# How to Measure Recommendation Strength in an AI Answer By AI Visibility editorial team AI-assisted article. An answer lists your company alongside four alternatives. Another says it would choose your company first for the customer’s stated needs. A third names it while explaining why it would look elsewhere. All three are mentions. They communicate different advice. A report that treats them as equally strong recommendations hides the difference a reader would notice immediately. Measure what the answer actually says about the company in context. Use a small set of categories with explicit rules and keep the wording that supports each choice. AI Visibility’s monthly reporting guide covers the broader visibility report. This walkthrough isolates the classification of an answer’s recommendation, including cases where the language is conditional, negative or mixed. AI Visibility monthly reporting guide (/article/what-should-a-monthly-ai-visibility-report-contain/) Separate being named from being recommended Start with an observed answer and a specific company. Verify that the wording refers to the right business before judging what it says. Then look for a statement that connects the company to the reader’s decision. A neutral description or an inclusion in a list can be useful visibility. A statement that would choose the company for a stated need is a more explicit recommendation. A citation can support a statement without recommending the company named on the cited page. Category What qualifies What to keep as evidence Listed or mentioned Identified without a clear choice in its favour The list entry or relevant passage Conditional recommendation Explicitly suggested if a stated condition applies The recommendation and the condition Explicit preference Clearly chosen first or preferred for the stated task The wording and scope of the comparison Discouraged Explicitly advised against for the task The limiting or negative statement Mixed or unclear Conflicting advice or no defensible single classification Both relevant passages and the reason for uncertainty Track absent and missing separately. Absent means the answer was observed and the target company did not appear under the coding rule. Missing means you could not inspect an eligible answer. An observed answer that discusses alternatives can establish absence. A failed capture cannot. If you include source-only references, record them in their own field so a citation does not silently become a recommendation. Keep a short rule for comparative language. The first entry in a list can result from formatting or an order the answer does not explain. Treat first position as an observation of order. Use the preference category when the answer explicitly chooses the business or identifies it as the preferred option for the task. If the wording applies only to a particular use, retain that condition in the record. Keep scope and supporting evidence beside the label Record the exact prompt, product surface, date and available context. A preference for a low-budget task can be different from a preference for an urgent task. A change in language, geography or follow-up question can also change which conditions the answer addresses. Google documents differences between its AI Mode and AI Overviews. Avoid pooling different surfaces into one result unless the protocol explains why and how they are combined. Google guidance on AI features in Search (https://developers.google.com/search/docs/appearance/ai-features) Read the sentence before and after the quoted recommendation. “I would choose this provider” can refer to a specific branch or a restricted situation in the preceding text. Copy the smallest passage that preserves the meaning and save the reason for the classification. If the answer recommends one brand for availability and another for price, code each brand relative to the decision in the prompt. Where neither receives an overall preference, leave that distinction visible. A dated Brewing Gadgets Instagram caption illustrates why qualifiers matter. The official post describes an offer for selected Compak grinders in the UAE. The word “selected” limits the scope. A reviewer reading an AI answer needs the same habit when the answer says a business is a good choice only under certain conditions. The social post is an example of a scoped publisher statement and does not count as an AI recommendation. Official Brewing Gadgets Instagram post (https://www.instagram.com/brewing_gadgets/p/DfWOw78j2CY/) Verify the recommendation’s underlying facts separately if the report depends on them. An answer may strongly prefer a company based on hours, pricing or service coverage that its own website does not confirm. A strong recommendation label describes the wording. A source check records whether the reason is supported. Keep both even when they point in different directions. The owner can then distinguish a visibility question from an accuracy problem. View the selected official Instagram post (https://www.instagram.com/brewing_gadgets/p/DfWOw78j2CY/) Official Brewing Gadgets promotion on selected Compak grinders. Make the classification reproducible Write a short codebook before reviewing the batch. Include a positive example, a near miss and a conflict for each category that needs judgment. Decide how to handle a recommendation that starts positive and ends by preferring another option. In a simple panel, a mixed or unclear category may be more useful than forcing it into a stronger or weaker class. Save the reason so a second reviewer can reproduce the decision. Ask a second reviewer to label an overlapping sample without seeing the first classification. Compare the rows where the reviewers differ and revise the rule if needed. Save the original and reconciled labels. NIST’s generative AI profile discusses structured human feedback and the limits of relying on quantitative measures without understanding their context. The categories here are a proposed editorial rubric for this task. They are not a NIST scoring standard. NIST generative AI profile (https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf) Avoid