# How to Version a Prompt After Its Meaning Changes By AI Visibility editorial team AI-assisted article. A team tests “Where can I get a leather bag repaired in Dubai?” It later adds “and collect it today.” The new question may still sit under leather repair in the spreadsheet, but it asks a provider to meet a deadline. If the business appeared less often after the change, the original result and the new one cannot be read as a simple visibility trend. Give the revised wording its own version and record which requirement changed. Keep the original answers attached to the wording that produced them. If both questions still matter, test them side by side under matched conditions. This gives the team two useful observations and a record of why a direct before-and-after comparison would be misleading. Compare the customer task before comparing the words Read each prompt as a set of conditions a useful answer must satisfy. Identify the service, location, entity, time requirement and expected next action. A comparison question may ask for a shortlist, while a booking question needs an actionable route. A prompt asking about Dubai may permit a citywide recommendation, while a district-specific prompt can make some branches irrelevant. A small text edit can alter the set of satisfactory answers. Write an intent note for each version. For the original leather-repair prompt, the customer wants to find a suitable Dubai provider. For the revised prompt, the customer wants a provider who can accept the work and return the item today. That second task needs evidence about timing for the relevant service. Ordinary opening hours can show when the shop is open, but they cannot establish that a repair will be ready before closing. Inspect changes to implied assumptions as well as explicit words. Switching from “Can I get a quote?” to “Where do I send photos for a quote?” assumes there is a photo-based route. The first question can be answered by explaining available options. The second asks for a particular route and could tempt an assistant to supply one even when it does not exist. A reviewer who can state that difference before seeing the answer is less likely to change the scoring rule after the fact. A useful review question is whether an answer could satisfy the old task and fail the new one. If it can, record the condition that separates them. Also ask whether the target business can qualify under both definitions. A new requirement can remove it from the eligible candidate set for valid operational reasons. The resulting absence is evidence about the narrower task, not a direct measure of lost exposure on the original one. Save the exact wording and the reason for the version Give related prompts a family ID and each submitted wording its own version ID. For example, REPAIR-FIND can group a citywide discovery question and a timing variant, while REPAIR-FIND-v1 and REPAIR-FIND-v2 identify the actual text. Preserve the exact submitted string, effective dates, author, reviewer and change reason. Link the version to the scoring instructions and the reference evidence it needs. For a display-only correction, such as editing a descriptive title that was never sent to the assistant, update the display field and retain the submitted prompt. If you change the submitted wording, preserve both strings even when you think the meaning is equivalent. Decide separately whether the resulting observations qualify for the same trend. Pew Research Center's questionnaire guidance explains why exact wording and context matter when tracking change. That principle applies here as a reason to preserve the observation conditions. Pew Research Center guidance on writing survey questions (https://www.pewresearch.org/writing-survey-questions/) Store the prompt version, run conditions and scoring version independently. A prompt can remain unchanged while a reviewer tightens the eligibility rule or starts requiring a different kind of evidence. That can also break a trend. If you later rescore an earlier answer, keep the original code and the new code with their rule versions. A recoded answer still came from the old prompt. Use an explicit version reference in every capture row. A spreadsheet that links all runs to one editable text cell can silently rewrite the history when someone edits that cell. A version field should point to a fixed record. Preserve the old evidence and the date the new version entered collection. An analyst can then trace an unexpected score to either a changed answer, a changed prompt or a changed interpretation. Freeze the review rules before testing the revision Define the shared outcomes you can compare across versions. You might record whether the target business was named in either answer. That is a common textual measure. You might separately record whether the answer supplied a verified same-day route. That criterion applies to the revised task and needs different evidence. Keep it separate from the common mention measure. An official service example can help confirm that a business has performed a type of work. The Shoe Doctor DXB Instagram post below presents colour-change work. The post can support a statement about the work it depicts. It cannot establish the capacity or turnaround for a new job. In a versioned prompt audit, link each claim to evidence that supports its specific requirement. A new qualifier often requires a new source. For time-sensitive qualifiers, specify how the team interprets the wording. “Today” depends on the market and local capture date. A business can accept a same-day request without guaranteeing completion. The reviewer needs to know which condition the question asks for and which evidence is available. Record insufficient verification separately from a confirmed failure to meet the requirement. Keep the scoring rule independent of which version you expect to perform better. A response that names the business but gives no timing information can be positive for mention and unconfirmed for same-day feasibility. An answer that offers a valid contact route without naming the tracked business can satisfy one task criterion and fail the mention rule. Distinct outcome fields make those combinations reviewable. View the selected official Instagram post (https://www.instagram.com/theshoedoctor.dxb/reel/DfCWKA3SXl2/) Previously verified official Shoe Doctor DXB colour-change example. Turnaround requires separate evidence. Compare the versions in the same window where possible If both tasks remain relevant, run them under matched conditions. Keep the product, market, language