How to measure whether AI recommends your company.

Start with the decision, not a visibility score

Write down what the baseline will help your team decide. A useful question is “Which comparison questions deserve an evidence review?” A less useful goal is “Make every AI score go up.” The first leads to inspectable answers and a manageable work queue; the second encourages an unexplained aggregate.

Define the company, relevant products, market, language and competitor set. Identify aliases carefully: a common product name may match unrelated text. Keep those matching rules alongside the classifications so that a changed rule does not look like a changed recommendation.

Build and freeze a buyer-question set

Start with roughly 20–25 questions, reviewed by someone who understands the product and its buyers. Treat this as a practical starting sample, not a statistically representative survey of demand. Mix discovery, comparison and decision questions and include questions that do not mention your brand.

Use your own real category and constraints. Do not manufacture favorable questions that only your product could satisfy and present that as broad market coverage. Assign IDs and versions, then record any later changes.

Collect comparable answers

Choose a named API/search surface or a documented consumer-surface collection method. Record the model, region, language, time and settings that you can actually control. Do not label a region assumption as a verified user location. Do not blend consumer results and API results without reporting the difference.

Target three to five runs per question per engine per cycle where the budget permits. Save each raw answer, its citations, tool calls and classification separately. Record failures and retries rather than quietly dropping inconvenient results. A transport failure is not evidence that the brand was absent.

A completed answer that does not recommend any option can still be informative. Keep it in the defined observed-answer set. An unusable or incomplete response should have an explicit exclusion reason and remain visible in a collection-quality summary.

Classify meaning, then calculate

Use absent, mentioned, shortlisted and recommended as separate states. “Acme exists” merely mentions it. “Acme and Northstar both meet these requirements” shortlists both. “For this situation I recommend Acme” explicitly endorses it. First position in an otherwise neutral list is not enough by itself.

Under the Mentionloom definitions, Recommendation Share counts shortlisted or recommended answers divided by the report’s total sampled answers. Lost Recommendation Share counts answers where at least one competitor qualifies while your brand is absent or only mentioned. Each answer contributes at most once to the lost count.

For a fictional 20-answer teaching example, six shortlisted plus four recommended answers produce 50% Recommendation Share. Seven qualifying competitor-win answers produce 35% Lost Recommendation Share. These are arithmetic examples, not Mentionloom results or benchmarks.

Interpret a baseline without overclaiming

Show counts alongside percentages. A two-answer difference in a tiny sample can produce a large percentage-point movement. Disclose the interval calculation and whether it accounts for repeated answers clustered within questions. Small samples do not justify universal conclusions about an engine or an entire market.

Review results by intent and surface before reading the overall number. A changed mix of easier questions can raise an aggregate even when no comparable question improves. Keep collection quality and the exact denominator visible.

Do not add referral sessions or crawler fetches to the numerator. They are useful but different observations. Nor does this process reveal how often real people asked those questions in private conversations.

Turn one result into a useful next step

Open one losing question, read the answer and inspect the sources actually cited. Record the gap as a hypothesis, not a causal explanation. A missing integration detail may justify documentation work; an unsupported requirement may instead indicate that your product is not a fit.

Before shipping, define the change, affected pages, comparison window and untreated question set. Re-run the same protocol afterward, document other changes, and avoid declaring success from a single favorable response. Continue with the competitor-evidence investigation guide.

A useful baseline handoff

The handoff should include the approved questions and versions, the collection protocol, raw-answer evidence, reviewed classifications, both recommendation metrics, failed-response counts, limitations, and one prioritized investigation. That makes the work repeatable by another teammate rather than dependent on a screenshot.

Mentionloom’s intended workflow connects these records to improvement and re-measurement. Public previews are illustrative; this guide can also be used to structure a manual evaluation.