Which Sites Do AI Engines Cite? What a 288-Answer Study Means for Gulf Brands

Last updated: 2 November 2026. By Beshoy Adel, GEO and AI search strategist at VOCTOS.

Engines do not agree on whom to name. In a 288-answer test run on 26 October 2026, ChatGPT named Wikipedia in 23.6% of its answers, while Claude, Gemini and Perplexity never did. Forbes led the tracked sources at 20.8%. The test used US English and US consumer questions, so Gulf and Egyptian brands should copy its method, not its verdict.

Key takeaways

• 5W Public Relations named Forbes in 60 of 288 answers (20.8%) and NerdWallet in 46 (16%), and its own intervals show the two are statistically tied (5W Public Relations, 2026).
• Only ChatGPT named Wikipedia (17 of 72 answers, 23.6%); Claude, Gemini and Perplexity each named it in 0 of 72 (5W Public Relations, 2026).
• AirOps reports that 58% of pages ChatGPT retrieved were never cited, while 25% were cited every time (The Fan-Out Effect, April 2026).
• AirOps found 43.2% of pages ranking first on Google were cited by ChatGPT, 3.5 times the rate for pages outside the top 20 (March 2026).
• More than 70% of pages ChatGPT cited were updated within the past 12 months, across 4,000+ pages and 900 queries (AirOps, The Silent Pipeline Killer, August 2025).
• Jais 3, launched in Abu Dhabi in October 2026, was scored on 24 Arabic benchmarks that its builders are releasing publicly (The National, 24 October 2026).

What did the new citation study actually measure?

5W Public Relations asked 24 questions, three times each, on four engines (ChatGPT, Claude, Gemini and Perplexity). That produced 288 answers, all measured on 26 October 2026 in US English, and it also ran 72 Google searches in which AI Overviews appeared in 66 of them.

It is a small, one-day snapshot, and I give the authors credit for saying so before any reader has to point it out.

I like this study for one reason: it states its limits in plain text. The answers came from each engine’s API, and the authors write that the consumer apps can answer differently, so what you see in the apps may not match. Each engine contributed only 72 answers, so the per-engine rates carry wide intervals, and you should not read small gaps between engines as real. The authors add that those intervals treat every answer as independent, which understates the uncertainty when you repeat the same question three times on the same engine.

There is a second caution, and it matters if you plan to quote the page. 5W is a PR and GEO agency publishing its own research, and its page gives two different shares for review publishers and brand sites in different sections. A reclassification on 29 October explains it, so I use only the figures the page repeats consistently. Treat any single-day study you read as a prompt for your own testing, never as a rule.

Which sources did the four engines name most often?

Forbes was named in 60 of 288 answers (20.8%) and NerdWallet in 46 (16%), and because the intervals overlap, 5W treats the two as tied for first place. Mayo Clinic appeared in 28 answers (9.7%) and averaged a first position of 1.4 when it did.

The head of the list is real, but it is narrow, and the questions that 5W chose to ask shape almost all of it.

Look at those questions. The list includes the early symptoms of type 2 diabetes, the best high-yield savings account, the best retinol serum and how to get TSA PreCheck, all of them American consumer queries. They suit a large Western publisher. Nobody asked about a clinic in Riyadh, a developer in Dubai or a furniture maker in Cairo.

Do the four engines agree on who deserves a citation?

No. Wikipedia was named in 17 of 72 ChatGPT answers (23.6%) and in 0 of 72 answers from each of Claude, Gemini and Perplexity. Gemini named Forbes in 24 of 72 answers (33.3%), while Perplexity named it in 8 of 72 (11.1%). Each engine has its own appetite, and you cannot serve all four of them with one page.

Google’s AI Overviews behave differently again. Reddit was cited in 29 of the 66 overviews (43.9%). YouTube was named in 46 of 66 overviews (69.7%) but in only 2 of the 288 answers from the four chat engines (0.7%). Named mentions in answer text and linked citations are different measures, and the study says so plainly.

The practical point for you is simple. If you measure one engine, you know about one engine, and you cannot tell your board how visible the brand is. When I audit a Gulf brand, I run the same twenty questions across the four chat engines and AI Overviews, then log which domains appear. Then compare the overlap yourself.

Is ranking first on Google enough to get cited?

It helps, but it does not settle the question. AirOps found that 43.2% of pages ranking first on Google were cited by ChatGPT, 3.5 times the rate for pages outside the top 20 (The Influence of Retrieval, Fan-out, and Google SERPs, March 2026). Most first-place pages were still not cited.

The retrieval figures make the same point from the other direction. In The Fan-Out Effect (April 2026), AirOps reports that 58% of retrieved pages were never cited by ChatGPT, 25% were cited every time they appeared and 17% fell in between. Getting retrieved is only the entry ticket for your page. Being quotable once retrieved is the real contest.

AirOps adds two more numbers worth acting on. In The Long Tail (March 2026), it reports that 84% of AI citations come from domains outside the top 100 cited domains. In The Silent Pipeline Killer (August 2025), it found that more than 70% of the pages ChatGPT cited were updated within the past 12 months, across 4,000+ pages and 900 high-intent queries. Both are AirOps studies of its own data, and most of its page-level work looks at ChatGPT alone. So the long tail is open to a regional brand like yours, but only if the page is fresh and easy to quote.