and capture window comparable. Use the same session reset rule, and record the order. Where prior answers could influence a follow-up, submit the versions in separate sessions under the agreed procedure. Preserve any variation in timing and availability that could affect interpretation. The following constructed example uses a fictitious repair business. The family contains two exact prompts. Version 1 asks “Where can I get a leather bag repaired in Dubai?” Version 2 asks “Where can I get a leather bag repaired in Dubai and collect it today?” The first measure is whether a scorable answer names the target. The example assigns one pair to each of four observation blocks under the same protocol. The outcomes illustrate the analysis and were not collected from an AI product. Observation block Version 1 mention Version 2 mention Paired comparison A Yes Yes Both named the business B Yes No Only the original named it C No No Neither named it D Yes Capture unavailable No complete pair The original wording has three mentions in four scorable answers. The revised wording has one in three. On the three complete pairs, the counts are two of three for the original and one of three for the revision. The fourth block contributes evidence for the original version and no comparison for the revised one. Report it as missing instead of adding a negative outcome to the revised denominator. The fixture contains one discordant complete pair. That is a reason to inspect the answer, including whether the same-day condition influenced which businesses were eligible. It does not establish a stable effect for every future run. Even matched observational captures can differ because answers vary. If you want to measure timing accuracy too, add a separate evidence field for verified pickup feasibility. The table records only whether the brand was named. Published AI Visibility reporting provides context for preserving the change. On one date in Google AI Mode across four Dubai sectors, a second asking retained 24 of 31 baseline website placements, while a question with one qualifier added retained 17 of 31, counted within sectors and summed. Those results concern cited-source retention in that sample. They do not predict what a same-day condition will do to business mentions. AI Visibility monthly reporting guide (/article/what-should-a-monthly-ai-visibility-report-contain/) Keep the historical series attached to its original meaning If version 1 still represents an important customer task, keep it in the existing trend and run version 2 as a separate task. A dashboard can show both. Label the new series with the tighter condition rather than giving it the generic family name. The first observation under version 2 can become its own baseline once collection conditions and scoring are established. If version 1 is obsolete, close its series with a dated reason. A branch move might make the original location requirement irrelevant. In that case, preserve the old series and show where the new one begins. The old answers can still document what happened under the former setup. Copying their scores into the new series would imply they tested the new location. When a change also affects the scoring rule, show both changes. For example, a new version may ask for completion today, and a stricter rubric may require an explicit verified statement about timing. A drop in confirmed passes can reflect either part of the change. If retained evidence supports a same-answer recode for a shared outcome, publish that separately. Recoding cannot turn an answer to the old question into a run of the new one. The public AI Visibility test-sheet guide covers preserving the underlying questions and observations. Use that capture discipline and add immutable version links to the working record. A revision note should explain the exact difference and whether the earlier observations remain comparable. AI Visibility test-sheet guide (/article/how-to-check-what-ai-says-about-your-business-free-test-sheet/) Give each version an effective date for collection and a separate date for approval. An editor may approve the change before the first run, and some scheduled observations may still use the old wording in the meantime. Record the version actually submitted in each row. Do not classify it from the run date alone when both versions are active. When several teams reuse a prompt, publish the change note where they collect results. A branch team may continue a local diagnostic under the old version while another team adopts the new one. Keep both sets under their own labels. A shared family ID can make them easy to find without making them interchangeable. Give reviewers an explicit comparison decision Before publishing, ask a second reviewer to explain the change using the two exact strings. They should be able to identify the new condition, name the evidence it requires, and say which outcomes can still be compared. If they disagree about whether the task changed, keep the versions separate while you resolve it. This is a practical check on the definition rather than an assessment of the assistant. Use a direct reporting sentence. For the example, “The same-day version was introduced as a separate task. In three complete matched blocks, the business appeared twice for the original wording and once for the same-day wording.” Add the missing capture and the scorable counts if the reader will use the result. This gives the manager a reason to inspect the time-sensitive task without presenting a general fall in visibility. If the report serves several audiences, keep the short version note beside the headline and the full comparison in an appendix. A branch owner may need to know that a different question was asked. An analyst needs the exact wording, conditions, pair IDs and scoring history. Both should be able to trace the number to the version that produced it. Prompt version record Save the family and version IDs, exact text, task definition, changed requirement, evidence needed, dates, reviewer, and separate scoring version. Attach each run to its fixed version and mark any missing paired observations. State whether the change creates a new task, a parallel diagnostic or a break in the trend. The record should let another analyst recover both what the team asked and what it was able to compare. AI Visibility is published by Lunasol, which offers GEO and AEO services. The repair prompts and paired outcomes are constructed examples. The historical source-retention figures are identified separately. About AI Visibility (https://aivisibility.ae/about/)