A Dubai business has tested the same four questions for three months. Customers start asking whether they can get a quote by sending photos, and that task has no place in the panel. If the team replaces one question immediately, the next visibility score can move simply because the questions changed. If it keeps the panel closed to new topics, it can miss a real customer need.
Freeze the prompts that support the trend and give new questions a separate route into the test. Keep exploratory results visible, review candidates against a stated rule, and run old and new versions together when the core needs to change. A frozen panel is a record of which questions you used for a comparison. It can coexist with a queue of questions you are still learning about.
Freeze the conditions that make the core comparable
Start by writing down the decision the core panel serves. It might monitor whether the right business appears for a few priority services, or whether a customer can find the correct branch and booking route. Choose questions that represent those decisions. Record the exact text, its language and intent, and an ID that stays with it across reporting periods.
Save the rest of the protocol with the wording. Name the surface, market, session rule, context or order if applicable, scoring definitions, and expected reporting window. A question asked after a conversation about the brand can behave differently from the same text submitted in a fresh session. Changing the interface locale or the way sources are captured can also affect comparability. The core should specify which conditions are fixed and which the analyst will simply record.
Survey researchers face a related continuity problem. Pew Research Center's guidance stresses keeping wording and context comparable when measuring change over time. An AI visibility panel can borrow that discipline without adopting survey sampling assumptions. Treat revised wording as a new version, even when the editorial intention is to make the sentence sound more natural. Pew Research Center guidance on writing survey questions
The existing AI Visibility test-sheet guide covers collecting observations about what assistants say. Keep the capture record for that work. A panel register adds the decision about which question versions belong in the recurring comparison and when a new question can enter it. AI Visibility test-sheet guide
Give new customer questions a visible intake route
Create an exploratory queue that can accept questions from contact forms, phone conversations, messages, sales notes and site search, when those records are available for the purpose. Keep a short description of the source and date. Remove customer-specific details from test wording. The aim is to reproduce a customer task without copying a private message into an external product.
Capture the underlying intent separately from the proposed prompt. “Can I send a picture?” might mean the customer wants a remote quote, a preliminary assessment, or confirmation that a job is within scope. Clarify what decision the customer is trying to make before drafting the test. Two similar strings can belong to different tasks, while two phrasings can test the same task.
The official Shoe Doctor DXB Instagram example below illustrates a visual service example a prospective customer may see before asking about assessment. It documents what the business published. To justify adding a remote assessment question to a recurring panel, look for separate evidence that the task matters to customers or operations. A dated public post can suggest a candidate without establishing how often customers ask it.
Give exploratory work its own capacity. A team can reserve a small number of slots per period for new questions and log them alongside the core with a distinct flag. If a question exposes an actionable problem, the business can investigate it immediately. The result can remain exploratory while the fix proceeds. Putting a useful finding into an action queue does not require putting the prompt into the trend.
Define when a candidate can join the core
Set the promotion rule before reviewing the candidate's AI results. Ask whether the task is relevant to the business decision, whether it fills a gap in the existing panel, and whether a reviewer can verify and score it consistently. A new high-impact booking route may qualify even with little request volume. A burst of seasonal questions may deserve a short-lived supplementary panel. Record which reason applies.
Check for overlap with the core. If the new question is a wording variant of a task you already track, it can remain a diagnostic. If it changes a condition such as branch, service constraint or contact method, decide whether that distinction warrants its own recurring observation. A small panel has a finite budget. Adding a candidate can mean retiring an obsolete prompt or reducing a different test condition.
Write the acceptance decision and effective date in the register. Preserve the candidate's exploratory observations as such. Promoting a question because the business performed well or poorly on its first run would make the panel sensitive to the outcomes it is supposed to measure. A reviewer should be able to see why the team added the task and which definition will apply in the next reporting cycle.
Use a bridge cycle when the core changes
A bridge cycle runs the outgoing and incoming panel definitions in the same reporting window. Submit the shared questions once per planned condition and use those observations in both views. Run both the retiring prompt and its replacement under matched conditions. Save the results and calculate both definitions from the same window. This shows how the panel change affects the reported score at that point in time.
