AI search recommendation visibility is not the same as being indexed, retrieved, cited or named in an answer. A brand can appear in a source list without being proposed to the user. It can be mentioned as a market participant while competitors receive the actual shortlist. It can also be recommended, but only for a narrow use case that the measurement programme failed to record.
That distinction matters for SEO and content teams. A dashboard that counts every mention as a win will overstate performance and obscure the work required to earn a credible recommendation. The practical question is simpler: when a person asks an AI system which product, provider or brand to choose, does the answer actively put your offer forward, for whom, and on what evidence?
This framework treats recommendation visibility as an observed answer outcome, not a proxy metric. It combines structured prompt sampling, explicit coding rules, evidence review, competitor analysis and conversion-oriented follow-up. It is designed for category, comparison and shortlist queries where commercial consideration is genuinely in play.
Start by separating presence from recommendation
Use a measurement taxonomy before calculating a rate. The categories below should be mutually exclusive at the primary-answer level, although one response may contain several secondary signals.
| Outcome | What it means | What it does not prove |
|---|---|---|
| Retrieval or source presence | Your domain, page or source appears among material available to the answer. | That the brand is shown or endorsed. |
| Mention | The answer names the brand or product. | That it fits the user’s need. |
| Citation | The answer links or attributes a claim to your content. | That the answer recommends your offer. |
| Qualified inclusion | The brand appears in a shortlist with a stated fit, limitation or use case. | That it is the preferred choice. |
| Active recommendation | The answer tells the user to consider or choose the brand, product or service for a defined need. | That the recommendation is accurate, durable or commercially valuable. |
This is compatible with the broader distinctions in AI Search Measurement Taxonomy. In professional reporting, I would never collapse citations and recommendations into one “visibility” number. They arise from different conditions and demand different remedies.
A citation can support a useful factual claim while a competitor is recommended because it better fits the stated budget, market or implementation constraint. Conversely, a recommendation without a visible citation may be influential but difficult to diagnose. Track both; do not mistake one for the other.
Define what counts as an active recommendation
Create an operational definition before collecting outputs. An active recommendation should meet all three tests:
- Selection language: the answer uses a directive or evaluative phrase, such as “choose,” “best for,” “a strong option,” or “I would shortlist.”
- Named entity: the recommendation identifies the brand, product, service or a clearly attributable offering.
- Fit statement: it connects that offering to a user need, constraint or scenario.
“Brand X is a project-management platform” is a mention. “Brand X is worth shortlisting for small teams that need simple client approvals” is a qualified recommendation. “Brand X is the best choice” is a stronger recommendation, but only if the answer explains the conditions under which that judgment applies.
Record negative and conditional language too. “Avoid Brand X if you need enterprise SSO” is not a recommendation, yet it is high-value intelligence. It reveals an exclusion criterion that may be accurate, outdated or unsupported.
Use a coder guide, not individual judgment
Give reviewers a one-page rubric with positive and negative examples. Require them to capture the exact recommendation phrase, the entity named, its use case, qualifying language and any cited source. Blindly double-code a small recurring sample. Resolve disagreements by changing the rubric rather than quietly changing the result.
This discipline prevents a familiar failure mode: a team sees its name in an answer, assumes success and spends months improving pages that were already being cited but not chosen.
Build a prompt set that represents decisions, not just keywords
Recommendation visibility should be measured on prompts that ask for a selection. Category prompts establish eligibility; comparison and shortlist prompts test preference. Build strata that reflect the commercial paths your audience actually takes:
- Category: “What are the best payroll platforms for a 50-person company?”
- Comparison: “Is Product A or Product B better for agencies with multiple clients?”
- Shortlist: “Give me three CRM options for a UK consultancy with a limited admin team.”
- Constraint-led: prompts involving price band, location, integration, accessibility, industry or scale.
- Follow-up: “Which of those has the fastest onboarding?” or “Why did you not include Brand X?”
Use a documented sample frame: source of query ideas, target market, language, device or interface where relevant, and the date each prompt entered the set. The method for building a representative prompt set deserves the same care as keyword research; see How to Build a Representative Prompt Set for AI Search Visibility Tracking.
Do not begin with branded prompts. They are useful for brand accuracy checks, but they answer a different question. Unbranded or lightly constrained prompts are the better test of whether the system independently introduces your brand into consideration.
Capture answers as repeatable observations
AI answers can vary by session, location, product interface and time. Treat each result as a dated observation, not a permanent ranking. For every capture, retain the full prompt, engine and interface, model or version label if shown, locale, language, signed-in state, date and time, run number, full answer, citations, links and screenshots or exports where permitted.
For a client programme, a defensible log might read: “Engine: [named product]; interface: web; version: label displayed in interface; locale: en-GB, London; signed out; 10–14 February 2025; 40 prompts; three independent runs per prompt; manual capture.” That makes a finding auditable. Without those fields, statements such as “the engine recommends us” are not reproducible evidence.
Google’s documentation on AI features and your website is the appropriate primary reference for Google-specific site participation guidance. It should not be read as a promise that any page, brand or source will be shown in an AI response. Similarly, Perplexity’s bot documentation is useful for understanding its published crawler controls, not for inferring recommendation placement.
