Google AI Mode Query Fan-Out (2026): Why One Dubai Search Becomes Ten Hidden Questions
89% of UAE searchers now start with an AI assistant, and AI Mode never searches what they typed. It fires around 10.7 hidden sub-queries. How fan-out picks pages, and how to own one.
- PUBLISHED
- 07 NOV 2026
- READ TIME
- 13 MIN
- AUTHOR
- AQUARIUS · DUBAI
- UNIT
- REV 2026.09
Short answer: Google AI Mode does not search what your customer typed. It rewrites that one question into a batch of hidden synthetic queries — around 10.7 of them on average in Gemini 3-era testing — runs them in parallel, scores individual passages from the results, and writes one answer from the winners. So the page that gets cited is rarely the page that ranks for the typed phrase. It is the page that holds the best passage for a sub-question nobody ever saw.
That single mechanic explains almost every confusing analytics screenshot a Dubai marketing manager has stared at this year: rankings flat, impressions up, clicks down. You are being read, decomposed and quoted without being visited.
Why this matters more in the UAE than almost anywhere else
Three numbers set the stakes, and they are unusually extreme here.
UAE searchers have already moved. YouGov’s Web Search and AI Report 2026 — online searchers across 19 markets, fielded 21 April to 29 May 2026 — found that 89% of UAE respondents now start queries with an AI assistant. That is the joint highest of all 19 markets, level with India and Indonesia, ahead of Saudi Arabia at 86% and Singapore at 82%, and nearly double Great Britain at 54% and the United States at 48%. Whatever timeline a global agency is selling you for “AI search readiness”, the UAE is a year or two ahead of it.
Adoption is not a niche. Microsoft’s 2026 AI Diffusion Report put the UAE top globally, with 70.1% of the working-age population (15 to 64) using AI for at least 90 minutes a month, against a global average of 17.8%. The Dubai Center for AI reports 67% of UAE consumers trust AI systems, versus 42% globally. High trust plus high usage is exactly the condition under which an AI answer substitutes for a click.
And the clicks are genuinely going. AI Overviews now appear on roughly 48% of Google queries. An Ahrefs study published in February 2026 measured a 58% reduction in click-through rate for top-ranking pages when an AI Overview is present — up from the 34.5% the same methodology found in April 2025. Google’s own numbers show the shape of it: in Q3 2026 Search revenue growth slowed from 17% to 12% and paid clicks decelerated from 13% to 8%, even as total query volume grew at its fastest rate in five years.
None of that is a reason to abandon Google. Google still holds about 93.5% of UAE search in September 2026 (Bing 3.96%, Yandex 1.8%), and roughly 98% on mobile and tablet. The traffic is still there. The path to it changed.
What query fan-out actually does, step by step
Google has described the architecture in a run of patents, the most explicit of them published in March 2026. Stripped of the legal language, the pipeline runs like this:
- Interpretation. The typed query is combined with context — device, location, prior turns in the session, inferred intent.
- Synthetic query generation. A generative model emits a set of new queries covering different facets: the explicit need, implicit needs, comparisons, and historically adjacent questions. These are not paraphrases. In testing, the sub-queries fired frequently contain none of the words the user typed.
- Parallel retrieval. Each synthetic query runs its own retrieval against Google’s index, producing its own candidate document set.
- Passage selection. This is the part that breaks old habits. The March 2026 patent describes building a separate set of relevant, diverse passages for the main query and for each related query, then selecting across those sets. Each chunk is scored on its own merits rather than inheriting the page’s authority.
- Synthesis. The surviving passages are merged into one answer, with links attached to whichever sources contributed.
The consequence is measurable. Only about 38% of AI Overview-cited pages still rank in the organic top 10. Roughly 31% sit at positions 11 to 100, and another 31% rank beyond position 100 for the query that triggered the answer. Position-one thinking does not describe this system.
