What 624 AI answers taught us about brand-name matching

260 phrase matches. Zero brand mentions attributed.

A case-insensitive search for the phrase “share of voice,” allowing whitespace between the words, matched 260 of 624 collected answers: 41.7%. The phrase appeared 385 times in total. Repetition within an answer does not create an extra answer-level match.

The saved analysis attributed no brand mentions to Share of Voice in the 623 answers it analyzed. The remaining answer, q4-1-ChatGPT-1, has no saved classification. It is unknown, not a brand absence. All 260 phrase-matching answers were in the analyzed set.

A metric can look like a brand name

In answer q1-0-ChatGPT-2, the phrase appears in this passage: “Maximum control, customizable metrics (share of voice in answers, accuracy, sentiment, source overlap).” In q1-0-Claude-0, a comparison table contains “Share of voice vs. rivals.” These inspected examples use the words as a metric, not as an identifiable reference to our product.

An identity check should consider the surrounding sentence, product description, domain or other distinguishing evidence. A short phrase match cannot settle the question. Conversely, an automated classifier’s rejection is not infallible; a disputed answer needs review against its original text.

No archived answer text contained the literal domain shareofvoice.app. That is a text-only observation. We do not claim that no attached source cited the site: some provider citation URLs are opaque redirects, and attached citation metadata is not part of this public extract.

Exactly what we collected

On 10 October 2026, from 11:51:46 to 12:41:39 UTC, we collected 624 successful, single-turn API answers. The scope was Global, the category was “ai answer visibility monitoring,” and the audience was marketing and PR teams. All prompts requested English. This was not a Switzerland-specific report.

The panel contained 26 question groups, two phrasings each and three repeats per provider: 52 prompts × 3 repeats × 4 providers = 624 answers. It used the older buyer-panel-v4 protocol: five overlapping buyer needs crossed with five generic comparison dimensions, plus a category discovery group. Those needs were not validated as representative of buyer demand.

Collection requested at most 300 words, with a 4,000-token output ceiling. OpenAI, Gemini and Claude used web-search tools; Perplexity used provider-native search. Requests allowed at most one search query if needed; Claude also had a one-use tool limit. OpenAI used low reasoning effort and low search context. These are requested settings, not proof that every response respected the word or search limits. Full settings are in the download.

The provider totals describe this archived collection only. They are not a quality league table. Answer length, search behavior and wording can affect phrase frequency. API responses are not personalized ChatGPT, Gemini, Claude or Perplexity consumer sessions.

What you can safely conclude

• Literal matching can confuse a common marketing expression with a brand name. Inspect identity before treating a match as brand visibility.
• Report collection coverage and analysis coverage separately. Here they are 624/624 and 623/624; an unclassified answer must remain visible as an exclusion.
• Zero attributed mentions describes this analysis of this prompt panel. It does not prove the product is never recommended, nor does it measure Swiss customers or audience market share.
• This is one collection of our own brand, with legacy prompts and automated labels. It does not establish accuracy across other brands or show that any SEO change caused an improvement.

Download the evidence and reproduce the count

The JSON contains all 624 original English answer texts, the prompts, timestamps, requested models and saved brand-attribution flags. It excludes billing, access links, usage records, search logs, attached citation metadata and the action plan. No other customer’s report is included. Historical vendor claims, prices and recommendations inside model answers are unverified output, not our advice.

• Archived prompts and answers · JSON · 2.2 MB
• One row per answer · CSV
• Offline recount · Python
• Method, fields and limitations · TXT

Save answers.json and recount.py in the same folder, then run: python3 recount.py answers.json. The script uses only Python’s standard library and makes no network or AI calls. Its regex is \bshare\s+of\s+voice\b with case-insensitive matching. It reports answer matches, total occurrences, provider breakdowns and the saved classification denominator. The checksum in README.txt identifies the exact archive used here.

Use this check before showing a client a score

Ask the supplier to show the answers behind a brand total, its alias rules and any unclassified responses. For ambiguous names, read both accepted and rejected examples. If identity is unresolved, label it as such; don’t turn uncertainty into a zero or count every matching noun as a brand mention.

Then choose the right denominator. Our calculator guide separates mention rate from share of brand appearances. The score reliability guide explains comparable runs, while mentions, citations and recommendations separates the signals. Our methodology sets out the wider report approach.