ChatGPT and Google AI build a recommendation from two sources. The first is what the model learned up to a certain date, and the second is what it finds by searching at the moment of the question. When a question is about businesses and services, search runs much more often. Then what matters most is what is written on the website of the business, on its Google profile and in texts others write about it. A place in such an answer cannot be bought or guaranteed, and the same question asked several times on the same day often gives different businesses.
Do ChatGPT and Google AI search the internet?
One Serbian question is “kako da me chatgpt preporuči”, meaning “how do I get ChatGPT to recommend me”. The Google AI Overview answers it by saying that “AI models do not search the internet in real time in the same way as classic search engines”. That is only partly true.
OpenAI writes in its description of ChatGPT search that ChatGPT starts a search on its own when an answer needs fresh information. Google states for AI Overviews and AI Mode that they take pages from the same pool used by regular search, through the same systems. Both features are officially available in Serbia, in Serbian in both Latin and Cyrillic script.
When there is no search, the model answers from what it learned earlier. The model ChatGPT uses for quick answers knows only what existed up to April 2026. In a large Ahrefs study, the free version of ChatGPT searched the internet in only about one in seven answers. For many questions, the chance of a business therefore depends on what the model already knows about it.
Questions about price, reviews and specific businesses still call for search more often. In the test of Serbian questions described below, ChatGPT searched in eight of nine answers. The only time it did not search, it failed to recognize a small Belgrade agency by name and guessed it was a medicine or a supplement. A small or new business therefore cannot count on the model already knowing it, and search reads only what has been published about the business.
“ChatGPT may search the web automatically when your question would benefit from current information.”
What does ChatGPT do with a question about a business?
ChatGPT rewrites the question of the user into one or more shorter queries and sends them to search engines. OpenAI gives the example of a question about good restaurants nearby, which together with the approximate location of the user becomes a query with the name of the city. If memory is on, ChatGPT also uses what it remembers about the user, so two people with the same question can get different businesses. Google AI splits a question into several related searches in a similar way.
ChatGPT can also check the website of the business itself. In the same Ahrefs study, when it limited a search to one website, that was most often the website of the business in question. For questions about price, services and business details, ChatGPT cited at least one official source in more than four of five cases.
In the test of Serbian questions, ChatGPT stated the basis for its choice itself. It wrote that it singled out agencies for an initial comparison “based on the available business profiles, reviews and service descriptions”. Next to each business it showed the rating, the number of reviews and the address. Then it asked about the line of business and the part of the city to narrow the choice. The final recommendation thus changes during the conversation, according to what the user says about themselves.
Why do articles by others about a business matter more than links?
A large Ahrefs analysis compared how often brands appear in AI answers with various measures of how well known they are. The brands that appeared most often were those mentioned a lot by YouTube channels and other websites. The number of links to the website had almost nothing to do with it. That does not prove that mentions alone earn a place in the answer, but it shows the direction.
Google writes that its AI features show what is said about a business on blogs, in videos and on forums. It adds that artificially commissioned mentions do not help as much as they seem to. Muck Rack shows the same after reviewing more than 25 million links from the answers of the assistants ChatGPT, Claude and Gemini. The great majority of citations were texts others had written about businesses, and paid and advertising content almost never appeared. Muck Rack sells a public relations tool, so this finding too should be read with that in mind.
A citation is still not the same as a recommendation. Lily Ray tracked lists of the “best” in which a business puts itself in first place. When the Google AI Overview cited such a list, it left the business that wrote it out of the recommendation in more than two thirds of cases.
The same was seen in Serbia. The website behind the most cited Serbian list of SEO agencies, on which the author ranked its own agency first, was often cited by Google AI. As an agency, however, the author was not recommended once. The AI recommended other agencies from that list. A guide by a Belgrade agency that neither names nor ranks competitors did better, but only in AI Overviews and only when the question was in English.
What did a test of 25 questions in Serbia show?
In October 2026, Atrigen asked Google AI Overviews and AI Mode 25 questions, each three times, and asked ChatGPT some of them too. These were questions people actually type, from short queries such as “seo agencija” to questions about price and about specific businesses. The browser had no login and was set to Serbia, and other AI assistants were not tested.