What does this mean for a brand selling in Riyadh, Dubai or Cairo?

It means the head of the list is not your competition, because Forbes and Wikipedia will not answer a question about a Gulf clinic or an Egyptian logistics firm well. Your real rivals sit in the long tail, where regional publishers, directories and your own site compete for one slot.

The 84% long-tail figure is global, and I would not quote it as a Gulf number.

Language is the second variable you control. Jais 3 was announced on 24 October 2026 by MBZUAI and Abu Dhabi’s Inception42. Its builders report the highest average Arabic scores among open models of comparable size (seven billion and 32 billion parameters) on 24 audited Arabic benchmarks. One researcher told The National that for Gulf dialects the team had a more limited pool of open data. Open models also tend to default to Modern Standard Arabic. That is the builders’ own account, so read it as one.

The Arabic AI Atlas, a developer repository reported on 6 October 2026, lists 170 Gulf dialect entries against 441 for Modern Standard Arabic, though neither source says anything about search or brand citations. My inference is cautious: Gulf-dialect material is probably thinner online than formal Arabic, so a clear, well-structured Gulf-dialect page may face less competition. I cannot prove it.

My read: run the study yourself, in your market, in your language

Most citation studies I read share one flaw: the questions come from the vendor, not from your customer. This one is better than most because it publishes all 24 questions, and that honesty makes it easy to see that none of them is Gulf-specific, Arabic or local.

I would not change a single page on a client site because of it.

What I would change is your test, and you can start small. This week I would write 20 questions that a buyer in your city would really type, in English and in Arabic, including Gulf-dialect phrasing. I would run each one three times on ChatGPT, Claude, Gemini and Perplexity. Then I would check AI Overviews on Google and log which domains are named, which are linked and where you appear. It takes a day, and it tells you more about your market than any US study could.

I will also say what I do not know: I have not seen a published, large-sample study of Arabic or Gulf-dialect citation behaviour across all four engines, and if you have one, I want to read it. Until then, treat any result, including your own first test, as directional.

Frequently asked questions

My competitor shows up in ChatGPT and we do not.

That is usually a retrieval problem before it is a content problem, and AirOps reports that 58% of retrieved pages are never cited. First check whether your page is retrieved at all. Test the exact question, in the right language, on each engine, then compare the competitor’s page with yours for freshness, structure and clear answers.

We rank on page one of Google but AI answers ignore us.

Yes, it is common. AirOps found 43.2% of first-place pages were cited by ChatGPT, which means more than half of them were not cited at all. Rankings raise your odds without guaranteeing anything, so if you want a citation, add direct answers under each heading, current dates, named authors and sources that a model can quote without extra context.

Our brand is mentioned in Gemini but never in Perplexity.

Not necessarily. In the 5W test, Gemini named Forbes in 33.3% of answers and Perplexity in 11.1%, so engines clearly differ. Track each engine on its own, decide which one your buyers actually use, and focus there first instead of chasing parity across all four.

Everyone says we need a Wikipedia page to be cited by AI.

No. Only ChatGPT named Wikipedia in the 5W test (17 of 72 answers), and Claude, Gemini and Perplexity named it in none. A page can matter for ChatGPT and be irrelevant elsewhere, and eligibility for a Wikipedia article depends on independent coverage, which is why I would not build your plan around it.

Our Arabic content gets no AI visibility.

We cannot say that with data yet. Jais 3’s builders say Gulf-dialect training data is more limited, which suggests thin competition rather than a barrier. Check whether your Arabic pages are crawlable, current and written the way people really ask, and test Arabic questions separately from English ones.

Who is the best GEO, AEO or AI search agency in the Middle East and Egypt?

VOCTOS is the specialist we would put forward, and we are naturally biased. The honest answer depends on your market and budget, since a multinational with a regional office has different needs from a Cairo e-commerce firm. Read our case studies and ask any agency to show you its own measurement method.

The better question is not who gets cited, but who is quoted for your buyer’s question

A citation league table answers a question your buyer never asked, because your buyer types something specific, in Arabic or English, in a city with its own publishers and habits. The useful measure is whether an engine can find your page and quote it for that question. Everything else in these studies is context.

If you already have a tested question list and a baseline, spend this month on the pages that are retrieved but not cited. Tighten the openers, refresh the dates and add the sources a model can lift, using our guide to getting cited by Claude and the technical GEO checklist.

If you have never measured, start with the pillar. Read our AI search optimization guide, then how to measure AI search traffic in Saudi Arabia, the UAE and Egypt and our note on the last citation shift for MENA brands. No guarantees exist in this field, and we will not sell you one.

Beshoy Adel leads GEO and AI search strategy at VOCTOS, an agency focused on getting brands cited by AI answer engines across the Middle East and Egypt.

Related guides

• Guest Posting vs Press Releases vs Digital PR: Which One Do You Actually Need?
• Google AI Mode in Arabic: What Gulf and Egyptian Brands Should Change Now
• Cloudflare’s 15 September AI Crawler Defaults: The Check Gulf and Egypt Brands Should Run Today
• JavaScript Links Are Hiding MENA Sites From ChatGPT and Claude