When someone asks Google to recommend a business, how many businesses is it really choosing between?
We explored that question by giving Google AI Mode a detailed request for an accountancy firm:
I run a small manufacturing company in Birmingham with 25 employees. I need an accountancy firm to handle year-end accounts, corporation tax, payroll and monthly management accounts. I use Xero and would prefer a firm offering fixed monthly fees. Which three Birmingham accountancy firms should I consider, and why?
We submitted the same query five times and recorded the recommendations. That created 15 recommendation positions, but only eight different firms occupied them.
Considering the number of accountants serving Birmingham, that was a surprisingly restricted group.
The recommendations changed between runs, but Google did not appear to range freely across the entire local market. It repeatedly drew from a relatively small collection of plausible candidates.
That suggests a useful way to think about AI visibility: before a business can win one of the recommendation positions, it may first need to enter what we will call the AI recommendation candidate pool.
This was a detailed customer requirement
The query did not simply ask for three accountants in Birmingham. It contained several separate requirements:
- Birmingham
- Manufacturing experience
- Suitability for a company with 25 employees
- Year-end accounts
- Corporation tax
- Payroll
- Monthly management accounts
- Xero
- Fixed monthly fees
A firm might be capable of meeting every requirement without stating all of that clearly on its website.
That distinction became important when we examined the results.
Eight firms filled the 15 recommendation positions
Across the five runs, the recommendations came from this group:
- Agile Accountants
- Prime Accountants
- Linford Grey
- Morgan Reach
- Inform Accounting
- Gondal Accountancy
- Tax Care Accountants
- Dains
This is not evidence that Google considered only these eight firms internally. We cannot see its retrieval or selection process, and five searches are not enough to establish the full population of possible recommendations.
Nevertheless, the visible result is clear: a large local accountancy market produced a much smaller recurring set of recommendations.
Why these firms?
To investigate, we reviewed the pages linked in the answers where the complete URLs could be identified. We then examined relevant supporting pages elsewhere on the recommended firms’ websites.
That distinction matters. The page linked in an AI answer did not always contain all the information that appeared to justify the recommendation. Sometimes stronger supporting evidence existed on another page belonging to the same firm.
The firms were different kinds of partial match
No principal page we examined clearly covered every component of the query. Instead, different firms provided strong evidence for different parts of it.
| Strongest visible evidence | Firms that demonstrated it particularly clearly |
|---|---|
| Manufacturing expertise | Prime Accountants and Dains |
| Xero and fixed monthly support | Agile Accountants and Linford Grey |
| Birmingham, Xero, payroll and fixed fees | Tax Care and Gondal Accountancy |
| Broad Birmingham accounting services | Morgan Reach |
| Xero and management information | Inform Accounting |
| One page clearly covering the complete prompt | None found among the results |
Prime Accountants explicitly discussed manufacturing and production businesses, including their cost structures, financial processes and need for strategic financial information. Its website also established a Birmingham office, year-end accounts, tax and payroll capability. However, the evidence for Xero, monthly management accounts and fixed monthly fees was less complete.
Dains’ manufacturing page was the strongest industry-specific page we found. It discussed production costs, overheads, stock management, financial reporting and corporate tax for manufacturers. Dains also provided Xero, payroll and management-accounting evidence elsewhere on its website, but the fixed-monthly-fee requirement was not clearly established.
Agile Accountants clearly promoted Xero and fixed monthly pricing. Supporting pages described monthly management accounts and a service package capable of handling payroll for 25 employees. But its manufacturing credentials were not equally clear on the principal page linked with the recommendation.
Linford Grey’s Birmingham page connected Birmingham businesses with Xero and fixed monthly support. Related pages described statutory accounts, corporation tax, payroll and management reporting. Again, manufacturing expertise was the weaker part of the match.
Tax Care’s Birmingham page contained clear evidence about its Birmingham presence, Xero expertise, payroll and fixed monthly payments. It was a good match for much of the commercial arrangement, but it did not clearly demonstrate manufacturing specialism or monthly management accounts.
Gondal Accountancy’s Xero page established Birmingham, Xero, bookkeeping, payroll, accounts and tax support. A separate small-business accounting page added fixed fees, year-end accounts, corporation tax and payroll for employers with up to 50 staff. Manufacturing experience and an explicit monthly management-accounts service remained unclear.
Morgan Reach demonstrated broad Birmingham accounting capability, including accounts, tax, payroll and management information across its website. It also referred to fixed-cost work and Xero. But its current manufacturing evidence and commitment to a fixed monthly package were less obvious.
Inform Accounting had strong Xero credentials and clearly understood management information and finance support. Its visible proposition referred to quarterly management accounts, however, while fixed monthly fees and manufacturing experience were not clearly demonstrated.
These were all credible recommendations in different ways. None was an obvious complete match.
What happens when every visible candidate is incomplete?
When no candidate has clearly superior evidence across the complete requirement, Google must still produce an answer.
One response might place greater weight on manufacturing expertise, making Prime or Dains attractive. Another might prioritise Xero, monthly reporting and a predictable fee, strengthening the case for Agile or Linford Grey. A further response might give more weight to Birmingham presence, payroll and broad small-business support, favouring Tax Care, Gondal or Morgan Reach.
We cannot confirm that this is how Google weighted the candidates internally. It is an interpretation of the pattern rather than a fact about Google’s system.
However, it offers a plausible explanation for some of the variation:
AI recommendation volatility may sometimes reflect an evidence gap in the market, not only variability within the AI system.
If several businesses are credible but incomplete matches, relatively small changes in how the requirements are interpreted could change the final three.
Capability is not the same as visible evidence
Many more Birmingham accountants may genuinely be able to serve this customer.
