Much of the discussion around AI search visibility focuses on a simple question:
Did the business appear in the answer?
That is obviously important.
If somebody asks an AI search engine to recommend an accountant, cleaning company or piece of software, a business would usually prefer to be included rather than invisible.
But our recent experiments exposed another question that may be just as important:
What did the AI actually tell the potential customer about the business?
A recommendation is not just a business name.
AI search systems usually explain why each business might be suitable. They may describe its specialisms, locations, pricing, experience, features, compliance knowledge or the type of customer it is best suited to.
So we decided to take some businesses that Google AI Mode had already recommended in our experiments and compare those descriptions with what the businesses actually publish on their own websites.
The results showed why measuring mentions alone can miss a large part of the picture.
Recommendation Visibility Is Only the Beginning
Our earlier experiments had been looking at how recommendations changed as a customer’s circumstances became more specific.
For example, a general search for a commercial cleaner in Leeds produced a fairly varied group of companies.
When we changed the customer to a dental practice in Leeds, a much more specialist set of businesses appeared.
Three of those companies were:
- Ashworth Commercial Cleaning
- It’s Clean
- Top Notch Cleans
We had also seen a similar effect with accountants.
EasyAccounts did not appear in any of our initial searches for a small limited company in Bristol, or our searches for a small construction company.
But once we specified that the construction company regularly used subcontractors, EasyAccounts appeared in all three runs.
That gave us four businesses with particularly interesting AI representations to examine.
The method was simple.
We took the sort of claims AI Mode was making or implying about each business and checked whether they were supported by the business’s own website.
We weren’t trying to prove whether every claim was objectively true.
That would require a much wider investigation.
We were asking a narrower question:
Does the way AI describes the business match the proposition the business itself publishes?
Case 1: Ashworth Commercial Cleaning
Ashworth appeared in all three of our searches for a commercial cleaning company for a small dental practice in Leeds.
Its own website provides a dedicated page for medical and dental practice cleaning.
A large proportion of AI Mode’s representation was strongly supported.
The website explicitly discusses:
- dental and medical practice cleaning;
- Leeds and West Yorkshire;
- colour-coded microfibre systems;
- preventing cross-contamination;
- COSHH procedures;
- cleaning reports;
- audit trails;
- treatment-room surfaces;
- DBS-checked staff;
- cleaning designed around CQC expectations.
So if AI describes Ashworth as a Leeds cleaner with specific experience and procedures suitable for dental and healthcare environments, there is substantial published information supporting that description.
But one phrase stood out.
HTM 01-05
HTM 01-05 repeatedly appeared in AI Mode’s discussion of what dental practices should consider when choosing a cleaner.
However, when we reviewed Ashworth’s dental-practice page, we could not find HTM 01-05 mentioned.
That does not mean Ashworth has no knowledge of it.
It means something much more limited:
We could not verify that particular association from Ashworth’s own website.
That distinction matters.
AI might have obtained the information from another source.
It might have inferred that a suitable dental cleaner would understand HTM 01-05.
Or it may have combined general information about dental-practice cleaning with information specifically about Ashworth.
Whatever the explanation, the potential customer sees the resulting synthesis as one answer.
AI Also Left Things Out
There was another side to the Ashworth example.
Its website contains some quite distinctive operational features, including things such as:
- GPS-verified attendance;
- photographic records;
- reports following cleaning visits;
- customer-accessible service records.
Those could potentially be very meaningful differentiators for someone managing a dental practice.
Yet they were not necessarily the features AI chose to emphasise.
So representation accuracy is not only about asking:
Did AI say anything wrong?
It can also involve:
Did AI leave out something important about why this business is different?
Case 2: It’s Clean
It’s Clean was another company recommended in all three dental-practice searches.
Again, the website strongly supported the core recommendation.
It explicitly discusses:
- dental-practice cleaning;
- Leeds;
- more than 20 years’ experience cleaning dental practices;
- treatment rooms and waiting areas;
- flexible scheduling;
- out-of-hours cleaning;
- CQC cleaning standards;
- Regulation 15.
Overall, AI Mode had identified the business proposition reasonably well.
But we again found examples where the AI wording could be stronger than the exact wording on the website.
For example, describing a service as being designed around particular regulatory requirements is not necessarily identical to saying that individual members of staff are specifically trained or certified in those regulations.
The distinction can look small when reading an AI-generated recommendation.
Commercially, it may not be small at all.
We again could not verify HTM 01-05 from the company’s own published dental-practice information.
The interesting finding was therefore not that the AI description was completely wrong.
It wasn’t.
