told me that it did not have access to the queries but could reconstruct them, which increases the credibility of those cases where it provided them.
Preliminary conclusions:
Long-tail keywords and answers to specific questions matter. The model often added the word “guide” to the search query, although of course it does not always do so, and it depends on what we are searching for.
When making product inquiries, ChatGPT uses the shopping layer:
The Shopping Layer in ChatGPT (a Separate Mechanism)
It is described in the official OpenAI guide, updated on April 24, 2026.
https://help.openai.com/en/articles/12911370-using-shopping-research-in-chatgpt
Shopping layer” is a colloquial name for an additional mechanism that activates when a question appears to be purchase-related. In other words, not merely: “what is an espresso machine?”, but rather: “which espresso machine should I buy for up to PLN 2,000?”, “compare these three models”, “find an inexpensive gift for a child”, “which laptop is suitable for work and gaming?
In such a situation, ChatGPT can show product cards in addition to an ordinary text response: images, names, prices, descriptions, ratings, features, and links to pages where users can check out or purchase the product. OpenAI explains that such results appear when a query suggests purchase intent, and products are shown with images, details, and links to sellers’ pages.
What is shopping research in ChatGPT?
Shopping research, or purchase research, is a more advanced mode than simply displaying products. It is intended for situations where making a choice requires comparison, trade-offs, and multiple criteria, e.g. “the best laptop for video editing under PLN 5,000,” “a quiet vacuum cleaner for a small apartment,” or “a gift for someone who likes cooking but already has basic equipment.” OpenAI describes shopping research as an interactive process of discovering and comparing products based on the user’s needs, preferences, and budget.
In this mode, ChatGPT may ask clarifying questions, such as about preferred brands, size, budget, style, comfort, performance, or price. The user’s answers narrow down the research and are intended to improve the relevance of the suggestions.
After a few minutes, this kind of research can produce something more akin to a mini buying guide than an ordinary product list: the best options, differences between them, trade-offs, strengths and weaknesses, up-to-date information, and links to retailers. OpenAI writes that shopping research searches the internet for current information such as prices, availability, reviews, specifications, and photos, and then creates a personalized buying guide.
Applications
This means that having a store on a website is becoming increasingly important (if we conduct sales). More precisely, having structured data is more important, and having a store and an SEO plugin helps with this.
Schema.org itself is not a magic way to ensure visibility, but it is a useful tool. It is important to remember that the description in Schema.org must be consistent with what the user sees.
How to Promote a Website in Language Models?
The research, journalism, and investigative aspect is becoming increasingly important when writing articles.
Texts written for company websites can be written in a typically marketing-oriented style. In such cases, the time spent writing the text is often calculated precisely. It is then useful to use secure solutions, and if something cannot be easily verified or clearly presented, the simplest approach is to omit it.
It is different in research work or serious journalism. Here, more time needs to be devoted, and the result of this work is not always immediately apparent.
Custom knowledge bases
Concrete takeaway: It is useful, for example, to create your own knowledge bases. The final text published on a blog is limited by its length and by the target audience it is written for. Creating a knowledge base allows you to use research conducted for one piece of content to create multiple articles. The point is not merely to speed up the writing process, but to fundamentally improve the quality of the content. Citing sources is also important.
This also translates into the way an SEO infrastructure is built.
A common practice is to build thematic portals that do not sell anything directly but link to company websites. This can be done in better or worse ways: sometimes these are spammy, automatically generated pages (although they have made less and less sense for many years), while at other times they are quite good SEO blogs containing useful, albeit derivative, information.
However, for ChatGPT, such portals have very little value. Unlike Google, ChatGPT does not create a single ranking of websites, and OpenAI also does not have its own data on website popularity. An SEO portal is therefore not considered useful if it does not contain genuinely unique content in terms of substance. These may even be good, enjoyable texts to read, but they are not particularly helpful to the model, because it is more convenient for ChatGPT to refer to original sources, especially since it can do so faster than a human.
For example, let us assume that, in order to promote technology companies, a portal about the technologies of the future, space, etc. was created. Such a portal can be supplied relatively easily with texts created on the basis of news from other websites. This can be done quite well, so users spend an appropriate amount of time on the site, and Google considers the site to be quite valuable.
ChatGPT, on the other hand, sees that all texts are derivative, and it is better to read directly: scientific research, news portals with technology sections, etc.
This of course does not mean that only scientific studies or texts written by credentialed journalists matter, nor that the only entity capable of providing SEO support can be a research institution. Although a language model can read long research PDFs, breaking a longer text down into shorter pieces concerning specific sections is also useful for LLMs. Quick texts created along the lines of “write a text based on this video about a meteorite that fell in Poland, and add a few examples of other news stories about meteorites” are not particularly valuable to ChatGPT. Even if the text itself was written manually, correctly, and without factual errors.
Things look different if, while creating the text, one has compiled, for example, some unique list—in this case, a list of the most recent meteorites to have fallen to Earth, or a list of cases in which the exact moment of a meteorite striking the Earth was recorded (which is very rare).
So how do you create GEO support sites?
Instead of creating a typical marketing-oriented thematic portal, it is better to create:
- Personal blog (relatively low cost)
- A niche portal based on a genuine community, with some content created, for example, by enthusiasts, in collaboration with organizations, creators who are well-known in their niches, etc.
- Collaboration with the media
- A somewhat more complex approach involves creating your own organizations, establishing associations, etc.
Of course, at this point there are no verified solutions; I wrote these points based on how ChatGPT evaluates sources.
So the entire approach is changing; SEO support could previously be created primarily to drive visits and page views. However, the cost of creating and developing an organization is so high that it would be beneficial if it also supported other needs along the way.
Social media and forums are becoming increasingly crucial.
