How to Measure a Portfolio of Brands, Products and Sub-Brands in AI

A group can have a highly visible corporate brand while several product lines rarely appear in AI answers. It can also seem to improve simply because it acquired a well-known brand. Without attribution rules, the aggregate mixes commercial outcomes, scope changes and answers counted more than once.

To measure a portfolio of brands, products and sub-brands in AI, preserve observations for each entity first, then calculate portfolio presence using unique answers. Keep explicit parent-company mentions, presence derived from owned brands and the distribution of product outcomes separate.

This guide is for corporate analytics and product marketing teams that already have an entity register and a sample of answers. The output is a reconcilable ledger: it explains what appeared, which entity received attribution and what changed during consolidation. The method and every number below are hypothetical examples, not customer data or an official assistant metric.

Decide what it means for the group to be visible

Start with the management question. “Do people see the parent company's name?” is not the same as “Are any of our offerings among the recommended alternatives?”. An acquired brand may answer the second question without ever naming its owner.

Use separately named indicators:

Derived presence is a classification in your system, not a statement the assistant made. When an answer names a subsidiary brand, your rule may include it in a consolidated measure, but it must not turn into an explicit holding-company mention.

Use the identity map as an input, not the output

Start from the brand entity audit: IDs, aliases and relationships must be approved before attributing answers. This process does not decide again whether a name is a brand, product or legal entity; it decides how that entity counts in a specific measurement.

Add an effective period, applicable market and analytical treatment to each relationship. Ownership may support consolidation; a distribution license does not automatically do so. A shared brand or joint venture needs an explicit rule, not two parent totals added by default.

Your reporting tree may simplify a more complex graph. Preserve the original relationship and assign one consolidation path in each additive view. If an exploratory view allows multiple memberships, disclose that its branches cannot be summed.

Do not confuse an umbrella brand with a variant family either. Google's ProductGroup documentation groups variants of one product and distinguishes common and variable properties. That model does not turn every product a company owns into a variant or establish a visibility-measurement rule. Applying this protocol does not require changing schema.

Build the ledger and eligible denominators

Store an answer once. Its identifier must distinguish the prompt, run and repetition under a recorded configuration. The appearances table references that ID and may contain several entities; it must not copy the answer as if each entity created an independent observation.

A minimum record retains:

• answer_id, prompt_id, run_id, date and surface;
• market, language, mode and prompt-bank version;
• entity_id and resolved level: brand, family or product;
• the passage supporting entity identification;
• state: mention, recommendation, rejection or ambiguous;
• extraction-rule and portfolio-scope versions;
• entities eligible for that question before inspecting the result.

Deduplicate repeated names within an answer for a binary measure. “Brand A” written five times produces one presence, not five. A genuine repeated query is another observation; importing the same capture again is not.

Each entity's denominator contains valid answers to questions where that entity could compete under the prior rules. Do not penalize a software line for missing a food-related question elsewhere in the group. A valid answer without mentions is an absence; a capture failure remains a failure and is reported separately.

An unresolved alias is not assigned to the most popular product. For confirmed presence, it does not enter the numerator and stays in the denominator when the answer is valid and eligible; also disclose ambiguous cases. Detailed attribute and availability measurement belongs in SKU-level recommendation tracking.

Reconcile answers before consolidating

Consider a hypothetical sample of 100 valid answers. Brands A and B are eligible alternatives in every answer. This example's consolidated rule includes explicit mentions of the parent and either brand.

A appears in 35 answers and B in 25. Adding them gives 60 entity-answer appearances, but only 45 distinct answers containing either brand: 35 + 25 - 15 = 45.

Explicit parent presence is 5/100, or 5%. Presence derived from the two brands is 45/100, or 45%. Presence across the full scope, which also admits the parent in this example, is 50/100, or 50%. These answer three different questions; none can be obtained by adding percentages without checking intersections.

In production, calculate a union of answer_id values, not a two-brand formula applied repeatedly. For each valid answer in the consolidated universe, record whether it contains at least one eligible entity in scope. Divide unique positive answers by unique eligible answers in that same universe.

If the scope admits only recommended offerings, a neutral parent mention does not qualify. Decide the rule before inspecting the result. Retain brand intersections too: they explain why the corporate total is smaller than the sum of the brand rows.

Separate unions, entity averages and business weights

A union asks “Did any entity appear?”. An average asks “How are the measured entities performing?”. They are not interchangeable.

This second example uses three products at the same analytical level, with comparable evaluation tasks and independent samples. It does not share the 100 answers from the previous example.

The observation-pooled average is 84/130 = 64.6%. The equal-product average is (80% + 20% + 0%)/3 = 33.3%. The first gives more influence to the product with more observations; the second treats products equally even though their estimates have different precision. Neither describes market share on its own.

