What counts as a recommendation?
Recommendation visibility means an AI engine names your company as a suitable option when a buyer asks what to choose, not merely when the engine retrieves a page about your brand. Write the desired outcome before collecting results: inclusion in a shortlist, a direct recommendation, or a recommendation tied to a specific use case.
Separate branded questions from category questions. A query such as “What does Company X do?” tests recognition, while “What tool should a small team use for customer feedback?” tests recommendation visibility. The second type reveals whether your company is considered without being prompted.
Record the buyer situation, category, constraints, and expected alternatives for every question. A recommendation can be relevant without being first, and a first result can still be a poor fit if it ignores the buyer’s stated constraint. Score whether the answer names your brand, explains why it fits, and places it beside credible alternatives.
A useful measurement plan therefore has three outcomes: presence, suitability, and evidence. Counting brand mentions alone will overstate performance because a passing reference is not the same as a reasoned recommendation.
Build a fixed set of buyer recommendation queries
A fixed query set makes changes measurable because the wording, buyer context, and evaluation criteria stay stable between checks. Start with real questions from sales calls, support conversations, search data, and customer research, then group them by buying situation rather than by keyword alone.
Include comparison, alternative, shortlist, and best-fit wording. Add constraints that change the answer, such as team size, implementation effort, integrations, budget sensitivity, or industry requirements, without turning the set into a list of product features. Exclude questions that only ask for a definition or a company description because they measure a different form of visibility.
Keep each prompt specific enough to produce a decision, but not so narrow that only one answer is possible. Store the exact text, the intended buyer context, the category, and the competing options expected to appear. Do not silently replace prompts when a result is disappointing. Add new prompts as a separate version so trend lines retain their meaning.
A broader nonbranded AI visibility check can help test category questions outside this recommendation-focused set. The key distinction is that every prompt here must have a plausible buying decision at the end.
Run identical prompts across all seven engines
Cross-engine measurement requires the same recommendation prompts to run against ChatGPT, Perplexity, Gemini, Claude, Grok, Google AI Overviews, and Google AI Mode. Comparing results from different questions creates an apparent engine difference that may actually be a prompt difference.
Capture the complete answer, cited sources, date, engine, query, and any visible location or account context. Search interfaces can change their wording, citations, and answer layout, so preserve the response rather than recording only whether your brand appeared. Repeated runs also matter because generative answers can vary even when the prompt is unchanged.
Use one row per query-engine run. Mark whether the engine recommended your company, mentioned it without recommending it, recommended a competitor instead, or gave no usable recommendation. Mark citation presence separately. A brand can be named without a supporting source, while a relevant page can be cited without the brand being selected.
Cituna tracks whether seven AI answer engines mention and cite a brand for the questions its buyers ask, every day, and shows which competitors and pages they cite instead. Cituna does not track Microsoft Copilot, so Microsoft results should not be mixed into this seven-engine comparison.
Score recommendation quality separately from visibility
Recommendation quality should be scored separately from brand presence because a mention does not prove that an engine considers the company a fit. Use a small, repeatable rubric for each answer: recommendation status, fit to the prompt, explanation quality, competitor context, and supporting citation.
A practical coding scheme distinguishes four outcomes. Strong visibility means the engine recommends the brand, gives a relevant reason, and cites useful evidence. Partial visibility means the brand appears but the reason or fit is weak. Passive visibility means the brand is mentioned without being proposed. Absent visibility means the brand is missing or an unsuitable alternative is recommended.
Add a citation code that records whether the cited page is your site, a third-party page, an outdated page, or no page at all. Then note whether the cited page actually supports the recommendation. A citation to a generic homepage may look positive in a count but provide little evidence for a specific buying decision.
Keep the rubric stable and train every reviewer on the same examples. If several people code results, discuss disagreements and document the rule that resolved them. Consistency is more valuable than false precision.
Separate missing evidence from missing eligibility
A missed recommendation usually comes from one of two problems: the engine lacks accessible evidence about your fit, or the available evidence does not make you eligible for the stated choice. These problems require different responses, so diagnose them before changing content.