Internet forums have lost popularity to social media and groups, but the development of language models may mean that they will become increasingly important. Even now, models like to retrieve content from Reddit and Quora, and citations from various forums are not uncommon either.
Content from Facebook or Twitter is treated differently because models do not have direct access to it; only some social media content is publicly available.
Reddit, despite not having the hierarchical structure known from forums, is much more convenient for searching than thematic groups on Facebook.
Promoting a Website in Language Models: A List of Tips.
General principles and operating philosophy.
-
Classic SEO. The page must be easy to find. However, ranking near the top is somewhat less important. People most often click the first result, and the top three positions are very important. A language model can check a dozen or so results and choose the one that fits best.
-
Don’t write for ChatGPT. Write for a person who uses ChatGPT.
There is no magic trick that increases visibility in language models. This principle also applies to conventional SEO; the time when writing for Google’s algorithms made sense has passed and will not return.
-
Authenticity matters. Avoid using tricks, mass-generating content, forcing phrases in, etc.
-
Promoting different industries and topics looks different. Therefore, it is worth asking a language model questions (in the way a potential customer searching for a particular product or service would). For example, if you run a hotel, it is worth asking ChatGPT about the best hotels in the area, travel itineraries, etc.
List of specific activities worth carrying out when promoting a website for GEO and AEO
-
Specific, up-to-date, well-structured, signed by the author, based on experience (or research), easy to quote, and supported by sources.
-
It is worth testing long-tail phrases, longer than those typically used for SEO, which can therefore consist of 4–8 words.
-
Dates indicating that the content is up to date. For example, on a company’s pricing page, it is a good idea to provide dates. In the case of articles, dates of updates should be provided.
-
Locations work similarly. Someone viewing a company’s locations can easily assess whether a given company is nearby, taking into account not only the straight-line distance but also the travel time.
-
Start the article with the answer, not with a long introduction. In traditional blogging, it is common to write a lengthy introduction to “ease the reader in.” For AI visibility, the following model works better: question → brief answer → elaboration → details → examples → FAQ.
-
Build articles around “one good answer.” Every piece of content should have a clear topic. Do not create an article that is about everything at once.
-
Write passages that can function independently
-
AI often does not quote the entire article, but uses an excerpt. Therefore, every important paragraph should be understandable without reading the entire page.
-
Take care of authorship and credibilityAI is more likely to rely on a page that looks like a credible source. This does not mean you have to be a major portal. But it should be clear: who wrote the text, why they know the subject, when the text was updated, where the data comes from, and whether the company actually operates in this industry. Google recommends showing who created the content, adding author bylines, and linking to information about the author, as this helps users understand the credibility of the content.
-
Create topic clusters, not individual random postsOne post is rarely enough to cover a given topic. It is better to create an entire set of related texts.
-
**Write so that your page becomes a “source of definitions”**Good topics for AI are those that organize concepts
-
A presence in many places beyond the website. In other words, building a brand. A person can check whether a company is legitimate by visiting its premises, calling the company, checking its National Court Register (KRS) details, or asking friends. Language models do not have this option. It is therefore worth taking care of your Google Business Profile (and genuine customer reviews), listings on other websites, listings on local portals and industry directories, etc., as well as a presence on social media (not only the company’s own accounts).
-
Social media are very important, especially those used for sharing information (such as Reddit, Wykop, and YouTube).
-
It is also worth developing a personal brand and maintaining non-business social media accounts. A company appears more authentic when the people behind it have their own social media profiles. Customer interaction is also important.
What Isn’t Gaming the System in SEO and GEO, Even Though It May Seem That Way?
Including dates in the title and phrases such as “step by step,” “guide,” “what is,” etc. is definitely good practice and is not simply trying to game the system (as long as the text is actually a guide). The same applies to properly planning headings.
I used to think that adding a date like this was forced (because it is one of the guidelines in SEO plugins), but adding a date is useful information, for example, because it indicates how up-to-date the content is.
The additional phrase must indicate the search intent. Therefore, doing SEO for typos makes less and less sense. However, the phrase “step by step” clearly defines the intent.
Text on the Internet differs from printed text. In the case of an article on a website, every heading, every piece of text, and every image can directly attract traffic to the site. In the case of printed text, people assess whether they want to read it based on the cover and description. Once someone has a book in their hands, they read it (or browse through it).
The Internet can be compared here to a vast card catalogue, where every paragraph in a book or newspaper has its own tab with a brief description. Thus, as they walk through the library’s aisles, users see not only the titles of texts, but all the headings.
Fun fact: How did ChatGPT search for content in the past?
Two years ago, I checked by asking ChatGPT to list the best bike rental shops in Szklarska Poręba. The model ranked the website of the rental shop I was promoting first, second, or third. Interestingly, in the case of other rental shops, it indicated a link to my article as the source, even though I had not mentioned those rental shops at all. I asked it why it had chosen those particular rental shops, and the model pointed to an article on my website reviewing the tourist infrastructure, in which I had not mentioned any of the rental shops it had listed.
Of course, the current ChatGPT handles this better, but other, less well-known models may still use simpler methods.
Conclusion:
The development of language models is forcing changes in traditional SEO. For several years now, it has been clear that stuffing content with keywords and links is becoming less effective, while content quality is increasingly important. However, when optimizing for ChatGPT, these changes need to be applied even more rigorously. This is evidenced by situations where a given website ranks highly on Google, but ChatGPT does not recommend it—and vice versa.
Although positioning for language models is an extension of activities and changes in current SEO, the difference is so significant that it necessitates carrying out entirely new tasks.
Sources:
https://help.openai.com/en/articles/12911370-using-shopping-research-in-chatgpt
https://www.similarweb.com/blog/marketing/geo/most-cited-domains-llms
Own mini research.