Businesses cannot currently obtain a reliable list of every long conversational question people ask in Google AI Mode. Fortunately, they may not need one.
AI visibility can be tested around commercially important customer needs rather than exact long-tail queries.
The aim is not to predict every sentence a potential customer might type. It is to select a representative group of natural questions and test whether Google connects the business with the underlying need.
Why exact AI query discovery is difficult
Traditional search research provides several clues about customer language. Autocomplete suggests common formulations, keyword tools group recognisable phrases, and Search Console shows some of the queries for which a website received impressions or clicks.
Conversational AI searches are harder to uncover. A customer can describe the same situation in dozens of ways, adding details about location, industry, company size, software, problems and commercial preferences. Each exact formulation may be rare even when the underlying need is common and valuable.
Google says AI Mode is particularly useful for nuanced questions involving exploration, reasoning and comparisons. It also says people can ask questions that might previously have required multiple searches. AI Mode may then use query fan-out, issuing related searches across subtopics and sources to construct its response. Google explains this process in its guidance on AI features and websites.
An owner searching for an accountant might previously have searched separately for:
- Manufacturing accountants Birmingham
- Xero accountants Birmingham
- Outsourced payroll Birmingham
- Monthly management accounts
- Fixed-fee accountants
In AI Mode, those requirements can become one question:
I run a 25-employee manufacturing company in Birmingham. I use Xero and need an accountant to handle payroll, year-end accounts, corporation tax and monthly management accounts for a fixed monthly fee. Which firms should I consider?
That exact sentence may have little measurable search history. But it combines several credible requirements that a real customer could previously have researched separately.
Google’s newer generative-AI Search Console reports provide information such as impressions, visible pages, countries, devices and dates. However, Google’s published description does not currently list the original AI Mode questions as an available dimension. The limitations of the report reinforce the difficulty of discovering complete conversational queries.
Demand research and visibility testing answer different questions
It helps to separate two issues:
- Does sufficient customer demand exist for this service?
- Does AI connect our business with this customer need?
Keyword research, customer interviews, enquiries and market research can help answer the first question.
AI visibility testing is primarily concerned with the second.
An accountancy firm that genuinely wants to attract established manufacturers using Xero does not need proof that hundreds of people submit one exact 30-word prompt before testing whether Google understands that proposition.
It does need a credible commercial theme grounded in evidence such as:
- Services the firm genuinely provides
- Problems raised by existing clients
- Questions asked by prospective customers
- Website enquiry forms and sales conversations
- Search Console data where available
- Industry discussions
- Case studies and testimonials
- The firm’s specialist experience
The question becomes:
When this credible customer need is expressed in several natural ways, does Google understand that our business is relevant?
Define a customer-need theme
Start by describing the commercial proposition for which the business wants to be considered:
Xero accounting support for established manufacturing businesses in Birmingham.
This is not a target keyword or finished prompt. It is a customer-need theme.
The theme can be broken into the elements that genuinely influence the customer’s decision:
| Element | Question to answer |
|---|---|
| Customer | Who needs the service? |
| Problem | What are they trying to resolve? |
| Service | What support do they require? |
| Sector | Does specialist industry knowledge matter? |
| Location | Is geographic proximity relevant? |
| Technology | Does particular software matter? |
| Commercial preference | Are fees, delivery or contract terms important? |
Not every theme needs every element. The purpose is to identify the details that matter to this particular customer—not to manufacture the longest possible prompt.
Create three kinds of query variation
The prompt group should contain more than several mechanical paraphrases. There are three useful forms of variation.
1. Paraphrase variants
These retain the same requirements but express them differently. They test whether ordinary changes in wording affect the results.
- Which Birmingham accountants specialise in Xero support for manufacturers?
- Which Xero accountancy firms work with manufacturing companies in Birmingham?
2. Emphasis variants
These retain the general need but foreground one important selection criterion.
- Which Birmingham accountants have strong manufacturing experience?
- Which accountants provide specialist Xero support to Birmingham manufacturers?
- Which fixed-fee accountants offer monthly reporting to manufacturing businesses?
3. Scenario variants
These represent different realistic combinations of requirements within the same customer theme.
- I run a 25-person Birmingham manufacturing company using Xero and need payroll, tax and monthly reporting. Which accountants should I consider?
- My manufacturing company struggles to understand stock, production costs and margins. Which Birmingham accountants could help us improve monthly reporting through Xero?
- We want to move an established manufacturing business from Sage to Xero. Which accountants can manage the migration and provide ongoing accounts and payroll support?
This distinction allows the test to investigate three separate questions:
- Do synonyms alone affect the recommendations?
- Does changing the emphasis alter which firms appear?
- Does adding a realistic customer scenario narrow or change the candidate pool?
Keep the first test unbranded
If ABC Accountants is the business being tested, its name should not normally appear in these prompts.
The purpose is to see whether Google independently discovers and recommends the firm for the customer need. Naming it would remove the discovery question and create a different test: whether AI considers a specified business suitable.
That named-business assessment is useful, but it should be conducted separately. The theme-based test asks:
Does Google independently connect our business with this need?
Similar meanings can still produce different recommendations
Google advises businesses that they do not need to create repetitive pages for every long-tail wording. Its systems can understand synonyms and the general meaning of what someone is seeking. Google includes this advice in its guidance for generative AI search.
That does not mean semantically similar questions will produce identical recommendations.
We have observed that changing the wording while retaining the broad meaning can alter the firms appearing in AI responses. Different formulations may affect which requirements receive the most emphasis, which related searches are generated through query fan-out, which pages are retrieved and which businesses enter the apparent recommendation candidate pool.
