A useful voice-versus-typing test begins with the question the system actually received. If the spoken version loses a service detail, changes a place name or adds a conversational preference, a different answer cannot automatically be attributed to voice input.
For a UAE business, the decision is whether a customer can express a purchase need clearly and receive relevant, supported information. Separate speech capture, recognised wording, the search experience and the final answer. Each can create a different problem and call for a different response.
The aim is a comparison another reviewer can interpret. It should show what the tester intended, what was spoken, what the interface recognised where visible, and what the completed response offered. Avoid treating a conversationally written prompt as actual voice data.
Research note, 5 November 2026: This is a researched comparison-design guide. No spoken query, recorded audio, recognised transcript or matched voice/typed answer pair was collected. The examples are proposed and unrun. We checked official platform guidance and a public Dubai repair catalogue; no microphone permissions were changed or customer booking attempted.
Decide which question the experiment should answer
There are at least three useful comparisons, and they should not share one result label.
| Comparison | What stays similar | What the test can investigate |
|---|---|---|
| Identical wording typed and spoken | Intended words and specified product surface | Input capture and outputs under recorded conditions |
| Natural typed and natural spoken questions | Customer need | How different real-world formulations perform |
| Typed search and a Live conversation | Broad task | A wider difference in surface, interaction and output |
The second comparison can be valuable even though wording changes. It simply cannot isolate an input-modality effect. The third changes more than the way words enter the system and should be reported as a comparison of experiences.
Choose the intended conclusion before collecting answers. If the business wants to understand a customer's natural question, do not force the customer into an artificial script and then claim to have measured natural behaviour. If the aim is a tighter input comparison, preserve the same script and record deviations.
Name the voice route rather than the microphone icon
Google's AI Mode guidance describes typed questions and microphone input. Its separate Voice Search instructions describe a general voice-search route and include language and regional caveats. A microphone symbol alone does not identify the resulting search experience.
Search Live guidance describes back-and-forth audio conversations in the Google app on Android and iOS, with web links on screen and microphone permission required. It is distinct from entering one microphone question into a desktop search interface.
Record the actual product, application and route used. Do not call a desktop microphone test a native Search Live test. Likewise, a voice command leading to ordinary search results is not automatically an AI-generated answer.
The AI Mode documentation lists UAE availability, and Search Live describes its scope in relation to supported AI Mode regions and languages. We did not verify native-app availability on a particular UAE phone or account. Documented scope and observed access remain separate.
Preserve the stages between intention and submission
A test record should retain several versions of the question where they are available. They answer different questions about what happened.
First preserve the intended purchase request. Then retain the words actually spoken, rather than assuming the tester read a script perfectly. Record the recognised transcript if the product exposes one, followed by the query or conversation turn that was submitted.
If the interface does not expose a transcript, mark that field unavailable. Do not reconstruct it from the answer or silently replace it with the intended wording. An assistant answering a related question does not prove which words it recognised.
Keep edits visible. If the tester corrects a place name before submission, the record concerns corrected voice input. If they submit the first recognition unchanged, it concerns the original capture. Both can be useful, but they should not be merged.
This layered record helps route a problem. A lost qualifier might be visible before the answer is generated. An accurate transcript followed by an unsupported business claim points to a different review task.
Use a purchase detail that can be checked
A local service example makes the design concrete. The Shoe Doctor DXB repair catalogue separately lists heel-tip replacement and heel replacement for men's formal shoes. We checked that public distinction; we did not obtain an assessment or confirm an item's suitability.
That gives a proposed test a meaningful qualifier to preserve. “Heel tips” and a broader “heel replacement” request should not be treated as interchangeable merely because both concern shoes.
Here is a proposed identical-wording prompt, not run:
Which Dubai shoe-repair businesses advertise heel-tip replacement? Link the service page and tell me what I should confirm before ordering.
The comparison should preserve the service qualifier. It should not invent an example in which a recogniser actually dropped “tip”, because no speech was collected here. The risk to check is clear without fabricating a failure.
The business was selected independently during research, not returned in an AI answer for this article. Its catalogue supplies a real service distinction for the protocol, not a discovery finding.
Keep natural wording as its own comparison
A customer might type a short phrase but speak a longer request. That is a useful difference to investigate when the goal is practical customer discovery, provided the report acknowledges the changed wording.
For example, a proposed natural-language question could be:
The little tips on my heels are worn down. Who in Dubai says they replace those, and what would I need to send them for an assessment?
This is an authored example, not a customer quotation or spoken transcript. It asks for a similar need using a different formulation. A typed short phrase and this longer request would change both wording and input style if one were spoken.
Keep the resulting study separate from the identical-wording pair. Do not use a better answer to the detailed request as evidence that voice itself performs better. The additional description may be relevant, but the design must not assign its effect without evidence.
If actual customers contribute wording later, preserve the source and agreed research use without collecting unnecessary personal details. Label whether the question was naturally produced, read from a script or edited for the study.