They may already work with manufacturers, process payrolls of this size, use Xero and provide monthly management accounts for an agreed monthly fee. But Google cannot rely on information that exists only inside the practice or in conversations with existing clients.
It needs discoverable evidence.
A traditional accountancy website might divide its information into separate pages for:
- Accounts
- Corporation tax
- Payroll
- Management accounts
- Xero
- Industries
- Fees
- Testimonials
The information may technically be present, but the customer’s situation crosses all those divisions. Google must find the relevant facts and connect them before it can present the firm as a convincing answer.
Some details may not be stated at all. A page might mention payroll without indicating the size of employer the firm can support. It might offer management accounts without saying whether they are produced monthly, quarterly or annually. It might list manufacturing among several industries without demonstrating any understanding of stock, production costs or margins.
The firm may be highly capable while its website provides only a weak case for recommending it in this particular situation.
What would a more complete answer look like?
Imagine a Birmingham accountancy firm with substantial manufacturing experience creating a genuinely useful resource for established manufacturers.
The page would not merely say, “We work with manufacturing businesses.” It could discuss the issues those businesses commonly need help with:
- Stock and work-in-progress valuation
- Material and labour costs
- Production overheads
- Margins by product or production line
- Cash tied up in inventory
- Machinery and capital expenditure
- Cash-flow forecasting
- Connecting production or inventory systems with Xero
It could then explain how the firm’s services address those problems.
Monthly management accounts might include margin analysis, stock movements, overhead recovery, cash-flow forecasts and operational key performance indicators. Payroll support could cover production, administrative and management employees. Year-end accounts and corporation tax could be presented as part of the same continuing service rather than as isolated compliance jobs.
The page could also state clearly:
- The types and sizes of manufacturer the firm supports
- How it uses Xero and connected systems
- What is included in its fixed monthly package
- How frequently reports and review meetings are provided
- Which parts of Birmingham and the West Midlands it serves
Case studies could show how the accountants helped a manufacturer identify a deteriorating margin, improve stock reporting or plan an investment. Testimonials from Birmingham manufacturing clients could provide independent evidence that the expertise had been applied successfully.
That would create a much more complete body of evidence than a generic industry paragraph.
This does not necessarily mean putting everything on one page
The lesson is not that every relevant fact must be forced onto one enormous landing page.
Google may be able to combine evidence from a connected collection of pages. A focused manufacturing page could link naturally to detailed resources about management reporting, payroll, Xero, pricing and relevant case studies.
The important change is from isolated service descriptions to a connected explanation of how the firm helps a particular type of customer.
| Traditional website structure | Connected customer-situation structure |
|---|---|
| Accounts | Accounting support for Birmingham manufacturers |
| Payroll | Handling production and administrative payroll |
| Xero | Using Xero for timely manufacturing information |
| Management accounts | Monthly stock, margin and cash-flow reporting |
| Testimonials | Results achieved for comparable manufacturers |
The existing service pages can remain useful. The customer-focused resource connects them into a coherent answer.
Would a more complete resource appear more regularly?
Logically, it should give Google stronger evidence that the firm is relevant.
If Google’s objective is to provide a useful and well-supported response, an indexed website that closely addresses all the material parts of a query should be a plausible source for that response.
But this experiment does not prove that creating such a resource would guarantee inclusion. Other factors may matter, including the wider authority of the website, external references, reviews, location evidence, indexing and Google’s assessment of the source’s reliability.
We therefore need to separate three propositions:
- A complete, connected resource would be more relevant to the query.
- Its evidence should be easier for an AI system to interpret.
- It would consequently appear more frequently in recommendations.
The first two are logical expectations. The third is a testable hypothesis, not yet a conclusion.
Our working hypothesis is:
A website that clearly connects location, industry, customer circumstances, required services, software, pricing and real customer evidence will have a better chance of appearing consistently than websites that provide only fragments of that evidence.
Testing that properly would require finding or creating a genuinely comprehensive resource, recording its visibility before and after indexing, and repeating the same commercially detailed prompt enough times to measure whether its recommendation and citation frequency changed.
What the AI recommendation candidate pool means for businesses
This experiment suggests that a business may not move directly from being one of hundreds of local providers to occupying one of three recommendation positions.
There may be an intermediate stage: becoming a plausible candidate that Google can connect with the query.
For a business testing its AI visibility, that creates two questions:
- Does the business appear to enter the candidate pool for this commercially valuable requirement?
- Once it appears within that group, how often is it recommended, cited and accurately described?
A single search cannot answer those questions. It only reveals the final selection on one occasion. Repeated searches begin to show whether the business appears at all and which other providers repeatedly compete for the recommendation positions.
This may be more useful than treating every local competitor as equally visible.
What the Birmingham experiment showed—and what it did not
The observed findings were:
- Five identical searches created 15 recommendation positions.
- Only eight firms occupied those positions.
- The recommended firms provided strong evidence for different parts of the prompt.
- No principal page examined clearly answered the complete requirement.
- Some of the strongest supporting evidence was located away from the page linked in the answer.
Our interpretation is that Google may have been selecting from a relatively restricted candidate pool of credible but incomplete matches.
Our unproven hypothesis is that a clearer and more complete body of evidence would improve a firm’s chance of entering that pool and appearing consistently.
That distinction matters. We have not proved that reorganising a website around customer situations will produce more AI recommendations. But the test gives businesses a good reason to examine whether their websites provide the evidence an AI system would need to understand exactly who they help, what problems they solve and why they fit a detailed customer requirement.
The next step is to find a market where one business already provides that complete answer—or observe what happens when a business builds it.