If approved business weights are 60%, 25% and 15%, the weighted indicator is 0.60 × 80% + 0.25 × 20% + 0.15 × 0% = 53%. Do not select whichever view looks best: explain the question each answers and keep P3 visible with its small sample.

To define the value of questions, use the prompt-bank weighting method. Do not add another entity weight without documenting both layers: you could duplicate business influence.

An entity without valid answers has an unavailable rate, not zero. Keep its row and disclose portfolio coverage. If an entity is excluded from an average, state how many entities and how much weight were omitted; do not silently redistribute its weight. A total of entity-answer opportunities is also not a count of unique answers when multiple products share the same sample.

Preserve two views when the portfolio changes

An acquisition, retirement or reorganization can change the aggregate without changing a single answer. Preserve ownership and relationships at the observation date, alongside the scope used by the current report.

You need two dated views:

1. Scope at each date: attributes observations using the structure effective then. It describes what the group included at that time.
2. Constant scope: recalculates both periods using a fixed entity selection and comparable rules. It isolates the arithmetic effect of entities entering or leaving the portfolio.

Constant scope cannot always be reconstructed. If a newly acquired brand was not measured before, its past is unknown; do not fill it with zeros. Present the common panel and additions separately.

Accompany changes with a reconciliation bridge: observed movement in the common panel, additions, removals, reclassifications and coverage changes. Calculation order can affect assigned contributions when intersections exist; document the order and reconcile using answer_id values. This bridge decomposes the indicator; it is not causal proof that marketing produced the improvement.

Do not mistake co-occurrence for cannibalization

Two group brands appearing together does not establish that one is taking demand from the other. They may meet different needs. Product substitution across two isolated captures does not establish lost sales either.

Describe the observed pattern first: A only, B only, both or neither, under the same task and configuration. Then assess whether the recommended offering fits the predefined segment. The outcome can be appropriate for the group and problematic for a strategic brand.

A small brand that stays at zero needs its own review even when the consolidated measure rises. A brand with insufficient data needs better observation before an absence diagnosis. Keep brand and product owners involved; the corporate metric does not replace their decisions.

Test the consolidation before using it with management

Check six invariants using small cases you can count manually:

• Importing a capture twice does not change the answer count.
• Repeating an entity's name does not increase its binary indicator.
• An answer naming parent, brand and product counts once in the group union.
• A parent mention does not propagate to every child.
• No numerator exceeds its eligible denominator.
• An ownership reclassification changes only views using that version.

In a nested hierarchy, parent and child rates may have different denominators: do not compare their percentages as if they shared a population. Test set inclusion within the same answer universe.

The final report needs one row per entity with numerator, denominator, rate, ambiguous cases, coverage, version and owner. Add corporate unions and averages as separate views, not another row to sum with the others. Retain an appendix that traces each percentage to the original answer.

FAQ

Does a company mention count for all its products?

No. Record the entity identified by the answer. A parent mention does not establish child presence; only an explicit rule permits upward consolidation, and the result must be labeled as derived presence.

How do I avoid counting an answer twice when it names several brands?

Keep a unique answer_id and calculate the union of positive answers within the same eligible universe. Deduplicate by answer when consolidating; do not add rates or appearances for brands that may co-occur.

Which average should I use to compare products?

Show each product's rate and sample, and distinguish observation-pooled, equal-entity and business-weighted averages. Each answers a different question, and none replaces distribution or coverage.

What should I do with a product that has no valid data?

Keep its row as unavailable and disclose missing coverage. Do not turn it into zero or redistribute its weight without notice. An observed absence in a valid answer differs from being unable to measure.

Can I include an acquired brand throughout the historical series?

Only in a clearly labeled constant-scope view when comparable historical observations exist. Preserve attribution effective at each date too; if earlier data is missing, that brand's past remains unknown.

Does two brands appearing together mean cannibalization?

No. Co-occurrence establishes a joint appearance in the sample, not a transfer of demand or sales. Assess conflict using equivalent tasks, segment fit and additional commercial evidence.

Start with a small group you can reconcile

Select a parent, two brands and a few offerings. Count a sample manually, fix the consolidation rule and verify that the system reproduces the union without inventing parent or child mentions. Expand scope when another person can reconstruct every number.

Review how your brand appears in AI answers with Mentio. The portfolio protocol described here is an analytical method to implement and validate with your data; it does not presume an automatic hierarchical consolidation feature in the platform.

Want to know if AI mentions your brand?

Discover your visibility in ChatGPT, Claude and Gemini in minutes.

Analyze your brand for free

Related articles