How ChatGPT finds its sources and recommends your business

ChatGPT is a conversational AI assistant and search engine developed by OpenAI. It understands questions written in natural language, Arabic included, and builds complete, direct answers instead of simply listing links to pages. When a user asks a question today, the system draws on one of two sources of information, depending on the question. It either answers from the language knowledge stored in its model in advance, or it searches live web pages at the moment of the question and builds its answer on recent sources whose links it places in front of the reader.

The letters GPT in the name stand for "Generative Pre-trained Transformer". It is the large language model that OpenAI built to understand the context of human speech, connect concepts to one another and write text in correct language. In this first route, OpenAI collects public web text with its crawler GPTBot to train its foundation generative models. The model gains a broad body of knowledge that lets it explain general concepts, summarise texts and help users write and research. But whatever it stored during training is limited to the time it was collected. On its own, it is not enough for an accurate answer to questions in a market that keeps changing, such as what a product costs today, when a business opens, or which local business to recommend in a particular city.

This is where the second route comes in, the one that matters to every site and store owner: a live search of the web through ChatGPT's search feature, which relies on retrieval-augmented generation. When your customer asks ChatGPT about a service in Riyadh, Dubai or Cairo, the engine does not rely only on what it learned in its old training. It goes out to the web at that moment to look for real, recent pages, which its search tool reads through the crawler OAI-SearchBot. It then summarises what it found on those pages and attaches their links as sources. Microsoft Advertising published guidance in October 2025 on optimising content for inclusion in AI search answers, and it points the same way: answer engines are built to draw on clear, specific pages from the web.

For a business owner, this split between trained memory and live search has a clear meaning. Your chance of being recommended does not depend on waiting for OpenAI to train a new model in months or years. Because the engine searches the live web when a buyer asks a commercial question, editing your service or product page today, and adding clear numbers and facts to it, changes what the engine reads when it looks for an answer for your customer.

ChatGPT answers an Arabic buying question by searching the web before it writes a single word. It turns your customer's question into one or more searches, reads a small number of the pages that come back, writes one answer in its own words and places links to the pages it relied on inside it. Those links are the citation, and they are the only direct route from the chat screen to your website.

When a buyer types a question, the engine runs a process known in modern search as query fan-out. It breaks the user's compound question into searches that cover each of its conditions, so it can collect candidate pages to read. This is where what we call a buying question, explained in our glossary of GEO terms, makes a decisive difference. A buying question changes the engine's behaviour and the kind of pages it picks, compared with an ordinary informational question. When someone in Riyadh asks for a florist that delivers the same day, the engine looks for pages that bring together four elements: the city, the delivery time, the price and the business name. It prefers the page that puts all of them in one clear place. When someone else asks what a general term means, one definition page is enough, and it needs no city and no price.

So the questions you choose to analyse decide which kind of pages the engine will read on your site and on your competitors' sites. There is a separate guide on exactly this, how to choose the questions you analyse your standing on. Google's own AI summary also reads the web before it writes the summary at the top of the results page, and its mechanism is explained in the guide to Google AI Overviews in Saudi Arabia. You can also see how the Gemini chat engine picks its sources in how Gemini chooses its sources in Arabic and how it cites your site.

As you follow your site, keep in mind that ChatGPT's answer is not fixed. Ask the same question several times in one hour and you will read different wordings, and sometimes different sources from one request to the next. That is not a fault in your site. It is how the engine works, because it searches the web again every time and the search results themselves move and reorder with context. A detailed explanation of this change is in why ChatGPT's answer changes, and when the change is real. For this reason one lucky answer is never counted as a change, and when an analysis spans a range between two numbers we give you the range rather than one number that pretends to be stable.

What we analyse here we call standing. For your business the deciding point is not that your name flashes briefly across a screen. It is that your site is the source the answer was built on and the one that received the link. The practical work that takes you there is optimising for AI answers (GEO), which is closer in essence to editing content and organising facts in the buyer's language than to complicated technical tricks.

ChatGPT chooses pages that give it a plain answer at the very start, name the business, its city and its service as text, show current prices and hours in a tidy structure that is easy to quote, and stay open to its search crawler. A citation is the link the engine puts next to a sentence in its answer to show that the fact came from that particular page. In our analyses of Arabic questions in September 2026, the pages that earned this link shared five practical traits, and you can build most of them by editing text in one working day, and the robots.txt and JavaScript points are for your developer.

