Each benchmark states its run date, market and prompt count; vendor share is the share of answers naming the vendor and is never blended with citations.

How to read a benchmark

Start with the vendor share. It's the share of answers that name the vendor, nothing fancier. If a vendor shows 40%, it was named in 40% of the answers we read. It's not a market share, and it's not a ranking of product quality.

Then look at the run date. Engines change, and every report states the model ids used, so you can see which versions answered. The engines as of 30 September 2026 are ChatGPT (GPT-5, with web search on), Claude (Claude Sonnet 5.5, with web search on), Gemini (Gemini 3.6 Flash, with search grounding on) and Perplexity (Sonar). A run from a different date may not match this one.

Next, the market. Each prompt is asked from the market's locale, so London, GB by default and New York, US on the US edition. The same buyer question can get different vendors in different places. Check which market you're reading before you compare anything.

Then the prompt count. Each benchmark page states how many buyer questions were asked, and each question goes to four engines. So 15 questions gives 60 answers, and 100 questions gives 400. More answers means a steadier picture, but only up to a point. Every answer is read, and every claim is quoted.

Also, keep mentions and citations apart. A mention means the brand is named in the prose of the answer. A citation means one of the vendor's own domains appears in the answer's source list. We never blend the two into a single score, because they tell you different things:

A vendor named often but cited rarely has a page problem.
A vendor cited often but named rarely has a different problem.
A vendor with both is doing the work on both fronts.

Finally, read the verdicts. Each vertical gets one of three labels, drawn from quoted claims: Recognised, not cited; Entrenched, where one rival appears in the majority of answers; or Open, where no vendor is named in more than about a quarter of answers. The thresholds are printed next to the counts, so you can check the label yourself.

What a benchmark does not show

A public run is a sample, not a census. Each prompt is asked once, in a fresh single-turn conversation, and the same question tomorrow can get a different answer. So treat a gap of a few points between two vendors as noise unless the counts say otherwise.

The answers also come through the API. They sit close to what the consumer apps show, but they aren't identical. If you open the app and get a different list, that doesn't mean the benchmark is wrong, it means you've asked a slightly different system.

Three things sit outside what this method can tell you:

Traffic. A citation is a source link, not a click, so a high citation count says nothing about visits.
Why a model named you. We record what the answer said, not the reasoning behind it.
Stability over time. One run is one moment, and a single date can't show you a trend.

Then there's the question set itself. A public benchmark uses one fixed set of buyer questions for the whole category. Your own buyers may phrase things differently, or ask about a narrower niche than the category page covers. So the category view is a starting point, and your own prompts, edited and approved by you, are where the detail lives.

Read the benchmark for what it is: a dated, quoted record of which vendors the engines named for a set of buyer questions. That's a narrow claim. It's also one you can check line by line.