A tester's answer is an observation from a particular interaction. It becomes a misleading business claim when a team removes the context and reports it as what every customer sees. Personalisation makes that risk worth examining, but it should not become a convenient explanation for every difference.
The goal is to separate the tester's history, an explicit conversation preference and a general market claim. That lets a UAE business decide whether to investigate a source, collect another observation or change its reporting. It also prevents a later answer shaped by the tester from being counted as independent evidence of broad visibility.
Research note, 5 November 2026: We collected two completed turns in one Google Search AI Mode conversation: an unbranded Dubai coworking question followed by an explicit JLT preference. The follow-up changed the prompt and returned a different output type. This is a live conversation-context demonstration, not a controlled test of saved-account personalisation. No account history was inspected or changed.
Distinguish three sources of apparent personalisation
An answer can reflect information written directly into the question, earlier turns in the conversation or documented account-based features. Keep these possibilities separate even when they produce a similarly tailored result.
Google's general Search guidance distinguishes personalised recommendations from other context, including location, language, device and current searches. It also notes that not all results are personalised and results can look similar with or without personalisation.
Its AI Mode documentation describes Personal Intelligence using saved activity under stated history and recommendation conditions. That capability description does not prove it influenced a particular answer. The same guidance lists UAE availability, supports follow-up questions and allows for web links instead of an AI answer.
| Possible influence | What can establish it in a test | What does not establish it |
|---|---|---|
| Explicit preference | The preserved prompt contains the preference | A guess about the tester's interests |
| Conversation context | Earlier turns are retained in the record | Calling a follow-up a fresh independent run |
| Saved-account context | Relevant documented conditions and observed evidence | A different answer alone |
| Other search context | Recorded location, language or device information | Describing signed out as context-free |
Use “not established” when you cannot identify the mechanism. An unexplained difference is still worth recording; it does not need an invented explanation to be useful.
Choose the claim before collecting answers
Define the decision your test is meant to support. A business might want to know whether a first unbranded question surfaces a relevant provider, whether a customer preference changes the response, or whether a source supports the statement attached to it.
Those questions have different units of observation. A first answer is not the same unit as a later conversation turn. A provider named in generated prose is not automatically equivalent to a web page listed after a follow-up. A source check is another unit again.
Write the intended claim in plain language before testing. “Describe what appeared in these two turns” is achievable with a small example. “Determine what most Dubai customers see” requires a much broader design and cannot be established by one tester's conversation.
This preliminary step also limits selective reporting. If the team starts with an unbranded discovery question, it should not replace the result with a more flattering follow-up after naming the business or narrowing the geography.
The live example began with an unbranded question
We asked Google Search AI Mode:
Which Dubai coworking spaces advertise day passes? Please link official sources and say what a visitor should confirm.
The observed product offered sign-in. The first completed reply presented five provider blocks: The Executive Centre, Cloud Spaces, The Bureau Dubai, Work From Park and WO-RK. It also included a map and visitor-check suggestions. We waited for the ready marker and complete response before continuing.
These names describe the collected output, not an independently assessed recommendation list. The answer contained prices, ratings and operational statements that we did not verify for this article. We do not adopt those details as booking guidance.
The question named no preferred provider or neighbourhood. That establishes the wording of this turn. It does not establish a perfectly neutral environment: we did not independently inspect every contextual signal or the session's location provenance.
Preserve that first result even if a follow-up seems more useful. It remains the observation corresponding to the original question and should not be overwritten by the conversation's final state.
An explicit JLT preference changed the next task
In the same conversation, we then submitted:
For this shortlist, prioritise spaces in JLT. Keep the day-pass requirement and use sources.
This is an intentional change in the task. It introduces a geographical preference into an existing conversation. It is not a discovery of a real user's private preference and does not demonstrate that the product remembered an account-level fact.
The settled output showed “10 sites” and a heading introducing top web results. It listed pages from sources including Spaces, Croissant, Letswork, directories and Servcorp. The interface again displayed a ready marker. We inspected the result type in both the accessibility output and a screenshot.
The important observation is not that a particular provider “won”. The first turn was a generated shortlist; the second was a web-results presentation after a changed instruction. Treating them as two interchangeable ranking lists would conceal both differences.
| Collected stage | Preserved condition | Appropriate description |
|---|---|---|
| First turn | Dubai-wide day-pass question | One generated answer with five provider blocks |
| Follow-up | Explicit JLT preference in the same thread | One completed web-results output containing ten sites |
| Mechanism | Prompt and conversational state changed | Saved-history personalisation was not tested |
| Wider inference | One sequence, no repeats | No market-wide visibility rate established |
We did not establish why the second turn returned web results rather than a generated shortlist. Do not turn that unexplained output choice into a claim about confidence, model behaviour or the effect of JLT wording.
Do not merge the two stages into a visibility percentage
A tempting report would combine all businesses and pages from the sequence into one visibility total. That would produce a number with an unclear denominator. It would mix an initial discovery answer, a preference-conditioned follow-up and two different output types.
Report the stages separately. If a business appears in generated prose at the first stage, describe that occurrence. If a later result links to a directory or booking platform, describe the page and its role. Do not quietly count every linked page as another direct provider recommendation.