• They answer the question directly. The first sentence on the page is the answer itself, not an introduction about the business's history and not a welcome paragraph. What we see in the analysis is that the engine usually quotes the shortest passage that settles the answer, so write that passage yourself at the top of the page. The guide on how to write a paragraph an AI engine quotes in its answer walks through the steps.
• They name the entity as text. In search, an entity is the specific identity of your business. On the pages the engine cited, the business name, the city and the service were written as plain text inside the page, not drawn inside an image or carried by the logo alone. The engine ties an answer to a name it can read in the text, and it does not guess who owns the page.
• They are well organised. The page has clear subheadings, short lists, a price table and a FAQ. The engine quotes a passage with a clear beginning and end, while a page written as one long, tangled paragraph does not give it a passage ready to quote.
• They are current and dated. The page shows a written update date, and its prices and opening hours match reality. A page that leaves last year's prices untouched is often read by the engine and then dropped in favour of another, because it tends to prefer what it can date and check for freshness over what it cannot age.
• They are open to the engine, with no barriers. They do not block the OAI-SearchBot crawler, the one that reads for ChatGPT search, in robots.txt, because OpenAI says that sites which block it are not included in ChatGPT search answers. Their main text is also not built only after several seconds of JavaScript running, and they do not sit behind a login or a window asking for an email address before reading.

A citation with a link places your page's link among the sources of the answer and takes the reader to your site directly. A mention without a link only writes your business name inside the text of the paragraph, with no link connecting the buyer to your page. The difference is practical and affects your marketing results. It is more than a matter of wording. A mention without a link makes the reader familiar with your name but leaves the road to your site cut, so the customer learns that you exist and then has to go back to searching. A citation with a link brings the visit at once. It is also the stronger signal that the engine treated your page as a live source it read today, not as old information it remembers from before.

That is why Tasadr records one of four exact outcomes for every answer the engine writes. Each has a fixed name that does not change from one report to the next, so you can read your site's position with complete clarity.

You let ChatGPT's crawlers read your site without obstacles by allowing the search crawler OAI-SearchBot to reach your pages in robots.txt, which is the plain text file that tells crawlers which parts of your site they may visit. You also write your main text and prices in the page's direct HTML code, not hide them behind JavaScript or login windows. The direct reason is that the engine cannot quote a paragraph or link a page that its search crawler was blocked from entering, or that it reached and found empty. Opening the technical road for the search crawler is therefore the first condition, before any edit to your wording.

According to OpenAI's official documentation of its crawlers, OpenAI uses three different bots to deal with websites. Each has its own separate job, and each is a separate decision that the site owner controls in robots.txt.

You take five steps in a clear working loop that runs from the question the buyer types to the edited page. You choose five buying questions, analyse where you stand on them now, read the pages ChatGPT cited instead of you, edit one page clearly, and then analyse again to confirm the result. The rational rule is not to start with random edits before the analysis. Otherwise you repair in your site what nobody has established is broken, and you miss the question on which your business is really losing the recommendation.

1. Choose the questions your customers really ask in Arabic. The unit of analysis in AI answers is the full question, not the single keyword. Write each question the way your customer types it on their phone, as a full sentence that includes the city and the condition that matters to them, such as same-day delivery or a price under a certain figure. Five precise questions are enough to start. If you have a Google Search Console account, connect it and we suggest questions from real searches that brought visitors to your site. Without it, we suggest them from your site's pages, or you write them yourself. The analysis works well either way.
2. Analyse before you change anything on your site. The first analysis is the baseline against which you compare every later step. Without it you will not know afterwards whether an improvement came from your edit or from the engine's usual movement. In this analysis we ask every question on every engine your plan includes and record one of the four outcomes for each answer. We also keep the excerpt that names you or your competitor from the answer, word for word as the engine wrote it, together with the links it cited. Keep this text. It is the evidence you will come back to in a month to compare, and it is also the best description of what your customer wants, because the engine restates their question in clearer words than yours. If you would like to start on your own before using any tool, the manual check is set out step by step in how to know whether your site leads ChatGPT and Gemini answers.
3. Read who was cited instead of you, and find out why. Open the pages the engine cited on the questions where you were absent and read them through the engine's eyes. Notice where the answer sits on the page, in which line and in which words, and whether it carries a clear number, a delivery time or a city name. In most cases you will find that the gap is not in the quality of the business but in ordering and plain speaking. The guides how to know which competitor ChatGPT recommends instead of you, and why and how to analyse your competitors in Google and AI answers explain how to read these gaps. In Tasadr, the why-they-lead analysis does this work for you: it compares your page with the page of the competitor that was cited and tells you what is on theirs and missing from yours. How many such analyses you get each month follows your plan.
4. Edit one page, and make the edit clear. Pick your most important buying question and the one page on your site that answers it, and change four specific things. Put the question itself as a heading. Write a two-line direct answer under it that carries the name, the city and the number. Add a table of prices or hours. Add a clear update date. Do not edit ten pages in one day. When the number moves afterwards you will not know which edit moved it, and you will repeat the one that did nothing.
5. Analyse again and trust confirmed change only. Let one full analysis cycle pass after your edit, then read the difference. A single answer that changed means nothing, because a chat engine swings from one request to the next. That is why an alert reaches you only when a change is confirmed. We do not call a competitor new until the change holds, we do not announce a lost citation until the absence holds, and a gain counts at once. So your dashboard does not flicker every morning, and whatever reaches you as an alert has already held. The factors that set the pace after your edit are in how long GEO takes to change the results.

