A customer writes “Google” in your enquiry form. The answer could mean a map listing, an ordinary search result or an answer shown inside search. Another customer tells your staff, “ChatGPT mentioned you, then I looked at Instagram.” If you store both responses in a dropdown with a single “online” option, you lose the distinction. If you automatically classify the first response as organic search, you add precision the customer did not provide.
Start by deciding which discovery question the business needs answered. Ask it in a neutral way, preserve what the customer said, and apply a source code only when the evidence supports it. Measure how often the question was offered and how many responses were usable. That gives a Dubai business a way to compare what customers recall across forms, phone calls and messages, even when web analytics cannot reconstruct the whole journey.
Decide what you want the customer to remember
“How did you first hear about us?” asks about discovery. “Where did you find our contact details today?” asks about a later step. “What made you decide to contact us?” asks about motivation. These questions can produce different answers for the same customer. Pick the question that serves the decision, and name the resulting field accordingly.
For a discovery record, a useful starting question is “How did you first hear about us?” It lets the customer mention a person, an assistant, a search tool, a shop sign or something you had not considered. If the team also needs the immediate contact route, ask about it in a separate field or follow-up. A customer can remember the route to your phone number while being unsure where they first heard your name. Preserve that difference.
The official Home Box UAE Instagram publication below illustrates a later touchpoint a customer might mention. Someone could encounter the brand in a shopping answer and then check a public Instagram post before enquiring. The post can establish what the brand published. The customer's response is the evidence for which step they recall and which they considered first.
Start with a neutral question and test how it lands
Use wording a customer can answer without knowing your channel taxonomy. A label like “acquisition source” may make sense to an analyst and little sense to someone trying to book a repair. Put any prompt about a specific assistant after the open response if you need it for a separate purpose. If a staff member starts with “Was it ChatGPT?”, retain that response as prompted and avoid combining it silently with spontaneous mentions.
Pew Research Center's questionnaire guidance explains how open and closed formats, response order and earlier questions can affect responses. For a business intake form, this means the same customer can answer differently if “AI assistant” appears prominently before they see the discovery question. If you use a fixed list for practical reasons, make the categories distinct, offer an “other” route, and allow for uncertainty or skipping. Document the options and their order. Pew Research Center on writing survey questions
Pretest the wording with the staff who will use it and a small group who can try answering it in realistic contexts. Ask what they thought “first” meant, whether they could recall a source, and how they would answer if two sources were involved. On a mobile form, check whether the question and any help text are visible together. On a call, check whether staff can ask it naturally without interrupting the enquiry. The U.S. Census Bureau describes cognitive interviews and debriefing as ways to investigate how respondents interpret questions. Adapt those ideas to the scale of your intake process. U.S. Census Bureau questionnaire testing methods
Review English and Arabic wording against the same intent. A fluent reviewer should be able to explain whether each version asks about discovery, contact route or motivation. If the translations shift the time reference or imply a specific source, revise the wording before comparing results. Keep the final text and the date it took effect.
Capture the response at a consistent point
Choose where the question fits in each intake channel. A form can ask it near the end as an optional field. A phone team can ask after it understands the reason for the call. A chat team can ask after addressing the opening request. Document when the question is normally offered. If one team asks only people who have already booked, its responses describe a different population from a team that asks all eligible enquiries.
Record exposure to the question separately from the answer. An optional field shown on a submitted form may have been left blank. A staff member may have skipped the question during a rushed call. A customer may explicitly decline or say they cannot remember. These are different states. Where a system cannot confirm that a question was displayed or asked, keep that state unresolved instead of assigning it from the existence of an empty field.
Keep one source response per reporting unit under a written rule. If a customer begins in chat and follows up by phone about the same job, link the channels to the enquiry where the match is reliable. If someone gives a later clarification, preserve the original response and add the clarification with its date. An operational note can tell the next reviewer which response was used without exposing contact details in the reporting export.