Treat language and place names as separate conditions
The site's existing Arabic-question study concerns typed language comparisons and explains its wording and method limits. It does not supply voice-recognition evidence. Its findings should not be repackaged as proof about spoken Arabic or accents.
For a voice comparison, record the spoken language and the relevant application settings where known. If one arm is English and the other Arabic, the test changes language as well as input. Design that as a separate question rather than burying it inside a voice-versus-typing label.
Preserve location words exactly where visible. A Dubai neighbourhood or branch name can materially narrow a service request. If recognition changes it, keep the original and corrected forms rather than making an undocumented repair.
One speaker's result also cannot establish performance across a language community. A broader claim requires an appropriate range of observations. A small pilot should describe its speakers and conditions at a proportionate, non-identifying level and keep generalisations narrow.
Preserve material and service qualifiers
A purchase question often contains more than a category and a city. Material, item type, deadline and exclusions can determine whether a provider is relevant. Record those details in both arms rather than comparing only the final business names.
An official Shoe Doctor DXB post from October 2026 discusses its material-first approach. The repair catalogue links the official account, and the post's author and caption were inspected. It is attributed business context, not independent material-care advice or voice-test evidence.
The post's accessible image date and visible permalink date differed by a day, so this guide uses the month rather than assigning an unsupported exact date. That discrepancy does not change its limited role as service context.
The lesson for the proposed test is to retain a material word if the question contains one. Do not move a claim about cleaning, restoration or assessment into a different service merely because the answer sounds plausible. Check the precise published offer.
Run the pair with a clear completion rule
For an actual identical-wording test, specify the product surface, device, account conditions and first-turn state. Record the intended sequence. Do not use the typed answer as context for the voice arm unless conversation continuity is part of the question being tested.
Define completion for each arm. A successful speech capture followed by an interrupted answer is different from a completed answer. An unavailable microphone or denied permission prevents that arm from producing comparable voice data; it does not establish that a business was absent.
Retain the first attempt under the declared rule. If the tester repeats a phrase, edits recognition or restarts after a failure, record the intervention. Repeated attempts can help diagnose a problem but should not disappear from the record after a successful result.
A single pair can describe what occurred in those conditions. Repeated observations are needed before discussing persistence, and repetition alone does not isolate every influence. Keep the design and conclusion aligned rather than promising a universal voice effect.
Assess the heard response and the screen separately
A spoken answer and its accompanying screen can provide different evidence to a customer. Search Live documentation describes audio responses with web links on screen; that makes both components relevant when that is the experience being tested.
Record which businesses or claims were heard and which appeared visually. Do not count a source visible only on screen as something the customer heard. Equally, a shorter audible answer does not prove the product supplied fewer sources overall.
If the research decision concerns hands-free use, define what information the listener needs without touching the screen. If it concerns a user who can inspect links, include that path explicitly. These are different usability questions.
We did not collect either kind of output here. The distinction is part of the proposed protocol, not a finding that one format is more useful or better at converting customers.
Use an evidence table with honest empty cells
Before execution, a planned comparison should look incomplete because the observations do not yet exist. Preserve that state rather than filling it with plausible answers.
| Field | Typed arm before execution | Voice arm before execution |
|---|---|---|
| Intended question | Proposed script preserved | Same proposed script, if testing identical wording |
| Actual input | Not submitted | Not spoken |
| Exposed transcript | Not applicable to typed entry | Not collected |
| Submitted wording | Not collected | Not collected |
| Completed response | None collected | None collected |
| Source-supported claims | Not assessed | Not assessed |
After execution, add the actual observations and the limitations. Keep unavailable fields visible. A reviewer should be able to distinguish an unrun arm, a failed capture, an interrupted answer and a completed result.
Keep transcription quality separate from answer quality. A correctly recognised query can receive an inaccurate answer; a repaired transcript can lead to a useful diagnostic response without proving the original voice path succeeded.
Route the finding to the right action
If a transcript visibly changes the service request, investigate capture and wording before assigning a content task. If the transcript preserves the need but the answer makes an unsupported service claim, inspect the relevant source. If a result is relevant but its link cannot be reached, review source access.
For the shoe-repair example, check whether a future answer actually connects a provider to heel-tip replacement. A generic repair page may be weaker support than the precise catalogue item. Do not infer suitability, price or turnaround for the customer's shoes from the service name alone.
If public information is unclear, confirm the facts with the business before proposing edits. A clear service label and assessment process may be more useful than a new article written around a supposed voice-search ranking factor.
The report should end with the narrow decision supported by the collected evidence. Until a real pair is run, this guide offers a method and verified service context, not a conclusion about which businesses voice search recommends.
Checked: 5 November 2026. Next review: 5 December 2026, or sooner after relevant input or transcript changes. No actual voice query, recognised transcript, matched answer pair, microphone permission change, booking or conversion measurement was collected.
Disclosure: AI Visibility is published by Lunasol, which offers related GEO and AEO services.