We publish no client numbers and invent no case studies. What follows comes from a real report for an electronics store in Egypt, analysed between 12 and 26 October 2026 on 15 questions its buyers ask in their own words, with every name hidden.

• ChatGPT answered all fifteen questions and cited the store's site on 3 of them, among them the price question «كم سعر تكييف حصان ونص في مواقع التسوق المصرية؟» ("how much does a 1.5-horsepower air conditioner cost on Egyptian shopping sites?"). The site was absent from the other twelve.
• On that question ChatGPT wrote a price list of 5 products, from 17,000 to 26,500 EGP, all from one online marketplace, and sent the buyer to the store in none of its 8 links. On the washing-machine price question the store was in 3 of 5 links.
• The price question the store won on ChatGPT it also won on Gemini, but not on Perplexity, so one engine's result does not tell the whole story.

Six common mistakes in how content is written and how a site is set up stop ChatGPT from citing your pages and building its answer on them, even when your business is strong and well known in its local market. The good news for a business owner is that none of these six requires rebuilding the site from scratch. All of them can be fixed in one working week once you pin them down.

• A home page that explains everything and answers no single question. Some sites gather all their services and branches on one home page full of general welcome phrases and never dedicate pages to the specific questions buyers ask, so the engine finds no focused passage to quote and link.
• The name, the city and the service living inside an image or a logo. When you put your business name, the city you serve and the details of your packages only inside designed images or banners, the engine cannot pick up these facts and tie them to your entity. The search crawler reads the text written on the page, not the design.
• Prices and hours replaced by an invitation to get in touch. A buying question by its nature asks for a figure the customer can compare before buying. When a page hides prices or the cost range and says only "contact us for the price", the engine skips it for a competing page that states the figure openly, and whoever writes the figure wins the question.
• Content that exists only after a full JavaScript render, or a site that blocks crawlers. A site whose text can be read only after JavaScript has fully run in the browser, or one that blocks OAI-SearchBot in robots.txt, shuts the technical door in the engine's face and then wonders why nothing cites it in the answers.
• Analysing in English for a market that asks in Arabic. Some marketing managers in Saudi Arabia, the UAE and Egypt rely on tools that test questions in English only. They end up with correct answers and sources about a different market from their own, because a buyer who asks in Arabic triggers an Arabic search that retrieves very different pages.
• Judging from a single answer. Opening the chat window and asking once is not an analysis you can base a decision on. A change in wording, or in the order of some sources, between two consecutive requests is normal in how the engine works, and on its own it signals neither a lasting gain nor a fixed loss.

Nobody can guarantee you a steady citation or a reserved seat in ChatGPT's answers, because the engine searches the web and chooses its sources afresh on every request, according to the clearest pages available at the moment of the question. There is no console today where anyone can buy a guaranteed place among the organic sources the engine cites inside its paragraph. What we can honestly offer is to analyse regularly and tell you where your site stands today, which other site took your place, and why its page was clearer to the engine than yours.

ChatGPT's paid ads are something quite different from the organic citation of your site as a source of information. The practical difference between the two routes, and which one serves your marketing goal, is explained in detail in the guide ChatGPT ads or being cited, what is the difference and which do you need.

The limits of the analysis itself are stated openly too. The Tasadr score out of 100 is our own analytical score. It sums up your site on the questions of your market and lets you compare your performance before and after an edit, and it is not an official score issued by OpenAI or Google. Because a chat engine words its answer differently every time, we separate chance from a lasting result, and we do not promise a specific number of visits. We analyse your place inside the answer and its sources.

Tasadr is an Arabic platform for competitive analysis in AI search. It accompanies you on a complete journey from analysis to execution until you lead its answers. You can test your site's position yourself with a free 7-day trial, with no credit card, that covers one site and up to 10 Arabic buying questions on ChatGPT and Gemini. All plans, their limits and their prices are published openly on the pricing page.