Arab AI Visibility Index: method, registered before data

The Arab AI Visibility Index tracks what AI assistants recommend in Arabic. Once a month, starting 17 November 2026, we ask ChatGPT and Google AI Mode the same 150 commercial questions about Saudi Arabia, the UAE and Egypt, and record which brands they name, whether those brands are local, which sites they cite, and how many of those sites are in Arabic.

This page is the registration. We wrote the method down and published its SHA-256 hash before the first round of answers. That way nobody, us included, can bend the rules after seeing the numbers. Results will appear on this same page every month from December 2026 to April 2027.

If you market a hospital, a lab, a developer, a training institute or a bank in the Gulf or Egypt, the index answers a practical question. When a customer asks an AI assistant in Arabic where to go, does the answer name businesses like yours, and whose pages does it trust to back that up?

What the index measures

Every answer is checked for the same four things, and each engine gets its own figures. We never average ChatGPT and Google AI Mode into one score.

Each round also reports secondary figures: how many brands an answer names, how often it cites no source at all, and what kind of site gets cited, from brand sites and government pages to media, directories, forums, social and video. It also shows how often sponsored units appear and how closely the two repeats of the same question agree.

The two engines

Both engines are used the way an ordinary visitor meets them, without an account.

• ChatGPT at chatgpt.com, logged out, on the free tier as served. Each question goes into a new chat through the normal message box, with nothing that forces a web search, because whether ChatGPT decides to search is part of what we measure.
• Google AI Mode, signed out, with one fresh search per question and Google’s interface language set to Arabic.

Other engines were tested in earlier VOCTOS work and left out because they need a sign-in or show no readable sources. Logged-out Perplexity hides its source list behind a sign-in, Gemini requires an account, and Brave’s AI summary carries no links to its sources. Adding an engine later would mean a separate series with its own registration.

The questions: 150, fixed for six months

The panel is 150 questions in Modern Standard Arabic, drawn with a fixed random seed from the 5,958-question Arabic bank VOCTOS built for its citation research. Nobody chose them by hand. Each drawn question is asked about all three countries in the same words, so when Saudi and Egyptian answers differ, the wording is not the reason.

The panel covers five kinds of question, 30 of each: the best provider in a named city, what to look for when choosing one, what something costs in the local currency, which of two options suits the country, and when you really need a service. Two examples from the Saudi set are «ما هو أفضل مستشفى خاص في الخبر؟» and «البنك الرقمي أم البنك التقليدي؟ ما الفرق بينهما وأيهما أنسب في السعودية؟».

Health gets twice as many questions so its figures can be reported on their own. It is also where VOCTOS has done most of its earlier measurement. In the overall figures, each of the four sectors still counts equally. Marketing and advertising is left out on purpose, because VOCTOS competes in it.

We capture answers in the browser, logged out, the same way we collected 200 answers for our Egypt study of ChatGPT’s local recommendations, and the capture and stop rules carry over from that study. The panel itself does not change during the series. A different panel would start a new series, and at least one round would run both so the two can be compared.

Countries, and where the questions are asked from

The three markets are Saudi Arabia, the United Arab Emirates and Egypt, with 50 questions each. Every capture runs on one machine in Egypt on its normal connection, with no VPN. The question names the market and the connection stays where it is.

That means a Saudi or Emirati answer in this index is what someone in Egypt sees when asking about Riyadh or Dubai, which can differ from what a person in Riyadh sees. Google prints its own guess at the user’s location under the results. Any capture where that line names a country other than Egypt is asked again once.

Repeats and the six rounds

Each round asks every question twice on each engine, once on the first capture day and again on the next working day in a new random order. That makes 600 answers a round. The second ask shows how much an answer moves when nothing has changed except the day, and that movement is the yardstick for calling anything a change between months.

Six rounds are scheduled. A reserve day follows each pair of capture days, and each round’s results are published within 10 working days of its last capture.

The gaps are uneven on purpose. Round 4 is taken before Ramadan, round 5 after Eid al-Fitr, and round 6 early enough to publish before LEAP in Riyadh, so the time between rounds runs from 19 to 42 days. Every release states its gap. A round moves only for an outage, a public holiday we did not plan for, or the operator being unavailable, and the move is logged before the round starts. No round is ever run again because its results look odd.

What counts as a change from one month to the next

Engines give different answers on different days even when nothing has changed, so most month-to-month differences are noise. We call a difference a change only when all five of these conditions hold.

