By Samuel Anan
A brand name can appear in an AI answer for several reasons. It may be the subject of a question, one item in a list, an example of a category or an option the answer presents as suitable for a particular task.
Those appearances should not all be reported as recommendations. The distinction changes what a business can reasonably conclude and what it should investigate next.
I would rather report a modest observation accurately than use a larger count built from incompatible meanings.
A mention establishes appearance in that answer
At its simplest, a mention means the name appeared in the recorded response. It says nothing by itself about tone, accuracy or suitability.
An answer might mention a company to explain that it does not provide a service. It might repeat an outdated description. It might name the brand only because the question already included it.
That last distinction matters in reporting. Branded questions can be useful for checking accuracy, but they should not be mixed with unbranded discovery questions as though both measure spontaneous recommendation.
The AI observation method records the exact question and context so these differences remain visible.
A description needs factual review
A description can be favourable and wrong. Before treating it as a positive result, compare the claims with the business's approved facts.
Check the offer, location, customer type and any consequential limitation. A correct name beside an incorrect service area can send an unsuitable enquiry to the business or mislead the reader.
I use the claim-by-claim description review rather than assigning one impressionistic score to the whole paragraph. A response can be accurate about identity and unsupported about outcomes at the same time.
A report should preserve that mixed result. Removing the error because the overall answer sounds positive defeats the purpose of the audit.
A recommendation requires a suitability claim
I use “recommendation” when the answer presents a business as a suitable option for the stated task. Even then, I record the conditions and the basis visible in the answer.
“Here are several companies in this category” is not necessarily the same as “this company fits your requirement because…”. The wording and context matter.
A recommendation can also be weakly supported. The answer may claim suitability based on a source that establishes only a general service description. It may overlook a constraint in the question.
The useful next step is not simply to celebrate the appearance. It is to inspect whether the stated reasons are accurate and relevant to the customer's decision.
A citation is a separate event
An answer may link to the business's site without recommending the business. It may also recommend the business while citing a third-party source.
Record the destination and the claim it appears to support. Do not assume that a visible link transfers approval to every statement in the response.
OpenAI's web-search explanation describes how sources can appear in its search experience. The reader still needs to open the source and check the relevant passage.
Google's AI features guidance concerns its own search experiences and website requirements. Neither source provides a basis for treating every citation as a guaranteed endorsement or a permanent placement.
Compare like with like
An illustrative audit might contain twelve recorded answers: four include the brand, two describe it accurately, one presents it as suitable and three contain a relevant source link. Those figures are hypothetical, and the categories can overlap.
Adding them together would not produce ten separate successes. The same answer may contribute to several categories.
Instead, keep an answer-level record. For each observation, note whether the name appeared, whether the key facts were correct, whether suitability was asserted and which sources were cited.
This allows the business to see what changed. More appearances with persistent factual errors suggest a different priority from fewer appearances with accurate, relevant descriptions.
Keep absence proportionate too
An absent brand in a recorded answer is useful information, but it does not establish that the service never mentions the brand or that a technical penalty exists.
The result may depend on the question, context, available sources and other factors you cannot fully observe. Investigate the public information and relevant access conditions rather than jumping to a single explanation.
The business-facts audit remains a sensible first step. Correcting a consequential public error is justified independently of whether it changes the next generated response.
Avoid repeatedly altering the question until the brand appears and then treating that as an improvement. Exploratory testing should remain separate from the pre-agreed comparison set.
Connect the observation to a business decision
A mention count is not a business outcome. A visitor, enquiry or sale needs separate evidence, and the attribution may remain incomplete.
A useful report can say that a particular question now produces an accurate description in the recorded observations, while no reliable referral or enquiry evidence is yet available. That is an honest account of progress at one level.
This distinction is central to how I think about Sparlab: profiles, useful content and reporting should help businesses understand their public presence, not encourage them to mistake every visible number for growth.
The vocabulary in a report determines the conclusions people carry away. Use “mention”, “description”, “recommendation” and “citation” for the different events they describe. The resulting picture is more complicated than a single score, but it is also more useful.