The following example is constructed. A fictitious Dubai leather repair business monitors whether it is named in answers. The outgoing core has four prompts, A through D. The business decides that a general counter question is less useful after a change to the customer intake process. It approves a photo-assessment prompt, E, after exploratory review. The bridge is run on one surface with a stable market, session and scoring rule. A yes means the business was named in a scorable answer.
| Prompt | Customer task | Earlier cycle | Bridge cycle | Next cycle |
|---|---|---|---|---|
| A | Find a repair provider | Yes | Yes | Yes |
| B | Check strap replacement | No | No | No |
| C | Find the Marina drop-off location | Yes | Yes | Yes |
| D | Ask about visiting a general counter | No | No | Outside new core |
| E | Ask how to get a photo assessment | Outside old core | Yes | Yes |
For a reproducible planning example, the exact fictional prompts could be as follows. A asks “Which Dubai businesses repair leather handbags?” B asks “Can Harbor Leather Repair replace a handbag strap?” C asks “Where is the Harbor Leather Repair drop-off point in Dubai Marina?” D asks “Can I visit the general counter at Harbor Leather Repair to ask about a handbag repair?” E asks “How do I send handbag photos to Harbor Leather Repair for an assessment?” These wordings were prepared for the example and were not submitted to an AI product. The table illustrates the scoring and version transition.
In the earlier cycle, the old core has two mentions in four answers. In the bridge cycle, the old core still has two of four. The new core, A through C plus E, has three of four in that same bridge window. The observed difference between the definitions is one mention or 25 percentage points. It reflects which question was counted in this fixture. Comparing two of four in the earlier cycle with three of four under the new definition would combine a different panel with a different period.
In the next cycle, the new core has three of four again. The shared three prompts have two mentions in all three cycles. Report that continuity separately from the full-core figures. The bridge required five distinct runs under the single condition. Adding the old and new panel sizes would give eight and count the three shared observations twice. If the test had two conditions, the same question union would require ten first-pass runs before any planned repeats.
The bridge measures a difference between definitions on one set of observed answers. It does not establish that the difference will always be 25 points. A future answer to D or E could change. Preserve the common-question comparison and start the new full-core series on its declared date. Avoid adding a fixed adjustment to past scores unless a separate, defensible method supports it.
Decide what to do when an overlap run is impossible
A branch closure, discontinued service or unavailable product can make an old prompt unsuitable to repeat. Record the reason and date. A question about a closed branch might still be useful as an error check, but it serves a different purpose from monitoring an active route. Keep that diagnostic outside the comparable core unless you redefine the panel explicitly.
Use the surviving common questions to preserve part of the comparison where the conditions still match. Show the overlap and how much of each panel it represents. If the overlap is too small to inform the decision, report a break. A change in the data collection setup can also create a break even when the prompt texts stay the same. Save the earlier protocol so the next analyst can trace why the series changed.
When an observation is missing, distinguish it from an out-of-scope prompt. In the table, E was outside the old core in the earlier period. It was not observed under that definition, so the team cannot score it absent. A capture failure on A would be a missing result for a prompt that was planned. Keep both statuses visible in the export and define which rows enter the denominator.
If you must change several prompts at once, list the reason for each replacement. A panel can lose coverage even when its number of prompts stays the same. Replacing two branch questions with two discovery questions changes the tasks represented. Keep a task-level inventory and check whether the surviving panel still supports the original decision. If it serves a different decision, give the new series a descriptive name and preserve the earlier record.
For a large panel, a bridge may cover only the questions that changed and a designated stable subset. Report that smaller scope explicitly. The overlap calculation then supports a narrower comparison, and the headline should state which questions it includes.
Report the core and exploratory results without mixing them
Publish the core version and the count of eligible answers near the headline metric. Add a short exploratory section with the tasks tested, why they were chosen and what was observed. Give any confirmed issue an owner. If a question is pending promotion, state the review date and evidence needed. That keeps new customer concerns visible without changing the recurring score.
The AI Visibility monthly reporting guide shows why the distinction matters. Its published example records 24 of 31 baseline website placements retained on a second asking, compared with 17 of 31 when a qualifier was added, summed across four Dubai sectors on one date in Google AI Mode. Those are source-retention figures for that study. They illustrate that exact repeats and wording variants can produce different observations. They do not predict the effect of adding a particular question to your panel. AI Visibility monthly reporting guide
Keep a small change log with the outgoing ID, incoming ID, effective date, reason and bridge scope. Record whether the customer task changed or the wording alone changed. Make the report distinguish a new finding about an existing question from a finding made possible by asking a new one. Those are both useful to the business, and each supports a different comparison.
Frozen prompt set record
Save the core version, exact prompts and conditions, scoring rules and effective dates. Keep the exploratory queue and promotion decisions alongside it. When the core changes, retain the old results, report the bridge or series break, and show the comparable overlap. This lets the next reader understand the trend and see which new customer tasks the business is investigating.
AI Visibility is published by Lunasol, which offers GEO and AEO services. The bridge panel and answer outcomes are constructed. The published source-retention figures are identified separately. About AI Visibility



