How to Benchmark Competitor Visibility and Share of Voice in LLMs

A growth lead in India asks ChatGPT and Gemini for vendors in a strategic B2B category. Both name the same competitor first, while your brand is missing or described using an outdated product claim. Leadership wants to know whether this is an isolated answer, a wider visibility problem, or evidence that the competitor has become the default recommendation. A conventional SEO dashboard cannot resolve that question.

Traditional SEO measures the visibility of URLs against queries, locations, devices, and search features. LLM visibility measures how a brand is represented inside a generated answer: whether it appears, where it appears, how strongly it is recommended, which sources support the answer, and whether the underlying claims are correct. The same prompt can also produce different results across interfaces, model variants, browsing settings, accounts, and dates.

Treat the benchmark as an outside-in monitoring system rather than a rank tracker. Its job is to answer three operational questions: where your brand wins or disappears against competitors, whether the resulting answers can be trusted, and which content, data, documentation, PR, or product tasks should be opened in response. This requires a stable prompt library and grading rubric, not a collection of screenshots from unplanned searches.^[6]

Start with the business decisions the benchmark must support. Record the products or categories in scope, buyer segments, funnel stages, Indian regions, languages, and review period. Keep raw results segmented before creating an overall score. Otherwise, a strong English-language result for a broad category can conceal weak visibility for a regional query or a high-intent technical evaluation.

Include interfaces that buyers and evaluators are likely to use, such as ChatGPT, Gemini, Claude, Perplexity, and Google surfaces that present AI-generated answers. Treat a consumer interface, a browsing-enabled mode, an API model, and an AI search result as separate test environments even when they share a model family. Features and availability change, so confirm the active interfaces at the start of every audit cycle and preserve the exact model or surface label captured during testing.

Build a controlled comparison set. Include your brand, direct Indian competitors, relevant global vendors, and alternatives that regularly appear in evaluation conversations, such as an in-house approach or an adjacent product category. Freeze this set for the reporting period. Add or remove a competitor through a documented review rather than changing the list whenever an unexpected name appears; emerging brands can be tracked separately until the next baseline reset.

India-specific scope should reflect actual buying conditions rather than adding a country name to generic prompts. Account for local vendor availability, pricing conventions such as rupees where relevant, procurement expectations, regional terminology, and questions involving Indian regulation or data handling. If Hindi or another Indian language matters commercially, test prompts written naturally in that language and use qualified reviewers; literal translations often change intent and specificity.

Organise prompts around work a buyer is trying to complete. Discovery prompts ask which approaches or vendors fit a problem. Comparison prompts test differences, trade-offs, and shortlist logic. Risk prompts cover security, data residency, integration, compliance, and commercial constraints. Implementation prompts ask about migration, architecture, onboarding, support, and operating effort. Post-purchase prompts can expose incorrect support guidance that affects service quality as well as acquisition.

For a B2B SaaS audit, a useful sequence might move from “Which platforms can solve this problem for an enterprise in India?” to a named comparison, then to an implementation question about existing systems, and finally to a risk question involving Indian data requirements. Preserve neutral wording in the core set. Prompts that assume your brand is the best option measure confirmation behaviour, not competitive visibility.

Create controlled variants for persona, company size, industry, region, language, and technical depth, but change one dimension at a time. Tag every prompt with a permanent identifier, journey stage, intent, market, language, persona, and business priority. Keep a fixed core library for trend reporting and a smaller rotating library for new products, competitor moves, seasonal concerns, and questions collected from sales or support.

Coverage must match grading capacity. A smaller prompt set reviewed carefully is more useful than a large library that receives inconsistent scoring. Before launch, have SEO, product marketing, sales engineering, support, and regional reviewers inspect the set for missing buyer questions, loaded wording, and duplicated intent. Lock the approved version for the run; proposed wording changes enter the next version rather than altering a live baseline.

Swap the bracketed parts for your own category, competitors and market. Keep the wording fixed once you start, so month-on-month changes reflect the models, not your prompts.

Copy these columns into a Google Sheet or Excel file. One row is one answer from one engine for one prompt on one date. Keeping that grain makes it easy to pivot by engine, journey stage or competitor later.

Score each eligible response on two separate axes. Visibility covers mention share, prominence, and recommendation share. Answer quality covers factual accuracy, hallucination severity, tone, citations, and provenance. Keeping these axes separate prevents a prominently recommended brand with incorrect claims from receiving an inflated performance score.^[2]

Use mention share, recommendation share, prominence, and recommendation strength to capture how each model represents your brand versus competitors. Report these metrics by journey stage, prompt, model, market, and language before applying any business-priority weighting.

Grade factual accuracy against an approved evidence set containing current product documentation, policies, availability, integrations, and other verified claims. Classify errors by operational consequence: a minor wording issue, a material product or commercial error, or a critical claim requiring immediate review. Score citation quality separately by checking whether a source is accessible, relevant, sufficiently authoritative for the claim, and actually supports the statement. A citation is a traceability signal, not proof that an answer is correct.^[5]

LLM-based graders can classify obvious mentions, extract cited domains, and flag answers for review, but they should use the same written rubric as human reviewers. Calibrate them against a human-scored sample and route low-confidence, disputed, regulated, or high-severity cases to people with subject expertise. Periodic double-grading helps reveal drift between reviewers. Track abstention and uncertainty as valid behaviours rather than rewarding confident guessing.^[3]

Share of voice shows what slice of the AI conversation in your category you own. Calculate it per engine and per journey stage, then as a blended figure across the whole prompt set.

Worked example: you track 40 prompts on ChatGPT, Gemini and Perplexity, which gives 120 answers. Your brand appears in 30 of them, so your mention rate is 25%. Four competitors are mentioned 70 times in total, so your mention share of voice is 30 ÷ (30 + 70) = 30%. Run the same prompts next month and compare like for like.

A spreadsheet works for a first audit. Once you track more than a few dozen prompts or several engines, a tool keeps runs consistent. Engines and starting prices below are as reported in October 2026 and change often, so check each vendor before buying.^[8][9]

Lumenario also reports sentiment and share of voice against the competitors you choose. See plans and pricing.

Spreadsheets become fragile when prompt versions, model outputs, evidence, reviewer decisions, and remediation history must remain aligned across several functions. Lumenario is relevant where an organisation wants a governed AI discovery and measurement layer that connects visibility monitoring with structured brand knowledge, while preserving human ownership of prompt design and critical grading.^[1]