The short answer
An AI mention means the brand appears in a usable answer and refers to the intended entity. An AI recommendation is narrower: the answer presents that brand, product, or service as a suitable choice for the user’s stated need. A source title, neutral comparison, historical reference, warning, rejected option, or brand named only because the user supplied it can be a mention without being a recommendation. Preserve the complete answer, classify the brand’s role with a written codebook, and keep recommendation role separate from order, factual accuracy, citations, and answer completeness. Report exact counts over a declared set of eligible answers instead of letting every string match become a commercial win.
Editorial method · evidence boundaries
This guide proposes a review method. Worked examples and pilot sizes are illustrations unless a dated study is explicitly identified. Provider documentation supports only the behavior and product surface it describes; it does not prove that an AIBL workflow improves recommendations or business outcomes.
The six-state codebook, decision tree, and worked example are AI Brand Lens editorial methods. They are not provider-native labels, an industry standard, or evidence that a provider intended a commercial endorsement. Teams should adapt and validate the rubric for their own buyer questions before automating it.
Historical AIBL records have limitations: completed jobs can contain truncated answers, and native citation metadata was not retained uniformly. Verify answer usability and source availability before scoring or comparing results. Missing saved metadata is not proof that a provider supplied no sources.
What businesses notice
Common signs of the problem
- A dashboard counts every brand-name string as a recommendation, including warnings, source titles, and options the answer rejects.
- Reviewers agree that the brand appeared but disagree about whether it was a first choice, a qualified option, or merely an example.
- A percentage improves after more branded fact-check questions are added, even though unaided buyer recommendations did not change.
- A cited brand is treated as recommended even when the cited material supports background context or another company’s selection.
01
Define the observation before assigning a label
The same sentence can play a different role under a different buyer question. Start with the complete observation—not a keyword hit—so the label stays connected to what the answer was trying to do.
Preserve the question and full answer
Store the exact question, provider surface, market or location context, date, complete returned text, visible sources, and collection status. A cropped sentence can conceal a qualification, later reversal, or explicit warning. If the answer is partial or completeness is unknown, apply the metric-specific eligibility rule before interpreting omission or final recommendation role.
Check incomplete-answer eligibility →Confirm the intended entity
A text match is not a valid brand mention when the surrounding product, domain, location, or company facts identify a namesake. Resolve entity identity before recommendation role. Otherwise a wrong-company answer can inflate both presence and recommendation metrics.
Resolve brand identity confusion →Name the buyer task
Label the question as unaided discovery, aided comparison, branded verification, troubleshooting, or another declared task. A brand supplied in a fact-check prompt has different evidentiary meaning from a brand that enters an unaided shortlist. Both answers may be useful, but they should not share an unlabeled recommendation denominator.
Choose the unit of analysis
Use one eligible question × provider × market × repeat answer as the basic cell. If the answer discusses several products, record brand and product roles separately. If a follow-up depends on conversation history, retain the path and do not compare it silently with a fresh standalone answer.
02
Use a six-state recommendation codebook
A binary mentioned/not-mentioned field is useful for broad presence, but it cannot describe the commercial role the answer gives the brand. Add one mutually exclusive role for each eligible entity-answer pair.
1. Absent
The intended entity does not appear in the usable answer. This is a negative presence observation only when the answer is eligible for omission. A failed, truncated, blocked, or wrong-question record belongs in collection quality—not automatically in absent.
2. Passing mention
The answer names the brand without treating it as a candidate for the stated need. Examples include background history, a source title, a customer example, a comparison reference, or a statement that another product integrates with it. Passing mentions count toward clearly labeled mention coverage, not recommendation coverage.
3. Considered option
The answer includes the brand in the decision set but does not affirm that it fits. It may invite the reader to investigate, list neutral pros and cons, or say the choice depends on missing facts. This state is commercially more meaningful than a passing mention, yet it is not an affirmative recommendation.
4. Qualified recommendation
The answer says the brand is suitable when an explicit condition holds—for example, a particular budget, integration, location, or use case. Preserve the condition with the label. Removing it can turn a bounded recommendation into a claim the answer did not make.