It was mostly right, with elements of interpretation added on top.
That may be a much more realistic representation problem than a spectacular hallucination.
Case 3: Top Notch Cleans
Top Notch Cleans gave us a different result.
When we checked its dental-practice material, many of the specific claims associated with it in AI Mode were directly supported.
Its website explicitly discusses:
- dental-practice cleaning in Leeds;
- HTM 01-05 expectations;
- infection-control procedures;
- DBS-checked cleaners;
- colour-coded equipment;
- clinical-grade products;
- photographs following visits;
- video walkthroughs;
- documented checklists;
- early-morning, evening and weekend cleaning.
So in this case, even the HTM 01-05 association that we could not verify for the previous two companies was clearly present.
That is important because this experiment shouldn’t become an exercise in looking for AI mistakes.
Sometimes the representation is very good.
But Top Notch Exposed Another Problem: Which Business?
We had already encountered another issue while investigating this company.
There are similarly named cleaning businesses online.
We initially found a London-based company called Top Notch Cleaners, which was not the same business as Top Notch Cleans being recommended in our Leeds dental-practice experiment.
That sounds like a simple research mistake.
But it exposes a potentially important problem for AI visibility monitoring.
If a reporting system simply records:
Top Notch
or even:
Top Notch Cleaners
has it definitely identified the correct company?
Business names are not always unique.
There may be:
- similarly named businesses;
- different regional businesses;
- trading names;
- parent companies;
- old websites;
- directory listings;
- alternative domains.
So AI recommendation monitoring may need to track the actual entity and domain, not merely the displayed name.
Otherwise a business might appear to have visibility that actually belongs to somebody else.
Case 4: EasyAccounts
We then moved away from dental cleaning to see whether the same representation questions appeared in a completely different sector.
Earlier, we had asked AI Mode to recommend an accountant for:
a small construction limited company in Bristol that regularly uses subcontractors.
EasyAccounts appeared in all three runs.
That was particularly interesting because it had appeared:
0/3 times for a general small limited company in Bristol,
0/3 times for a small construction company,
but:
3/3 times once subcontractors were introduced.
Only afterwards did we inspect the website.
What we found was remarkably specific.
EasyAccounts doesn’t simply say it works with construction businesses.
Its website explicitly discusses:
- construction businesses;
- CIS contractors;
- CIS subcontractors;
- subcontractor verification;
- CIS deductions;
- monthly CIS returns;
- CIS payroll;
- construction VAT;
- cloud accounting software;
- fixed-fee services.
Its navigation even specifically refers to CIS Contractors & Subcontractors.
In this example there was very little need for AI to infer that EasyAccounts might be appropriate for a construction business regularly using subcontractors.
The business publishes that proposition very clearly.
Supported Does Not Necessarily Mean Independently Verified
EasyAccounts also helped expose another important distinction.
Suppose a business website says:
We are the highest-rated accountants in Bristol.
If AI repeats that statement, we could say:
The AI representation is supported by the company’s website.
But that does not automatically mean:
We have independently verified that the business really is the highest-rated accountant in Bristol.
Those are different tests.
Our representation audit is primarily about whether AI accurately reflects what a business publishes.
Some claims — especially rankings, review statistics, success rates, savings figures or performance claims — may require separate independent verification.
That means a useful AI representation audit probably needs more than a simple true / false classification.
A Better Representation-Accuracy Framework
From these examples, we ended up with several useful categories.
Directly Supported
The business’s own website clearly makes the claim.
For example:
Top Notch Cleans supports dental practices with cleaning designed around HTM 01-05 expectations.
That information is explicitly published.
Reasonably Inferred
The website provides relevant evidence, but the AI has interpreted or strengthened it.
For example, a business may say its cleaning procedures help practices meet regulatory requirements, while AI describes its staff more specifically as being trained in a particular regulation.
The general idea may be reasonable.
The precise wording may go beyond what was actually published.
Not Verified From the Website
We cannot find supporting information on the company’s own site.
This does not automatically mean the claim is false.
AI may be using another source.
But it means the claim cannot be confirmed from the business’s own published proposition.
Contradicted
The business’s own information directly conflicts with the AI description.
We did not find a strong example of this in these four audits, but it is an important category to monitor.
Published but Not Independently Verified
AI accurately repeats something the business claims, but we have not independently established whether the underlying claim is true.
Important Information Omitted
AI’s description may be factually accurate while still leaving out something commercially significant.
For a business owner, that could be nearly as important as an incorrect fact.
An AI Recommendation Is a Synthesis
These examples reinforce something we have repeatedly observed in our AI Mode experiments.