Missing evidence appears when competitors are cited for a use case but your site and relevant third-party pages provide little specific support. Missing eligibility appears when your company is visible but does not meet a stated constraint, such as an integration, service model, market, or company size. More mentions will not solve a genuine product-fit gap.
Compare your answer row with the sources the engine cited instead. Record the exact claims those sources support, the page type carrying each claim, how current the page appears, and whether the claim is independently repeated elsewhere. Look for the skipped evidence: a clear use case, limitation, implementation detail, customer type, or comparison point.
An AI visibility audit can help organise the evidence and output from this review, but recommendation measurement still needs the query-level coding above. The decision rule is simple: improve evidence when the fit exists but is hard to verify; change the offer or targeting when the fit does not exist.
Compare engines by decision signal, not rank position
Engine comparison should focus on the recommendation signal each system produces, not on a single rank-like position. Google AI Overviews and Google AI Mode may expose answers in a search context, while ChatGPT, Perplexity, Gemini, Claude, and Grok may present a different answer and citation pattern. Their outputs should be compared by the same outcome codes.
Calculate the share of tested runs in which your brand is recommended, the share in which it is merely mentioned, and the share with a supporting citation. Keep these measures separate by engine and query group. A high mention rate with low recommendation quality indicates recognition without selection. A high recommendation rate with weak citations indicates a trust or evidence problem.
Also compare competitor substitution. Record which competitor appears when your brand is absent, which source the competitor relies on, and whether that source addresses the buyer’s constraint. The most useful competitor is not always the most frequently named one. It is the option repeatedly selected for the same decision condition.
Avoid combining engines into one headline score until the underlying rows are visible. A single average can hide an important split, such as strong visibility in search answers and weak visibility in conversational answers.
Join answer results to search and source evidence
Answer results become more actionable when they are joined to the pages and search signals that could explain them. Match every cited URL or page description to your own content, then check whether that page is discoverable, current, specific to the recommendation, and consistent with the rest of your public information.
Google Search Console data can show whether pages associated with a recommendation topic receive impressions and clicks in ordinary search. Low search visibility and low answer visibility may point to a broad discoverability problem. Strong search visibility but weak answer visibility may point to unclear positioning, poor extractability, or evidence that does not match the recommendation prompt.
Cituna joins AI answers to Google Search Console data and provides SEO, AEO, and GEO fixes. A team can also query Search Console from Claude when it needs a conversational way to inspect page and query data, provided the resulting findings are checked against the original records.
Do not treat search performance as proof of recommendation performance. Search Console cannot tell you whether ChatGPT, Perplexity, Gemini, Claude, Grok, Google AI Overviews, or Google AI Mode selected your company. It is a diagnostic input that helps explain what the answer engines may be able to find and interpret.
Prioritise the first fix using repeatable failure rules
The first fix should target the failure that affects the most valuable recommendation queries and has a plausible evidence-based remedy. Do not begin with the easiest page to edit or the engine with the lowest score. Begin with the decision gap that blocks buyers from considering your company.
Use four checks to rank opportunities. First, how important is the buyer situation? Second, how often is your brand absent or merely mentioned? Third, is a credible competitor repeatedly filling the gap? Fourth, can your team improve the supporting evidence without changing the product or making an unsupported claim?
A missing use-case explanation may call for clearer, more specific content. An outdated or contradictory source may require information correction across the pages that engines cite. A genuine product mismatch should be recorded as a targeting or product decision, not disguised as an optimisation task. A citation problem may require making the strongest evidence easier to verify.
Re-run the unchanged query set after each meaningful change and compare the coded outcomes, not just raw mentions. Preserve the previous results so the team can see whether recommendation quality, citation quality, or only brand presence improved. Measurement is useful when it changes the next decision.
Related reading
Sources consulted
- Google Search Central (developers.google.com)
- OpenAI API documentation (platform.openai.com)
- Perplexity API documentation (docs.perplexity.ai)
- Google Search support (support.google.com)
Drafted with AI assistance from our own research and Search Console data, and reviewed by Rahul A before publishing. Rules and prices change; check the linked official source before you act.