5. Primary recommendation
The answer presents the brand as the preferred, strongest, or leading fit for the stated need. Use this only when the selection is explicit. Do not infer primary status from first appearance, paragraph length, or visual order when the answer describes an unordered set.
6. Cautioned or rejected
The answer says the brand is unsuitable, ineligible, risky, or not preferred for the stated need. Keep the negative state visible; it still counts as an answer-body mention, but counting it as an affirmative recommendation reverses the answer’s meaning.
03
Apply a repeatable decision tree
Reviewers should reach the role through observable language and context rather than intuition about whether the answer feels positive.
First ask whether the entity appears
Check the answer body and any declared source fields separately. A brand that appears only in a source title may be a source-presence observation without being an answer-body mention. Document the search fields so two systems do not produce different counts from the same output.
Then ask whether it is a candidate
Would a reasonable reader understand that the answer is evaluating the brand as an option for this specific task? If not, use passing mention. Generic praise, factual description, or category membership is not enough when the answer never connects the brand to the user’s decision.
Then ask whether suitability is affirmed
If the answer merely places the brand under consideration, use considered option. If it says the brand fits subject to an explicit condition, use qualified recommendation. If it clearly selects the brand over the available alternatives for the stated need, use primary recommendation. Save the supporting sentence span for review.
Finally inspect reversals and caveats
Read the whole answer for disqualifying context. “Popular, but not available in your region,” “consider it only if,” and “avoid this option when” assign different roles. If mixed language cannot be resolved under the codebook, mark the case ambiguous for adjudication instead of choosing the label that favors the tracked brand.
04
Keep recommendation separate from nearby evidence
Recommendation role answers one question: how did this answer position the entity for this task? Other fields are necessary, but none can substitute for that role.
Do not infer rank from layout alone
Record first choice or numeric position only when the language or structure supplies a defensible order. A brand appearing in the first paragraph or first bullet is not automatically the preferred recommendation. Unordered sets should remain unordered.
Verify factual accuracy independently
A primary recommendation with an incorrect price, location, product capability, eligibility rule, or company identity is both visible and inaccurate. Preserve both findings. Collapsing them into one favorable score hides the item most likely to require urgent correction.
Audit citations independently
A link can support a definition, market fact, or competitor claim without supporting the brand’s suitability. Conversely, an answer may recommend a brand without citing its website. Record source presence and claim support separately from recommendation role.
Audit claim-to-source support →Keep sentiment and recommendation apart
Positive language does not always recommend, and a cautious recommendation is not necessarily negative. “A respected provider” can be a passing description; “the best fit if you require feature X, but verify the contract terms” remains a qualified recommendation. Code the decision role first, then any separate tone field.
05
Work through a transparent fictional example
This eight-answer illustration shows how one brand can look strong under a mention metric and much weaker under a recommendation metric. It contains no measured provider, customer, or market outcome.
Classify every eligible answer
Suppose eight completed eligible answers produce these roles for ExampleCo: one absent, two passing mentions, two considered options, two qualified recommendations, and one primary recommendation. The saved ledger also contains the exact question, answer, supporting span, reviewer, and any ambiguity note.
Calculate mention coverage
ExampleCo appears in seven of eight eligible answers, so mention coverage is 7/8, or 87.5%. That accurately describes broad presence in this fictional sample. It does not say that seven answers endorsed the company or that the company has 87.5% of an AI market.
Calculate recommendation coverage
If the declared recommendation metric includes qualified and primary roles, ExampleCo is recommended in three of eight eligible answers, or 37.5%. Considered options remain visible as 2/8 but do not enter the numerator. Report the counts beside the percentages so readers can see the role distribution.
Bound the conclusion
The defensible result is that ExampleCo appeared in seven and received an affirmative recommendation in three of these eight eligible cells. NIST distinguishes performance on a fixed benchmark from claims generalized to a wider population. This example therefore cannot establish buyer preference, market share, provider-wide treatment, or future behavior.
