How does ChatGPT pick which businesses to recommend?

The Tasadr team · 22 September 2026 · 13 min read

ChatGPT reads the web before it answers, then writes one paragraph and cites two or three sources. It recommends your business when one page on your site answers the same question in Arabic, names the business, and is structured, current and reachable. The rest is measurement: ask, read who was cited instead of you, fix the page.

How does ChatGPT answer an Arabic buying question?
ChatGPT searches the web before it writes. It turns your customer's question into one or more searches, reads a small number of the pages that come back, writes a single answer in its own words, and puts links to the pages it relied on inside it. Those links are the citation, and they are the only route a reader has to your site.
A buying question changes the engine's behaviour. When someone in Riyadh asks for a florist that delivers the same day, the engine looks for pages carrying a city, a delivery time, a price and a business name, and prefers the one that has them together. When someone asks what a term means, a definition page is enough. So the questions you pick decide which kind of page gets read, which is the subject of how to choose the questions you measure on. The other surface that reads the web before it answers is covered in the guide to Google AI Overviews in Saudi Arabia.
The answer is not fixed. Ask the same question several times in one hour and you get different wordings, sometimes different sets of sources. That is how the engine works: it searches again every time, and the search results themselves move. So we measure every question repeatedly, one lucky answer is never counted as a change, and when a measurement spans a range we show you the range rather than one number that pretends to be stable.
What we measure here we call standing, because the point is not that someone sees you on a screen, it is that your site is the source the answer was built on. The work that gets you there is optimising for AI answers, closer to editing than to technical tricks.
Which pages does ChatGPT cite?
A citation is a link the engine puts beside a sentence to say that this fact came from that page. In our measurements on Arabic questions in September 2026, the pages that earn it share five traits, all of them buildable in a working day without a developer.
- They answer the question directly. The first sentence is the answer itself, not a welcome or a company history. The engine quotes the shortest passage that will do, so write that passage yourself.
- They name the entity. The business name, the city and the service are text on the page, not locked inside an image or carried only by the logo. The engine attributes an answer to a name it can read.
- They are structured. Subheadings, short lists, a price table, a FAQ. A quotable passage needs edges that start and stop, and one long paragraph offers no edges at all.
- They are current. A visible updated date, with prices and hours that match reality. A page still quoting last year's prices gets read and then dropped: the engine prefers material it can date.
- They sit on a domain the engine can reach. No crawler block in robots, no text that exists only after a full JavaScript render, no page behind a login or an email wall.
What is not on the list
Word count is not on it. Neither is keyword density or domain age. The page that wins the answer is not the longest, it is the clearest, and one precise page beats ten that circle the topic from a distance.
Nor is your number one ranking. Ranking puts your page in front of the engine; it does not decide the outcome. The engine reads the first results and picks whichever answers, passing over the first for the fourth because the fourth stated a price.
What is the difference between being mentioned and being cited?
We record one of four outcomes for every answer, and each has a fixed name:
- Cited: the engine wrote your name and put a link to your site in its answer.
- Mentioned without a link: it wrote your name with no route back to you.
- Absent: it wrote an answer and you were not in it.
- No answer: the engine wrote no answer to that question on that run.
Why a citation counts for more
The difference is practical, not verbal. A mention without a link builds familiarity with your name and leaves the road to your site cut: the reader learns you exist and goes back to searching. A citation is the visit, and the stronger signal: the engine treated your page as a source it read today.
So the Tasadr score is one number out of 100 saying where your site stands inside the engines' answers, with its range beside it whenever a measurement spans one. What each run gives you is in how we measure AI answers.
No answer is never counted against you. An engine that said nothing said nothing about you, and charging you for its silence would punish a site nobody was asked about.
Five steps from the question to the edited page
The method is one loop you repeat monthly, weekly or daily, depending on your plan. Do not start with the edit. Start with the measurement, or you are repairing what nobody established was broken.
1. Choose the questions your customers actually ask