The answer presented to the searcher does not necessarily behave like a simple summary of one webpage.
AI may combine:
- the company’s own website;
- third-party websites;
- reviews;
- directories;
- general industry knowledge;
- other search results;
- inferred customer requirements.
The resulting description can therefore contain different types of information blended together.
A prospective customer doesn’t necessarily know where one fact ends and an inference begins.
They simply see something such as:
Best for dental practices needing strong compliance procedures.
That short statement may contain:
- a genuine website claim;
- an AI interpretation;
- an industry assumption;
- information from a third party;
- or some mixture of all four.
This is why checking the citation alone may not tell the whole story either.
Visibility and Representation Are Different Measurements
The experiment has helped clarify something important about AI visibility testing.
There are at least two separate questions.
Recommendation Visibility
Does AI recommend the business?
We can measure this by repeatedly running commercially relevant queries and recording how often the company appears.
Representation Accuracy
When AI recommends the business, what does it tell the customer about it?
That requires looking at:
- which characteristics AI highlights;
- whether those characteristics are supported;
- whether anything has been strengthened or inferred;
- whether important differentiators are omitted;
- whether the correct business entity has been identified.
A company could therefore have excellent recommendation visibility but poor representation.
Imagine a business appearing in 10 out of 10 relevant searches.
That sounds fantastic.
But suppose AI repeatedly tells customers:
the company only serves large businesses
when it actually specialises in small companies.
Or:
pricing starts at £500 per month
when no such pricing exists.
Or:
the company does not provide weekend services
when weekend availability is one of its biggest selling points.
The visibility score alone would look excellent.
The customer experience could be very different.
There May Be a Third Measurement Too: Citation Visibility
This also connects with something else we have seen.
A business can be recommended without its own website necessarily being the source used to support the recommendation.
That means we may eventually need to distinguish between at least three different measurements:
Recommendation visibility
How often is the business recommended?
Citation visibility
How often is the business’s own website cited or used as a source?
Representation accuracy
How accurately is the business described?
Those three numbers could tell very different stories.
A company might frequently be recommended but rarely cited.
Or frequently cited while not being directly recommended.
Or frequently recommended while being inaccurately represented.
Simply reporting one overall AI visibility percentage would hide those differences.
Why This Matters Commercially
There is a temptation to think that once AI starts mentioning your business, the job is done.
But consider the actual customer journey.
Someone asks:
Which commercial cleaner should I use for my dental practice in Leeds?
AI recommends three companies.
The customer may never visit all three websites.
They may initially decide which one looks most suitable based entirely on the AI-generated descriptions.
That means the few sentences AI writes about each company can influence which website gets clicked.
The business therefore has an interest not only in appearing, but in being understood correctly.
And this links directly with another finding from our previous experiments.
Make It Clear When Your Business Is the Right Fit
Across accountants, cleaners and software products, we repeatedly found that businesses gaining visibility for more specific searches often had websites that clearly described those customer situations.
EasyAccounts is a particularly strong example.
Its site doesn’t merely say:
We are accountants.
It explains that it works with construction businesses, CIS contractors and subcontractors, monthly returns, subcontractor verification, payroll and the other operational issues that come with that customer situation.
Top Notch Cleans does something similar for dental-practice cleaning.
This doesn’t prove that publishing those pages causes an AI recommendation.
But it does demonstrate something businesses can control:
How clearly their website explains when they are the right choice.
The clearer that proposition is, the more accurate information is available for customers — and potentially for AI systems — to understand.
A Practical AI Representation Audit
A business could run a simple manual test today.
Start with several genuine customer queries for which you would reasonably hope to be recommended.
Run each query several times.
If your business appears, record exactly what the AI says about you.
Then compare every meaningful claim with:
- your own website;
- the cited source;
- other authoritative information where necessary.
For each claim, ask:
- Is it directly supported?
- Has AI interpreted or strengthened it?
- Can we verify it at all?
- Is it actually about the correct business?
- Has AI omitted something important?
- Would the description give a prospective customer an accurate impression of us?
That final question may be the most important one.
The Bigger Lesson
Our experiments started with a relatively simple question:
Does this business appear in AI search?
The more we test, the less sufficient that question appears.
Businesses do not simply need to know whether they have AI visibility.
They may also need to understand what that visibility looks like.
Because there is a major difference between:
AI mentioned our company
and:
AI understood our company correctly and explained to the customer why we were relevant.
That suggests a broader principle for AI visibility testing:
Don’t just measure whether AI talks about your business. Measure what it says when it does.
A recommendation is valuable.
An accurate recommendation is better.