How Mention Impossible measures AI recommendations

Method version 1.0 · applies to measurements from 1 November 2026 · last reviewed 2 November 2026

Every week, Mention Impossible asks four AI assistants (ChatGPT, Gemini, Claude and Perplexity) the same buyer questions, then counts which brands they name and which websites they cite as sources. This page explains exactly how, so anyone can check or quote our numbers.

Summary

• 48 fixed English buyer questions, 8 in each of 6 software categories.
• 4 AI assistants, queried through their official APIs with web search switched on.
• 240 answers per week: ChatGPT 24, Gemini 48, Claude 24, Perplexity 144 (48 questions × 3 runs).
• Two separate measures: mentions (brands named in the answer) and citations (web pages linked as sources).
• Every published number shows its sample size (n).

The questions

• 48 questions in English, 8 in each category: CRM, email marketing, project management, ecommerce platforms, website builders and AI writing tools.
• They are phrased the way buyers ask. Examples: "What is the best CRM for a B2B SaaS startup with a 10-person sales team?" and "Which website builder is best for a portfolio site?"
• Some questions name brands on purpose, because buyers ask that way too (for example "Jira vs Linear vs ClickUp: which should a 20-person engineering team use?" or "What are the best Mailchimp alternatives in 2026?"). This raises the named brands' mention counts for that question, and we point it out when it matters.
• The list is fixed so weeks can be compared. New questions are added as a new version and reported separately for their first weeks.

The panel

• All runs for a week are made on the same day (1 November 2026 for the first full week). English questions use an approximate location of the United States where the API supports it.
• Per category, the 240 answers give n = 40: ChatGPT 4, Gemini 8, Claude 4 and Perplexity 24. Not every assistant answers every question, to keep costs sensible.
• We record the model name every week, because providers change their models often.

What we count

Mention. A brand is mentioned when its name, or a known alternative spelling, appears in the answer. We match names against a list of brands in each category. New names that the assistants use are reviewed by hand every week and added when they are real brands.

Share of mention. The percentage of answers in a category that name the brand at least once. Example: HubSpot was named in 37 of 40 CRM answers, a share of mention of 93% (n = 40).

Pooled share of mention. The share across all four assistants combined. Because Perplexity answers three times per question, it supplies 24 of the 40 answers in each category and carries the most weight. We also publish per-assistant figures.

Named first (top pick). The brand that appears first in the answer. Average position is the mean order in which a brand appears when it is named (1 = first).

Citation. A web page that the assistant attaches to its answer as a source, taken from the API's own citation data. Each URL is reduced to its domain (for example blog.hubspot.com).

Cited in X% of answers. The share of answers that cite a domain at least once. A domain counts once per answer, however many times that answer links to it. We do this because Gemini attaches a source to almost every sentence, and raw link counts would overweight it.

Owned vs third-party citations. A citation is owned when the domain belongs to a brand in the same list (the vendor's own website). Everything else is third-party: review sites, publishers, forums such as Reddit, video sites such as YouTube, and vendors that are not on our list. The owned share is calculated over unique domain-per-answer citations (1,394 in the week of 1 November 2026).

Agreement between assistants. On the 24 questions that all four assistants answer, we check whether all four named the same brand first. For Perplexity we use its first run.

Stability. For Perplexity we compare its three runs of each question and count how often the same brand is named first in all three.

Week-over-week change. The difference in percentage points from the previous week, on the same questions and the same panel.

What the numbers do not show

• API answers are not the consumer apps. The ChatGPT, Gemini, Claude and Perplexity apps use different settings, models, personalisation and location. Our numbers are a consistent benchmark, not a copy of what any one person sees.
• Gemini API is not Google AI Mode or AI Overviews. Those are separate products and can cite very different sources.
• Perplexity's "fast" preset is not necessarily Perplexity's own Sonar model. This week it was served by openai/gpt-6-luna.
• Small samples. ChatGPT and Claude give 4 answers per category per week. Treat per-assistant figures as signals, not verdicts. A change of under about 10 percentage points between weeks is usually noise.
• One day per week. Answers can change from day to day and from run to run.
• English and US only. The questions are in English with a US location hint. Results for other languages and countries can differ.
• Name matching. A brand that is not yet on our list is missed until it is added. Common words used as brand names can be under-counted.
• Mentions are not endorsements. A brand can be named as an option to avoid. We count names and order, not sentiment.
• Six software categories only. The results say nothing about other markets.

Honesty rules

• Every number comes from that week's measurement. Nothing is estimated or invented.
• n is shown next to every percentage, and small samples are labelled as small.
• If an assistant could not be measured in a given week, we say so.
• Sponsored placements, if there ever are any, are labelled and never change the data.

How to cite

Mention Impossible, "AI Visibility Index", week of 1 November 2026. Method: https://data.mentionimpossible.com/method/

Questions and corrections

Reply to any issue of the newsletter, or email hello@mentionimpossible.com. We read every message and correct errors publicly.

— The Mention Impossible team