Read NIST AI 800-3 ↗06
Calibrate reviewers before automating the label
Recommendation language is often nuanced. A useful codebook includes examples, boundary cases, and an escalation path—not just state names.
Build a small adjudication set
Select representative clear and ambiguous answers from the actual workflow. Have at least two reviewers label them independently, compare disagreements, and revise definitions with positive, negative, and borderline examples. Freeze the codebook version used for each reporting period.
Use classification, not a vague quality prompt
OpenAI’s evaluation guidance recommends defining an objective and metric, using task-specific criteria, and favoring classification or scoring over open-ended judgment. Anthropic likewise recommends specific, measurable success criteria and well-defined qualitative scales. Those principles support a bounded role classifier; they do not validate this particular six-state taxonomy.
Review OpenAI evaluation guidance ↗Validate any automated grader
An automated rule or model grader can speed up clear cases, but test it against human-adjudicated labels before relying on it. Review errors by state, provider format, answer length, and question type. Route low-confidence or conflicting cases to human review and retain the original output and grader version.
Do not optimize the rubric to the result
Define whether considered options count before seeing brand performance. Apply the same rule to tracked brands and competitors. If the business later needs a different metric, calculate it as a separately named view or recode all comparable periods under the new version.
07
Turn the distinction into a useful report
The point of the taxonomy is not to manufacture a more impressive score. It is to show where the brand enters the decision and what evidence or risk deserves attention next.
Publish the role distribution
For each important question group, show eligible answers, absent, passing, considered, qualified, primary, and cautioned or rejected counts. Add provider and market views when they support a real decision. Do not hide considered or negative roles inside a blended “visibility” total.
Read the transitions behind a trend
A move from passing mention to considered option is different from considered option to primary recommendation. Compare the same eligible cells and inspect the full answers, reasons, facts, and sources before attributing the change to brand work.
Compare AI visibility over time →Connect each gap to the right action
Absence may trigger source and category discovery work; considered status may expose missing fit evidence; qualified recommendations may reveal a product or market constraint; cautions may require factual correction, product review, or no action when the caveat is accurate. Treat the observed role as a diagnostic starting point, not a causal explanation.
Start with two complete answers
Run one representative buyer question through the free AI Visibility Snapshot, then read the full ChatGPT and Gemini answers. Confirm the entity, classify the role, record any explicit order, verify material facts, and inspect visible sources. One snapshot cannot represent the market, but it is enough to test whether your team can distinguish mention from recommendation consistently.
Run a free AI visibility snapshot →Continue the research
Related AI visibility guides
Primary sources
Primary documentation used for this guide
AI products, search behavior, and platform policies change. Check these maintained first-party sources before making technical decisions.
- Evaluation best practices ↗
OpenAI API Documentation
First-party guidance on defining evaluation objectives and metrics, using task-specific tests, logging outputs, calibrating automated scoring with human feedback, and favoring classification or scoring against explicit criteria over open-ended judgment. The recommendation-role codebook is an AI Brand Lens application of those general principles.
- Define success criteria and build evaluations ↗
Anthropic Claude Platform Documentation
First-party guidance on specific, measurable, relevant criteria; multidimensional evaluation; task-specific test sets; qualitative scales; detailed rubrics; and validation before scaling automated grading. It does not define brand recommendation roles.
- Expanding the AI Evaluation Toolbox with Statistical Models ↗
National Institute of Standards and Technology
NIST AI 800-3 distinguishes performance on fixed benchmark items from broader generalized claims. This guide uses that boundary to keep mention and recommendation rates attached to the tested answer cells; it does not prescribe NIST’s statistical models.
Measure your visibility
Turn the questions in this guide into an evidence-backed baseline.
Get AI Visibility ScoreTalk to us
Have a visibility question?
Tell us what your team is trying to measure or improve.
Contact AI Brand Lens →Keep learning
Explore every guide.
Browse practical answers about AI visibility, competitors, and measurement.
View all guides →