If a business’s website is cited in an AI-generated answer, does that mean the business is actually being recommended?
I tested this by running the same commercially focused accountancy query through Google AI Mode ten times.
The relationship turned out to be surprisingly one-sided.
29 of the 30 recommendation positions included the recommended firm’s own website among the sources.
But when I looked in the opposite direction, only 29 of 60 accountancy-firm domain appearances belonged to firms actually recommended in that response.
In this experiment, recommendation usually came with citation.
Citation did not reliably mean recommendation.
That distinction matters if businesses are using citation counts as a measure of their AI visibility.
Why I ran this experiment
In an earlier article, Does Being Cited by AI Mean Your Business Is Being Recommended?, I looked at examples where citations and recommendations appeared to separate.
A business could be recommended while other websites supplied some of the supporting information.
Equally, a business website could be cited even though another company received the recommendation.
That left an obvious question:
How closely are citations and recommendations actually related?
This experiment was designed to put some numbers around that question.
The query I tested
I used the same customer scenario for all ten searches:
“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?”
This was deliberately more specific than a broad search such as:
“Best accountants in Birmingham.”
The query identified:
- the location;
- the industry;
- the number of employees;
- the services required;
- the accounting software already used;
- a preference for fixed monthly pricing.
It therefore represented a plausible commercial situation in which somebody could genuinely be looking for a new accountant.
I ran the unchanged query ten times through Google AI Mode on 9 August 2026.
Each run started as a fresh search.
I was not attempting to decide whether the recommended firms were genuinely the best choices, nor did I verify every claim AI Mode made about their services, qualifications, prices or sector experience.
The question was narrower:
Which firms did AI Mode recommend, and which accountancy-business websites appeared among the sources used in the response?
How I counted the citations
Some domains appeared several times within an individual response.
Counting every numbered citation separately would therefore have inflated the apparent visibility of a website simply because Google cited it repeatedly in the same answer.
Instead, I counted an accountancy-firm domain a maximum of once per response.
If the same accountancy website appeared as citations [2], [5] and [8] in one answer, that counted as one domain appearance for that run.
I also excluded obvious non-accountancy-business sources such as Google, Xero, LinkedIn and general reference or directory sites.
Where it was not sufficiently clear from the results that a domain represented an accountancy business, I excluded it rather than making an assumption.
The resulting figures should therefore be viewed as a conservative classification of the source set.
There is one other important distinction.
I counted a recommended firm’s website as cited if its domain appeared among the sources in that response.
That does not necessarily mean every claim made about the firm was supported by its own website, or that its website was the reason the firm was selected.
Several answers combined information from the recommended firm’s own site with third-party and competitor sources.
That became one of the most interesting observations from the experiment.
When a firm was recommended, was its own website cited?
Each response contained three recommendations.
Across ten runs, that created:
30 recommendation positions.
The recommended firm’s own website appeared among the sources in 29 of those 30 positions.
| Measure | Result |
|---|---|
| Recommendation positions | 30 |
| Recommended firm had own website cited | 29 |
| Recommended firm did not have own website cited | 1 |
| Recommendation accompanied by own-site citation | 96.7% |
That is a very strong relationship within this experiment.
If AI Mode recommended a firm, its own website was almost always somewhere within the source set.
The only exception occurred in the first response, when The Accountancy Solutions was recommended but its own website did not appear among the sources I recorded.
So recommendation without an own-site citation was possible.
It was simply unusual in these ten searches.
What happened when I reversed the question?
The more revealing result came from asking:
If an accountancy firm’s website appeared among the citations, was that firm actually being recommended?
Across the ten responses I recorded 60 accountancy-firm domain appearances.
Only 29 belonged to businesses that were one of the three recommendations in that particular response.