Code the evidence at the level it supports
Use a small codebook that staff can read while working. Keep the source wording or a brief source-only summary in the approved system where appropriate. Store the normalized source code in a separate field. A reviewer can then see when a response was specific enough to name a platform and when it only described a broad category. Code the discovery and contact route separately where both are known.
| Illustrative response | Discovery code | What else to record |
|---|---|---|
| “ChatGPT suggested you, then I checked Instagram.” | Named assistant | Instagram was a later touchpoint |
| “A colleague sent your number.” | Personal recommendation or sharing | Original discovery beyond the colleague is unknown |
| “Google.” | Google, surface unspecified | Clarify neutrally if useful and feasible |
| “The Google Maps listing showed your address.” | Maps if stated as discovery | Treat as a contact route if that is what the customer meant |
| “An AI answer listed you, but I don't remember which.” | Assistant reported, provider unspecified | Retain the customer's description |
| “I don't remember.” | Unknown recall | Question answered, source unresolved |
For ambiguous responses, use one neutral clarification when practical. “What do you remember seeing?” can help distinguish a map, link list and assistant answer without naming a preferred platform. A response can still remain unresolved. If a customer says an assistant named the business but cannot identify which one, your codebook can count it under assistant-reported discovery while leaving provider blank. A generic “Google” is usable for a broad Google-source total, but its surface stays unknown if the metric requires distinguishing Maps, ordinary search and an AI answer.
Periodically have two reviewers code the same small set. Discuss where they differed and update the examples. If the reporting definition changes, date the change and decide whether earlier responses can be recoded from retained evidence. Keep both views labeled if the old and new figures serve different comparisons. Changing the codebook can move totals even when customers are answering the same way.
Report coverage before you interpret the source mix
In this constructed example, a business logs twelve distinct enquiries eligible for the discovery question. It records offering the question ten times. Eight customers give a response, and six replies are specific enough to classify under the report's rules. Two of those six describe AI discovery and four identify another source. The remaining two replies are unresolved. Two customers leave the offered question unanswered and two enquiries were never offered it.
Show the four useful denominators beside the outcome. The question was offered on ten of twelve eligible enquiries. Eight of ten offered enquiries produced a response. Six of eight replies were usable for the intended classification. Two of twelve eligible enquiries had reported AI discovery. If you also report two of six usable replies, label it as a result for the usable subset. The six enquiries without a usable source could have a different discovery mix.
Compare collection rates by channel when interpreting a shift. Suppose the first month includes mostly forms with an optional question and the second month includes more phone enquiries handled by staff who ask routinely. A rise in reported assistant discovery could partly reflect the change in how often the question was asked. Keep the intake mix and response counts visible. If you revise the prompt, preserve the date and version so the team can see when the change entered the series.
A compact note can say, “Of twelve eligible enquiries, ten were offered the discovery question and six gave a classifiable answer. Two reported AI discovery.” Add the reporting period and which enquiries were eligible. This communicates the finding without borrowing confidence from the enquiries whose source is missing.
Treat comparisons between question versions as a separate exercise. If you try an open prompt on some enquiries and a fixed list on others, keep the version with each response and check how often each was offered. A change in the proportion naming an assistant could reflect the options customers saw. Keep the version-specific results visible until you have a defensible basis for combining them.
Use self-report alongside measured acquisition
A customer can report discovering the business through an assistant and then reach its website through ordinary search. Analytics and the enquiry response can both be useful because they describe different moments. Keep the self-reported discovery source and the measured session source in separate fields. If the records can be connected under your approved workflow, report how many were matched. If no link is established, keep them as separate aggregate views.
When the records differ, read the original evidence before changing either one. An assistant mention in the response could refer to research that happened days earlier. A measured campaign could describe only the latest recorded visit. The discrepancy can help you understand the journey. It does not establish which touchpoint caused the enquiry.
The existing AI Visibility referral-to-enquiry guide covers how to bring contact records and analytics into a broader report. Use the discovery questionnaire to make one part of that report consistent across forms, calls and messages. The team can also use recurring uncoded responses to decide which source descriptions need better guidance. AI Visibility referral-to-enquiry guide
Customer discovery question record
Save the exact question text by language and channel, date and version, point in the intake flow, coding rules, and examples of ambiguous answers. Publish eligible, offered, answered and usable counts next to the number reporting AI discovery. Keep the measure tied to customer recollection and retain an explicit unknown state. That allows the next reader to see which difference reflects what customers said and which might reflect how the team collected it.
AI Visibility is published by Lunasol, which offers GEO and AEO services. The sample responses and totals are constructed to illustrate the method. About AI Visibility