1. It passes a permutation test that reshuffles each question’s four answers (two this month, two last month) between the months 10,000 times, which builds normal day-to-day variation into the comparison.
2. It survives a Holm correction across the six headline tests, three rates on each of the two engines, at the 5% level.
3. The difference is at least 5 percentage points.
4. The difference is larger than the gap between the two repeats of the same figure inside either month.
5. Both months have at least 90% valid answers in each repeat for that engine. A round below that is published as incomplete and is not tested.

Anything else is reported as no change detected, with the difference and its 95% interval shown. Round 1 is the baseline and makes no change claims.

A change that coincides with a model or product update we logged before the round is labelled as an engine update. We only call something a trend when it shows up in two monthly comparisons in a row, or against the round three months earlier. A domain moving from fourth to second place is shown in the table and never treated as a change. Cuts by sector, country and question type are tested the same way, labelled exploratory, and kept out of the headlines.

How VOCTOS clients are flagged

VOCTOS sells the SEO and AI search services whose demand this index measures, and some of our clients work in the sectors on the panel. Three are known to fall inside it, all in Saudi Arabia: Delta Medical Labs and Mouwasat in health, and HyperPay in finance.

Client brands and domains stay in the data and are never removed. Every pick and every cited page that belongs to a client is marked in the released files, and each release says how often client brands were named and cited, including when the count is zero. The coders never see the client list.

Limitations we are declaring now

• Every question is asked from one machine in Egypt, so the Saudi and Emirati results describe what an asker in Egypt sees.
• Only two engines are measured, both logged out. Signed-in users, paid plans, the mobile apps and other engines may get different answers.
• The panel holds 150 commercial questions in four sectors. It is not the whole Arab market, and it includes no news, civic or purely informational questions.
• All questions are in Modern Standard Arabic, so the index says nothing about how dialect or English questions are answered.
• Each question gets two draws a month, and both engines vary. In earlier VOCTOS runs the first pick held in only about half of repeated questions, so brand and domain tables are noisier than the headline rates.
• Engines change without notice. We label the changes we can see, and some will stay invisible to us.
• Google AI Mode links its sources through redirect tokens that expire within minutes. We publish the share we could not resolve, so the gap is visible.
• Brand origin and page language are coded by written rules with known edge cases, such as licensed brands and bilingual pages. The rules and the coding files are public.
• The two coders are separate model instances working independently. Their agreement is published every round. We would welcome an outside reviewer to recode a sample, and none is arranged yet.

Registration files and their hashes

Anyone can check that the method described here is the one we registered. Download the two files, run sha256sum on them, and compare the result with the hashes below. If anything has to change after round 1 starts, it goes into a dated addendum at the end of the registration, and the next release reports it.

• Download the registered method (PREREG.md, version 1.0)
• Download the question panel (panel.csv, 150 questions)

Each round’s data, covering the answers, the citations, both coders’ files and the analysis scripts, goes on Zenodo under CC BY 4.0 with its own DOI, linked from this page.

Where the results will appear

Results for each round will be published on this page, newest first, with earlier rounds in a table further down and every correction dated. The address stays the same for the whole series, so a link or a citation to this page keeps working from round 1 to round 6.

To cite the method: VOCTOS (2026). Arab AI Visibility Index: method, registered before data. Version 1.0, available at https://www.voctos.com/research/arab-ai-visibility-index/

Questions marketers ask about the index

Why ask ChatGPT while logged out?

We measure what anyone gets by default, without an account, a chat history or a paid plan. Signed-in and paying users may see different answers, which is listed among the limitations above.

Why are Saudi and Emirati questions asked from Egypt?

Every capture runs on one machine in Egypt on its normal connection, and we never use a VPN or switch networks to get answers. The Saudi and Emirati figures are therefore what someone in Egypt sees about Riyadh or Dubai. A capture from inside Saudi Arabia would have to be registered as a separate series before it runs.

What happens in a month when nothing changes?

We publish it anyway. A month where no headline figure passes the change rule gets the same release as any other, and the sentence saying so is shown as prominently as a change would be.

Can a brand ask to be added to the panel?

No. The 150 questions were drawn by seed before any data was collected, and they stay fixed for the whole series. A brand enters the index only when ChatGPT or Google AI Mode names it in an answer.

Can I reuse the index data?

Yes. Each round is released under CC BY 4.0 with its own DOI, so you can quote it, chart it or reanalyse it, as long as you credit VOCTOS and link to the source.