The remaining 31 belonged to accountancy firms that were not being recommended.
| Measure | Result |
|---|---|
| Accountancy-firm domain appearances | 60 |
| Appeared for a recommended firm | 29 |
| Appeared for a non-recommended firm | 31 |
| Cited firm-domain appearances associated with recommendation | 48.3% |
| Cited firm-domain appearances without recommendation | 51.7% |
Put another way:
More than half of the accountancy-business domain appearances occurred in responses where that business was not one of the three firms being recommended.
That is very different from the 96.7% figure in the other direction.
The relationship was asymmetrical
The experiment produced two very different results:
| Question | Result |
|---|---|
| If a firm was recommended, did its own website appear among the sources? | 96.7% |
| If an accountancy firm’s website appeared among the sources, was that firm recommended? | 48.3% |
So the practical conclusion is not simply that citations and recommendations are different.
It is more specific:
Recommendation was usually accompanied by citation in this experiment. Citation was not reliable evidence of recommendation.
That matters considerably when interpreting AI visibility reports.
The difference appeared throughout the ten runs
This was not caused by one unusual response.
| Run | Accountancy-firm domains cited | Belonged to recommended firms | Belonged to non-recommended firms |
|---|---|---|---|
| 1 | 4 | 2 | 2 |
| 2 | 5 | 3 | 2 |
| 3 | 6 | 3 | 3 |
| 4 | 5 | 3 | 2 |
| 5 | 6 | 3 | 3 |
| 6 | 9 | 3 | 6 |
| 7 | 4 | 3 | 1 |
| 8 | 9 | 3 | 6 |
| 9 | 5 | 3 | 2 |
| 10 | 7 | 3 | 4 |
| Total | 60 | 29 | 31 |
Runs 6 and 8 were especially striking.
Each answer recommended three accountancy firms.
But each source set also contained websites belonging to six additional accountancy businesses.
A report counting all of those domain appearances as equivalent AI visibility would therefore combine two very different outcomes:
- three businesses were actively presented to the prospective customer as firms to consider;
- six others supplied information to the answer without entering the recommendation shortlist.
Both may have some value.
They are not the same kind of value.
Companies999 shows the difference clearly
Companies999 provides perhaps the clearest individual example from the experiment.
It was recommended in three of the ten responses.
Its domain appeared among the accountancy-business sources in seven of the ten responses.
That gives two very different measurements:
Recommendation visibility: 30%
Citation presence: 70%
Neither figure is necessarily wrong.
They simply answer different questions.
If Companies999 monitored only whether its domain appeared among AI Mode’s sources, it could reasonably report visibility in seven out of ten searches.
But a prospective customer was actually told to consider Companies999 in only three of them.
In several other responses, its website contributed information while different accountancy firms received the recommendation.
Companies999 was not an isolated example. Other firms also moved between being recommended and merely appearing within the source pool across different runs.
Not all citations play the same role
This is probably the most important practical lesson from the experiment.
A citation can represent several different situations.
Recommended and cited
AI Mode recommends Firm A and uses Firm A’s website while explaining that recommendation.
This is citation visibility closely connected with recommendation visibility.
Cited but not recommended
AI Mode recommends Firms A, B and C.
Firm D’s website appears among the sources because one of its pages provides useful information about payroll, Xero, management accounts or another relevant subject.
Firm D has received a citation.
It has not entered the customer’s shortlist.
Cited while another firm is recommended
A firm’s content may also contribute information used within an answer that ultimately promotes one of its competitors.
That is still technically citation visibility.
Commercially, however, it is a very different outcome from being the business AI tells the customer to consider.
This suggests that AI visibility measurement needs to look beyond whether a domain simply appeared.
It also needs to consider citation context.
The responses drew on a much wider source pool than the recommendation shortlist
The answers did not look like simple summaries of three recommended firms’ websites.
Instead, they drew on a broader collection of sources covering subjects such as:
- manufacturing accounting;
- payroll;
- workplace pensions;
- Xero;
- management accounts;
- stock and work-in-progress;
- fixed-fee accountancy;
- tax relief;
- Birmingham accountancy services.