This also prevents a common interpretation error: treating an appearance after several hints as proof that an unprompted customer would encounter the business. The later turn may be useful for diagnosis or a particular customer need. It simply answers a narrower question.
Where a numerical summary is needed, define the eligible observations and inclusion rule first. Retain output failures and alternative result types as their own categories. Excluding inconvenient outputs after seeing them can make a tidy metric less informative.
Verify the business facts independently of selection
We inspected Letswork's official day-pass page, which publicly advertises a day-pass product across workspace categories. That is a narrow primary-source check of the product's existence.
The follow-up included a Letswork result for RSO Workplace in JLT. The general day-pass page does not independently verify that result's particular rate, current capacity or venue-specific inclusions. We did not book a pass or inspect an accepted quotation.
A factual source check and a selection observation answer different questions. A provider can publish a relevant service without appearing in the first answer. An answer can surface a provider while making a detail that still needs verification. Neither fact should be used to erase the other.
Apply the same standard to a preferred business and its competitors. A tester's familiarity with a company is not evidence that its claims are accurate, and an unfamiliar result is not evidence of an error. Open the precise source needed for the decision.
Keep social context within its evidential role
The Letswork day-pass page directly linked its official Instagram account. We inspected an official announcement dated 2 November 2026 introducing Work From Park, a name that also appeared in the first generated answer.
The announcement supplies limited business context. It does not verify the AI answer's prices, opening arrangements, partnership statements or current desk availability. We inspected the author, caption and date; we did not independently assess the full video or visit the workspace.
Nor does finding this relevant post during our research prove the AI retrieved it. The answer's visible source material must support any retrieval claim. A brand's public announcement, a source cited by an assistant and a customer's actual booking are distinct evidence.
This distinction matters when a team already follows the brand. Familiarity can make a result feel expected or trustworthy. Keep that reaction separate from the recorded output and the source check so the report remains useful to someone who does not share the tester's history.
Design a saved-history test as a separate study
If the business needs to investigate saved-account history, design that investigation explicitly. The live example above cannot supply its missing evidence because it changed the current instruction rather than controlling an account-history condition.
Start by defining the particular feature and product surface under review. Record the relevant conditions that are actually observable, the same customer question and the intended comparison. Do not assume that signing out, starting a new thread or using a temporary conversation isolates every possible influence.
A stronger design would retain repeated observations, record the execution sequence and preserve differences in feature access, timing and other context. It would also state what remains uncontrolled. Two answers that differ after a setting change can be described accurately without claiming the change alone caused the difference.
Do not inspect private history, connect personal apps or alter a person's settings merely to complete a form. Use an authorised research setup appropriate to the question. The reporting standard should make missing evidence visible rather than encourage unnecessary access.
Any such study remains unperformed here. Its possible value does not justify presenting a small conversation demonstration as an account-level experiment.
Prevent context from leaking between test categories
Keep first-turn discovery tests separate from exploratory conversations. Once a tester has named the target business, supplied its website or expressed a preference, label the subsequent output accordingly.
A practical sequence can preserve the first unbranded answer, then allow diagnostic follow-ups in a clearly marked section. The diagnostic work may reveal useful questions about service scope or source support. It should not flow back into the original first-answer count.
If another tester repeats the question, record their conditions rather than treating their result as confirmation by default. Similar answers do not prove identical processing, and different answers do not prove personalisation. The evidence may support only a description of the observed variation.
For teams maintaining a programme, agree on category definitions before reviewing client results. This reduces the chance that one analyst counts a linked directory as a recommendation while another counts only provider names in generated prose.
Write conclusions at the level the evidence supports
Lead a report with its narrowest defensible finding. In this example, the conclusion is that a two-turn conversation changed from a generated Dubai-wide shortlist to a web-results output after an explicit JLT instruction. Saved-history personalisation and population preferences were not measured.
A stronger-sounding claim such as “Google prefers these businesses for JLT users” would add an untested user group, a stable preference and an inferred mechanism. The collected sequence supports none of those additions.
Similarly, avoid claiming that a business's content failed because it disappeared after a changed question. Check whether the later task still fits the service and whether the output type is comparable. The appropriate action may be to classify the observation differently rather than commission a new page.
If a source claim is wrong, assign a source-review task. If the result is simply conditioned by an explicit customer need, assess that need on its own terms. If the mechanism remains unclear, preserve the uncertainty and decide whether further research would change a business decision.
Maintain a record that another analyst can interpret
Retain the exact prompt sequence, product surface, visible account evidence, output completion and result type. Store source checks alongside the claims they support. Keep the initial answer intact even after the conversation develops.
Review the protocol when personalisation or source interfaces change. An earlier observation should remain dated evidence of what was collected, not be silently rewritten to match a later feature description.
For the broader priorities, use our Google AI Search guide for UAE businesses. Personalisation research is useful when it narrows a real uncertainty; it becomes misleading when one tester's history or conversation is presented as the market.
Checked: 5 November 2026. Next review: 5 December 2026, or sooner after relevant product changes. One two-turn sequence was collected; no saved-history experiment, representative user sample, API run, booking or conversion measurement was performed.
Disclosure: AI Visibility is published by Lunasol, which offers related GEO and AEO services.