The question is the unit of measurement, not the keyword. Write each one the way a customer types it on a phone: a full sentence, with the city and the condition that matters to them, such as same-day delivery or a price under a figure. Five precise questions are enough to start.
If you have Search Console, connect it and we suggest questions from real searches that reached your site. Without it we suggest them from your own pages, or you write them by hand. The measurement works either way.
2. Measure before you change anything
The first run is your baseline; without it you will not know whether an improvement was yours or the engine's. We ask every question on every engine your plan includes, record one of the four outcomes per answer, and keep the text exactly as the engine wrote it with the links it cited.
Keep that text: it is what you compare against in a month, and the best description of what your customer wants. The engine restates their question in clearer words than yours.
3. Read who was cited instead of you, and why
Open the pages the engine cited where you were absent, and read them the way the engine does. Where is the answer? Does it carry a number, a time, a city? In most cases the gap is not quality, it is ordering and plain speaking.
The competitor analysis does this for you: it puts your page beside the cited rival's and says what is on theirs and missing from yours. How many competitor analyses you get a month follows your plan.
4. Edit one page, and make the edit visible
Take the highest-value question and the one page that answers it, and change four things: put the question itself as a heading, write a two-line answer under it carrying the name, the city and the number, add a price or hours table, and add an updated date.
Do not edit ten pages in a day. When the score moves you will not know which edit moved it, and you will repeat the one that did nothing.
5. Measure again, and trust confirmed change only
Let a full cycle pass, then read the difference. One changed answer is not news: chat engines fluctuate, which is why an alert reaches you only when a change is confirmed. A rival is not called new until the change holds, a citation is not called lost until the absence holds, and a gain counts at once.
So the headline and the alert do not flicker every morning: whatever reaches you as an alert has already held.
Try this on your own site
Add your site, write 5 questions your customers ask, and see who leads the answer today. No credit card.
Start free
An example from our September 2026 measurements
We publish no client numbers and invent no case studies. What follows is what we saw ourselves when we asked AI engines Arabic questions from the Gulf and Egypt in September 2026.
- Google AI Overviews writes a great many Arabic answers citing nobody. We record them as they are: real answers in which neither you nor your rival was cited, and we show them to you rather than hide them.
- On Gulf Arabic questions we asked in September 2026, Google AI Mode answered every one, and some answers carried no source at all. An answer without a source has not been won by anyone yet, neither you nor your rival.
- On a question from Cairo, Grok answered and cited a restaurant-listing site rather than the restaurant's own site. Listings win because they are built in the shape of an answer: name, district, cuisine, in one quotable line.
The lesson we take from it
Engines prefer the page that answers in one line over the page that introduces itself in a paragraph. Directories win because they were built that way from day one, and you can build one page on the same logic for every question that matters: question as heading, answer under it, detail after.
To be plain: these are observations from a limited set of measurements on questions we chose ourselves, not a law. We measure again every month and correct what we say when the numbers change. The eight engines we ask are described on the AI engines we measure in Arabic.
The common mistakes that keep you out of the answer
Six common mistakes, all fixable in a week:
- A home page that explains everything and answers no single question, so the engine finds no passage worth quoting.
- The name, the city and the service living inside an image or a logo. The engine reads text, not design.
- Prices and hours replaced by an invitation to get in touch. A buying question wants a figure, and whoever writes it wins it.
- Content that exists only after a full JavaScript render, or a site that blocks crawlers and then wonders why nothing cites it.
- Measuring in English for a market that asks in Arabic, which gives you correct answers about the wrong market.
- Judging from a single answer. One answer is not a measurement, and a different wording between two requests is normal.
What nobody can promise you
Nobody can guarantee a citation. ChatGPT picks its sources afresh on every request, and no panel sells a place inside its answer. What we can do is measure regularly and tell you where you stand, who took your place, and why their page is clearer.
The limits are stated too: the Tasadr score is ours, not OpenAI's and not Google's; a chat engine words its answer differently every time, so every question is measured repeatedly and one lucky answer is never counted as a change; and we do not promise visits, we measure your place in the answer. Every plan and price is published on pricing in Saudi riyal.

Frequently asked questions