A website belonging to an accountancy firm could therefore contain information useful to the answer without that firm ultimately being selected as one of the three recommendations.
This helps explain why citation visibility was considerably broader than recommendation visibility.
It also explains why simply counting sources can be misleading.
A citation total needs context
Imagine an AI visibility report showing:
Your website received 70 AI citations this month.
At first glance that sounds encouraging.
But several questions remain unanswered.
Were those citations attached to recommendations of the business?
Did the website merely provide information about a general subject?
Did it support an answer in which another company received the recommendation?
Was the brand even mentioned alongside the citation?
A citation total alone cannot answer those questions.
That does not make citations unimportant.
It means they need context.
Citation visibility can still be valuable
A business does not necessarily need to be recommended for a citation to have value.
Being cited could still:
- create a potential click;
- expose the website or brand to a customer;
- show that the site contains useful information;
- contribute evidence to an AI-generated answer;
- create visibility earlier in a customer’s research journey.
The mistake would be treating all of those appearances as though they represented the same commercial outcome.
That links directly to the RAMP Framework.
The useful question is not simply whether visibility occurred.
It is whether that visibility happened in a commercially relevant query, what role the business played in the answer, and whether the response created a realistic pathway towards the business.
The recommendations themselves also changed
The shortlist was far from fixed.
Across the 30 available recommendation positions, 14 different accountancy firms appeared.
That reinforces the recommendation variability demonstrated in my earlier ten-run Google AI Mode experiment.
But recommendation variability is not the main finding here.
The important additional observation is that accountancy websites belonging to businesses outside the changing shortlist repeatedly remained within the source pool.
A business could therefore lose its recommendation position while retaining citation visibility.
What should we measure instead?
This experiment suggests that at least three different measures should be kept separate.
| Metric | Question |
|---|---|
| Recommendation visibility | Was the business actually presented as a suitable choice? |
| Citation visibility | Did the business’s domain appear among the sources? |
| Citation context | Why was the domain cited, and was the business itself being recommended? |
That final measure is particularly important.
Two businesses could receive the same number of citations while having completely different outcomes.
One might repeatedly be cited alongside recommendations of itself.
The other might repeatedly supply information to answers recommending competitors.
A citation count alone would make those situations look identical.
They are not.
Limitations of the experiment
These percentages should not be treated as universal Google AI Mode benchmarks.
The experiment used:
- one customer situation;
- one industry;
- one location;
- one AI platform;
- ten repeated searches;
- testing conducted during a short period on 9 August 2026.
A different business sector, query, location or testing date could produce very different results.
The accountancy-domain classification was also deliberately conservative. Where the nature of a source was unclear from the results, I excluded it rather than guessing.
The figures therefore describe this experiment.
They do not demonstrate that 96.7% of all AI recommendations will include the business’s website, or that only 48.3% of business citations generally relate to recommendations.
What they demonstrate is why the two measurements should not be assumed to mean the same thing.
What the ten searches revealed
The experiment began with a simple question:
When Google AI Mode recommends a business, does it cite that business’s website?
In these ten searches, overwhelmingly yes.
29 of the 30 recommendation positions included the recommended firm’s own website somewhere among the sources.
But the reverse question produced a very different answer:
If an accountancy firm’s website was cited, did that mean Google AI Mode was recommending the firm?
No.
There were 60 accountancy-firm domain appearances across the ten responses.
Only 29 belonged to firms actually recommended in that particular answer.
The remaining 31 — 51.7% — belonged to firms that were not recommended.
The practical conclusion is therefore:
Being recommended usually came with citation in this experiment. Being cited did not mean the business was being recommended.
So when evaluating AI visibility, don’t stop at:
Was our website cited?
Ask:
Why was it cited, what role did our business play in the answer, and who did AI actually recommend to the customer?
Because appearing in the source list and entering the customer’s consideration set are two very different outcomes.
Testing conducted using Google AI Mode on 9 August 2026. AI-generated answers, recommendations and source selection can change over time.