How many hidden queries? The number worth planning around
Seer Interactive’s November 2025 measurement put Gemini 3 at an average of 10.7 sub-queries per prompt. Independent observation is lower and more variable — Gemini firing 3 to 8 searches, ChatGPT 1 to 6, depending on query type. Treat 10.7 as a practitioner benchmark, not a Google-published constant. The planning implication holds either way: one commercial question in Dubai is being answered from five to eleven different retrievals, and you can win or lose each one separately.
Work an example. A Jumeirah clinic owner asks AI Mode whether it is worth building their own booking system or using an off-the-shelf app. The fan-out plausibly includes build-versus-buy cost comparison, DHA patient-data rules, typical no-show rates, integration with insurance claims, Arabic interface support, and ongoing maintenance cost. Six sub-questions. A single “clinic booking system Dubai” service page answers at most two of them. A clinic that has also published a maintenance-cost breakdown, a DHA compliance note and an Arabic UX page is eligible for four.
Making a page fan-out-eligible
The shift is from ranking a page to seeding retrievable passages. Concretely:
| Habit | Classic SEO (still true) | Fan-out SEO (what changes) |
|---|---|---|
| Unit of competition | The page, for one keyword | The passage, for an unseen sub-query |
| Keyword research | Volume and difficulty on the head term | Map the decision: every implicit, comparative and objection question around it |
| Page structure | One H1, keyword in the first 100 words | Self-contained H2/H3 blocks that answer in the first two sentences and survive being quoted alone |
| Specificity | Helpful prose | Named figures — AED ranges, dates, percentages, regulator names — because unquantified claims lose passage scoring |
| Language | English, Arabic “later” | Both, because Modern Standard Arabic fan-out is a separate retrieval pool with a fraction of the competition |
| Rendering | Client-side JS tolerated | Server-render the answer text, not just the shell |
The Arabic line deserves its own sentence, because it is the cheapest win on the list. AI Mode reached Modern Standard Arabic in October 2025 as part of a 38-language expansion covering more than 200 countries and territories, after launching in English across MENA in August 2025. Arabic fan-out therefore runs today, against a far thinner corpus. If your site is English-only you are invisible to one entire retrieval pool in your own market. Our notes on Arabic and RTL design for UAE sites cover the build side; the content side is simply that translated-and-reviewed beats machine-dumped, and nothing beats absent.
How to measure something Google hides
Until mid-2026 there was no way to isolate AI performance: AI Overviews and AI Mode impressions were folded into the ordinary Search Console Performance report. On 3 June 2026 Google announced a dedicated Generative AI performance report in Search Console, rolled out first to a subset of UK site owners under pressure from the UK Competition and Markets Authority, then wider.
Know its limits before you build a dashboard on it:
- It covers three surfaces — AI Overviews in Search, AI Mode in Search, and generative AI features in Discover.
- It reports impressions only, broken down by page, country, device and date. There are no clicks, no CTR and no query data. You can see visibility, not its traffic value.
- Every link inside an AI Overview shares a single position value, so per-link rank optimisation is not a thing.
- It is a breakout of data already counted, with no historical backfill — reporting begins from around 18 May 2026.
- A separate opt-out toggle can exclude content from AI Overviews, AI Mode and Discover AI without touching normal Search ranking. Almost no UAE business should use it, but know it exists.
Practically, your AI-search scorecard is three things: generative impressions from Search Console, referral sessions tagged by AI host in analytics, and a manual citation log — run your twenty money questions through AI Mode monthly in both English and Arabic and record who gets quoted. We walk through that loop in how to get cited by AI search, and the surrounding market shift in what AI Overviews mean for UAE businesses.