The Google AI Overview appeared on every search, even on short queries such as “seo agencija” and “izrada sajta beograd”. For Serbian questions about services, an AI answer thus almost always sits above the first regular result.
The websites of the businesses themselves were cited most. In AI Overviews they made up roughly three of four citations, most often the home page, a service page and the price list. Articles, guides and agency catalogs were much rarer, and Reddit appeared as a source on only one question.
The Google profile of a business was a source in roughly every third AI Mode answer. For one business, the AI carried over the profile description almost word for word. The businesses ChatGPT recommended had almost perfect ratings, and some only a few reviews. The sample is small, but it shows that a business with few good reviews was not excluded.
“Google Business Profiles can help your products and services to be visible in both AI responses and other Google Search results.”
Why is there no fixed first place in an AI answer?
OpenAI writes that ChatGPT ranks results by several criteria and that placement in the results is not guaranteed. Tests also show why. In the Serbian test only about one in four sources in AI Mode appeared in all three repeats of the same question, and more than half only once. For the question “digitalna agencija beograd”, AI Mode named three almost completely different groups of agencies across three repeats.
Tests abroad give the same picture. In research by SparkToro, where the same questions were asked thousands of times, the same list of brands repeated less than once in a hundred answers. SISTRIX tracked the same queries for months and found that Google replaces more than half of the sources in AI answers every week, and ChatGPT almost three quarters.
ChatGPT says so itself. In the Serbian test it described its list as a shortlist based on publicly described services and added that it was not an independent ranking of quality. An offer of a “guaranteed first place” in AI answers therefore promises something not even OpenAI promises.
Visibility is therefore measured as a percentage. A fair test takes 20 to 30 questions the way customers actually write them. Each question is asked at least three times, in a new conversation and without login, and then the answers that include the business are counted. A small difference between two tests can easily be chance. Google also offers a report on performance in AI features in Search Console, its free tool for website owners. It also writes that no outside tool has insight into how Google ranks.
“Visibility % across dozens to hundreds of prompts run multiple times is a reasonable metric.”
What can a business do, and what counts as spam?
The first step is technical. OpenAI states that a website must let in the program ChatGPT uses to search websites and that hosting must not block it. A blocked website does not show up in answers. Google asks for no extra requirements for AI Overviews and AI Mode and writes that preparing for AI search is still ordinary SEO, that is preparing a website for search.
The second step is the content of the website. Services, city and prices should be written as plain text, on a page search can find. ChatGPT most often looks for price and service details exactly on the website of the business. In the Serbian test, for a question about the price of a website the AI cited several different price lists and texts about price. An example of such a text is How much does a website for a small business in Serbia cost?
The third step is the Google profile. Google writes that local results depend on how well the business matches the search, how close it is and how well known it is. A better place cannot be paid for, and according to Google a Business Profile helps visibility in AI answers too. An accurate category, address, opening hours and real reviews help a business get into the local results the AI chooses from. This is covered in more detail in the text How can a business show up better on Google Maps? No test has yet proven that extra reviews alone raise an AI recommendation.
The fourth step is getting others to write about the business. Texts by local media, associations, partners and clients, as well as videos, follow AI visibility more closely than links, and they come from work worth paying attention to.
Some things do harm. In May 2026 Google added to its spam policies that an attempt to manipulate AI answers in Google search is also spam. A separate page for every variant of a question, made to influence AI, breaks the rule on content made in bulk. Hidden text on a page has long been spam. Microsoft security researchers also described AI recommendation poisoning. These are “Summarize with AI” buttons that secretly write an instruction into the memory of an assistant to recommend a certain business. The researchers found such instructions at about thirty companies.
Fake reviews are also a legal problem. The Consumer Protection Law from 2026 counts posting fake reviews and commissioning such reviews from others as a misleading business practice. The fine for a company can reach two million dinars.
All in all, an AI recommendation is not built by one trick, but by what is written about the business on its website, on its Google profile and by others. The basis of everything is a website that Google and AI can read, and how a website is prepared for search is described on the SEO optimization page.