assigning numerical strength scores unless you have a defined purpose and a defensible interpretation. The difference between a mention and a conditional recommendation need not equal the difference between a conditional recommendation and an explicit preference. Negative and mixed cases also do not fit cleanly on a single line. Report category counts first. If you publish a composite measure, explain the weighting and show the uncombined observations beside it. Check whether the language of the rubric works in translation. A literal equivalent of “try” can express different force depending on the sentence and context. Let a bilingual reviewer read the complete passage and record why it qualifies. If English and Arabic questions serve different customer needs, report their results separately. A difference in prompt purpose can produce different categories even when the platform is unchanged. For answers that mention the same business several times, decide which passages are relevant to the requested task. Keep each passage in the evidence record and assign the final category at answer-and-company level. If a later sentence narrows a claim, retain the qualifier. If one section recommends the business and another advises against it for the same conditions, use the mixed label and record both passages. Counting each sentence as a fresh recommendation would give longer answers more weight. A recommendation can also be explicitly scoped to a ranked list. An answer might call one company the best choice among the options it considered, while making no statement about the broader market. Code the stated preference and preserve that boundary. The report should let a reader see whether the wording supports a preference within a shortlist, a conditional choice for one task or a broader statement. Work through a small example with explicit denominators A fictional panel used seven planned answer records for Example Workshop. Six were observed. One answer listed the workshop as another option, one explicitly chose it if the customer needed a Saturday appointment and one chose it first for the stated requirements. One advised against it, one did not name it and one mixed a possible mention with advice to start elsewhere. The seventh run was missing. Using the categories above, the local tally produced one listed, one conditional, one preferred, one discouraged, one absent and one mixed observed answer. The company appeared in five of six observed answers. The two explicit positive cases account for two of six observed answers, or two of five answers naming the company. Both versions of the rate can be useful when clearly labelled. The first asks how often the sampled answer gives an explicit positive recommendation. The second asks how often a named appearance becomes one. Capture coverage was six out of seven planned runs. The missing answer did not add a positive, negative or absent classification. The arithmetic check used the supplied sample categories and did not ask a model to classify the text. The rows are fictional and the percentages illustrate denominator choices. They are not findings about a real company or platform. The historical evidence has a different role. Lunasol’s published twelve-run Google AI Mode study counted named businesses and citations and explicitly left the definition of a recommendation unresolved. It reports that one run included a labelled line of praise for each listed business. That is a useful example of wording worth saving, but praise alone does not establish that each business was chosen above others. The historical runs were not recoded for this article. Lunasol’s published 12-run study (/article/does-google-ai-mode-cite-the-businesses-it-names-a-12-run-lunasol-study/) Report what changed without implying a commercial result Use a Recommendation wording record with the answer ID, company ID, prompt, surface, raw passage, category, qualifier, reviewer and any unresolved disagreement. Keep order and citation evidence in their own fields. Include a version for the rubric so the next report can distinguish new answer wording from a change in coding policy. Compare like with like across reporting periods. If the number of explicit preferences increased, check whether the panel included more prompts asking for a single choice. A broad prompt that asks for several options invites a different response from a prompt that asks which one to pick. Keep the question mix fixed where the purpose is comparison. If it changes, identify which part of the report can still be compared on the same basis. Present the counts next to the percentages. State whether the denominator includes all observed answers or only answers that name the company. For a small batch, show how one changed label would affect the rate. If several answers could plausibly be classified as mixed, state that before drawing a conclusion from a one-case difference. A clear rule and a small evidence table are usually more helpful than an extra decimal place. Visits and enquiries need their own records. An authorised analytics or CRM export can describe the outcomes it captures and the attribution gaps it leaves. It cannot show that a particular recommendation was seen unless there is evidence connecting the answer to that customer. This article used no private exports and ran no fresh assistant prompts. The evidence supports a coding method and a historical illustration, with commercial outcomes outside its scope. Include a row that challenges each category when handing the method to another team. For example, give them a first-listed brand without any endorsement and a second-listed brand explicitly preferred for the stated need. Ask for the evidence passage as well as the label. A reviewer who chooses by list position will produce a different result, and the saved wording makes the reason for the disagreement visible. AI Visibility is published by Lunasol, which offers GEO and AEO services. The published study is related-party research and the coding example is fictional. About AI Visibility (https://aivisibility.ae/about/)