What it costs in Dubai, and what it returns
This is less a new budget line than a reallocation of an existing one. Typical Dubai ranges, all subject to 5% VAT:
| Work | Typical Dubai cost (AED) | What you get |
|---|---|---|
| One-off fan-out and passage audit | 6,000 – 15,000 | Sub-query map for your top commercial questions, passage rewrite list, schema and rendering gaps |
| Monthly SEO and AEO retainer (SME) | 3,000 – 8,000 / month | The standard Dubai band; bilingual content and heavier technical work sits at the top end |
| Bilingual content programme | 8,000 – 20,000 / month | English plus reviewed Modern Standard Arabic — the second retrieval pool |
| Citation monitoring and reporting | 1,500 – 4,000 / month | Monthly AI Mode citation log, generative impression tracking, competitor share of answer |
The return argument rests on one contested but consistent finding: AI-sourced visitors convert better, because the model has already done the qualifying. Semrush’s 2026 benchmark puts AI search visitors at roughly 4.4x the conversion rate of standard organic. Seer Interactive measured ChatGPT referral traffic converting at 15.9% against 1.76% for Google organic. Be honest about the dispute: Amsive’s statistical analysis found no significant overall difference, and the multiple swings by industry and measurement method. The defensible version is that AI referrals are low-volume and high-intent — precisely the traffic a Dubai services business wants.
The cost of inaction is easier to size. If AI Overviews already sit on about half of queries and cut top-position click-through by 58%, then a Dubai business holding rank 1 on its money term is losing roughly a quarter of the clicks that position used to produce — without a single ranking moving. That is the number to put in front of a finance director, not a vague appeal to “AI readiness”.
How Aquarius approaches it
We start by mapping the fan-out, not the keyword. We take your five highest-value commercial questions, run them through AI Mode in English and Arabic, and record the sub-questions the answer actually resolves plus who is quoted for each. That produces a gap list of passages you do not own. Then we fix the two mechanical blockers almost every UAE site has — answer text that only exists after JavaScript runs, and missing Arabic — before writing a word of new content. The rendering and speed work overlaps directly with Core Web Vitals, so it pays twice. If you are starting from scratch on fundamentals, begin with SEO for Dubai businesses instead.
No one can promise you a citation, and anyone who does is guessing — AI Mode output is not deterministic and changes between runs. What is promisable is eligibility: server-rendered answers, quotable passages carrying real numbers, both languages live, schema that removes ambiguity, and a monthly log that shows whether your share of answers is moving. That is the whole job.
Frequently asked questions
Is query fan-out the same as AI Overviews?
No. AI Overviews is the summary block on a normal results page. AI Mode is the separate conversational search surface. Query fan-out is the retrieval mechanism both lean on — decomposing one question into multiple synthetic queries, retrieving in parallel, then synthesising. Fan-out is the plumbing; the two surfaces are taps.
Can I see the sub-queries Google fired?
Not officially. Search Console’s Generative AI report carries no query-level data at all. In practice you infer them: read what the answer actually resolves, note the follow-up suggestions AI Mode offers, and reverse-engineer the facets from the sources it cited. That inference is good enough to plan content against.
Do I need llms.txt or a special file for AI Mode?
No. Google has said it does not use llms.txt, and AI Mode retrieves from the ordinary Google index. The requirements are the ordinary ones, enforced harder: crawlable, server-rendered, fast, well-structured, specific. Blocking Google-Extended affects Gemini grounding rather than Search indexing — check what you have blocked before assuming you are eligible.
Does ranking still matter?
Yes, just less exclusively. Around 38% of cited pages rank in the organic top 10 — a strong correlation, not a requirement, given that 31% rank beyond position 100. Rank remains the cheapest proxy for being retrievable. It stopped being the finish line.
How long before this shows results?
Expect 8 to 12 weeks before generative impressions move meaningfully, assuming the rendering and Arabic blockers are fixed in month one. Citation share is noisier than rankings and needs a monthly log rather than a daily dashboard — judge it on the trend across a quarter.
If you want to know which hidden questions your Dubai customers are really asking — and which of them a competitor is currently answering for you — tell us your five money questions and we will run the fan-out map. Build and retainer ranges sit on our pricing page.