Calculate the core recommendation metrics
Report metrics by query stratum and then as a weighted total. An unweighted aggregate can be misleading if low-value, broad category prompts outnumber high-intent comparison prompts.
| Metric | Calculation | Use |
|---|---|---|
| Recommendation rate | Prompts with an active recommendation ÷ eligible prompts | Primary measure of being put forward. |
| Qualified inclusion rate | Prompts where the brand appears with a fit statement ÷ eligible prompts | Shows shortlist access before preference. |
| Mean recommendation position | Average ordinal placement among recommended options | Indicates prominence; retain the distribution too. |
| First-choice rate | Prompts naming the brand first or as explicit best fit ÷ eligible prompts | Separates participation from leadership. |
| Evidence-supported recommendation rate | Recommendations whose material claims pass review ÷ recommendations | Protects against celebrating weak assertions. |
| Competitor co-inclusion rate | Brand recommendations that include named competitors ÷ brand recommendations | Shows the competitive set the answer constructs. |
Position requires a clear convention. Count the order in a ranked list. For prose, code explicit preference first, then order of introduction only when options are presented as comparable. Do not manufacture precision from an unordered list. A four-brand shortlist where every option is framed as suitable is not a ranking.
Report numerator and denominator beside every percentage. “Recommended on 6 of 20 eligible UK agency prompts” is more useful than “30% visibility.” With small samples, variation matters; use repeated runs and intervals rather than declaring a one- or two-prompt movement a trend.
Score qualification and evidence, not just recommendation language
A recommendation has quality dimensions. Code each one on a simple 0–2 scale: fit (none, generic, specific), evidence (none, partial, adequate), currency (unclear, dated, current for the claim), and constraint handling (ignored, partly addressed, addressed). Keep the raw notes. A composite score can guide prioritisation, but it must not hide why a recommendation is weak.
Review the claims that justify selection: pricing, availability, integrations, performance, suitability, certifications and geographic coverage. Ask whether the cited or discoverable source entails the claim, whether the source is current and whether the answer omitted a material limitation. For a detailed method, use the citation accuracy evidence-entailment framework.
Structured data can make page entities and attributes clearer to machines, but it is not a recommendation-control mechanism. Google’s structured data introduction describes structured data as a way to provide information about a page and enable eligibility for certain search features; it does not promise display or recommendation. Treat it as implementation hygiene where the markup accurately reflects visible content.
Turn competitor inclusion into actionable diagnosis
Build a competitor matrix: prompts as rows, brands as columns, with cells for mention, qualified inclusion, recommendation, first choice and cited supporting domains. Then annotate the stated reasons for selection. The useful question is not merely who appears more often; it is which qualification criteria repeatedly favour them.
Common patterns are revealing. A competitor may win budget-constrained prompts because its pricing is easier to verify. Another may dominate regional prompts because third-party directories substantiate local coverage. Your brand may be cited for technical detail but excluded from shortlists because onboarding, support or fit for a segment is not clearly evidenced.
Prioritise the gap by commercial relevance, frequency, evidence deficiency and feasibility. Improve the underlying product information before publishing comparison pages that make claims you cannot support. Where authoritative sources disagree, preserve the disagreement and state the conditions rather than trying to erase it.
Connect recommendation visibility to business impact carefully
Recommendation visibility is an intermediate outcome. It can influence consideration even when the interface sends no referral click, but influence is not directly proven by the observation alone. Combine answer records with branded-search movement, direct traffic, assisted conversions, sales-call self-reports and CRM fields where your measurement design can reasonably connect periods, markets and campaigns.
Use holdouts or controlled changes when practical, and label conclusions proportionately: association, plausible contribution or measured incremental effect. The last label requires a credible counterfactual. For measurement approaches when citations do not produce click data, see How to Measure AI Search Assisted Conversions When Citations Send No Clicks.
In my judgment, the most useful executive report has two layers: a stable scorecard of recommendation and evidence metrics, followed by a small set of answer excerpts showing why the brand won, lost or was excluded. The scorecard establishes scale. The excerpts preserve the decision context.
FAQ and conclusion
Is a citation an AI recommendation?
No. A citation attributes or links information. Code it separately unless the answer also names the brand and expresses a fit-based choice.
How often should teams measure?
Start monthly for stable categories and more often during launches, pricing changes or competitive disruption. Re-run a fixed control set alongside new prompts so shifts are interpretable.
What is a good recommendation rate?
There is no universal benchmark. Compare against your prior, documented baseline within the same prompt strata, locale, interface and capture method. A rate without its denominator and sampling context is weak evidence.
Should we optimise pages only for the brand to be named?
No. Build accurate, current evidence for the criteria that drive selection: fit, limitations, price, availability and implementation. Naming without justification is fragile visibility.
Conclusion: Measure AI search recommendation visibility as a chain: eligible prompt, active recommendation, position, qualification, evidence quality, competitor context and downstream outcome. Keep mentions and citations visible in the dataset, but do not let them inflate the core recommendation metric. Repeatable captures, transparent coding and claim-level evidence review turn a vague visibility claim into a programme a content, SEO and commercial team can act on.