We cannot see Google’s internal weighting or hidden fan-out searches. This is therefore a plausible interpretation of the results rather than a proven explanation.
It does show why one carefully chosen prompt cannot represent an entire customer theme.
Repeat every query under consistent conditions
Wording is not the only source of variation. Repeating the exact same AI Mode query can also produce different recommendations.
If two query variants are each run only once, different answers do not prove that the wording caused the difference. The same fluctuation might have occurred if either query had simply been repeated.
A stronger test measures:
- Within-query variability: What changes when the exact prompt is repeated?
- Between-query variability: What changes when the same customer theme is expressed differently?
For a practical initial test, six prompts run five times each would create 30 responses. This is a diagnostic sample, not a statistically representative measurement of everything Google might show.
The testing conditions should be kept as consistent as possible:
- Use the same AI platform and mode.
- Start each run in a fresh conversation.
- Keep location and account conditions consistent.
- Use the exact recorded wording for every repetition.
- Run each prompt the same number of times.
- Complete the test within a reasonably short period.
- Record the date, recommendations, positions, descriptions and citations.
- Avoid changing the business’s website during the test period.
These controls cannot eliminate every source of variation, but they make comparisons more meaningful.
Measure visibility across the theme
The test should record more than whether the business appeared once.
| Measure | Question answered |
|---|---|
| Theme mention rate | How often was the business mentioned across all responses? |
| Query coverage | Across how many prompt variants did it appear? |
| Stable query visibility | For which prompts did it appear repeatedly? |
| Recommendation position | Where was it normally placed? |
| Citation rate | How often was its own website linked? |
| Description accuracy | Was the business represented correctly? |
| Competitor frequency | Which other businesses appeared most often? |
| Candidate-pool breadth | How many firms filled all available recommendation positions? |
If the business appears in 12 of 30 responses, its theme mention rate is 40%. That figure can help compare controlled tests, but it should not be presented as the percentage of all users who would see the business. The sample is too small and the system too variable for that claim.
Query coverage may be more revealing than the overall percentage. A firm appearing across five prompt types may have broader thematic visibility than one whose appearances are concentrated around a single formulation.
Use patterns to investigate evidence gaps
The results can identify questions that deserve further investigation.
| Pattern observed | Question to investigate |
|---|---|
| Appears for Xero prompts but not manufacturing prompts | Does the website prove sector expertise or merely list manufacturing? |
| Appears for manufacturing but not payroll | Is payroll capability and scale stated clearly? |
| Appears for broad prompts but not detailed scenarios | Is the evidence too general for specific customer circumstances? |
| Mentioned regularly but own site is rarely cited | Is Google relying on stronger external evidence? |
| Appears only for one formulation | Is the association dependent on a narrow wording or emphasis? |
| Appears but is described inaccurately | Is the published information ambiguous, inconsistent or outdated? |
These patterns do not diagnose the cause by themselves. A missing recommendation could also reflect stronger competitors, indexing, external authority, query fan-out or the limited sample. The next step is to examine the pages and citations supporting the results.
Do not construct the test backwards from the website
A business should not read its own website, extract every favourable detail and then write a prompt that reproduces those claims perfectly.
That would risk testing whether Google can match an engineered question with the source from which it was derived.
Define the customer need independently:
- Identify a commercially important customer group.
- Record the problems and requirements that group genuinely has.
- Construct natural unbranded prompts from different customer perspectives.
- Fix the prompts and method before viewing the results.
- Run every prompt the same number of times.
- Record recommendations and citations consistently.
- Only then compare the results with the business’s published evidence.
This keeps the exercise focused on customer relevance rather than creating an artificial success for the site being tested.
This is not a page-per-prompt content strategy
The purpose is not to create a near-duplicate webpage for every query variation.
A useful page or connected collection of pages should provide evidence relevant to several ways of expressing the same customer need. For a manufacturing accountant, that evidence might include:
- Demonstrated manufacturing experience
- Xero certifications and trained staff
- Relevant software integrations
- Payroll capability for established employers
- Monthly reporting services
- Knowledge of stock, production costs and margins
- Clear pricing arrangements
- Case studies and testimonials from comparable clients
The business is not trying to repeat every possible phrase. It is building a credible body of evidence around the theme it wants to be associated with.
A repeatable theme-based testing method
The complete method is:
- Define the commercial theme. Identify the customer need for which the business wants to be considered.
- Validate its relevance. Use enquiries, client experience, market knowledge and available search evidence.
- Create representative variations. Include paraphrase, emphasis and scenario variants.
- Keep the initial prompts unbranded. Test whether Google independently discovers the business.
- Control the conditions. Use fresh conversations, consistent settings and equal repetitions.
- Record more than mentions. Capture positions, citations, descriptions and competitors.
- Assess visibility across the theme. Look for broad and repeated association, not success for one preferred prompt.
- Investigate the evidence. Examine what the website proves, what it merely claims and what cannot be confirmed.
- Repeat after meaningful changes. Allow revised evidence to be indexed before rerunning the controlled test.
Test whether AI understands the need, not one sentence
No practical test can cover every way a customer might express the same idea. That does not make the exercise worthless. It means the chosen prompts must be treated as a representative sample rather than a definitive list of real searches.
The useful question is not:
Do we appear for this one exact AI prompt?
It is:
Across a representative group of commercially important customer questions, does AI understand that our business is relevant?
That acknowledges the limits of available query data, the variety of customer language and the fluctuation within AI-generated answers. At the same time, it gives businesses a disciplined way to test whether their intended market position is reflected in Google’s recommendations.
We may not know every question customers ask. But we can still identify the customer needs that matter and test whether AI connects those needs with the business.