How the benchmark is built

What we measured

The benchmark pools … AI answers from … AI search studies run by We AI & Data Advisors between …. Each study tracked a set of questions (… in total) that real people ask about an industry, and collected every answer each AI engine gave, every day or week, with the web sources it retrieved and cited and the brands it named. The studies were run on Peec AI, Profound and Meltwater.

Every row of every export was loaded, and nothing was sampled. Duplicate exports were removed, and truncated exports were rebuilt from the source platform.

Dimensions you can filter by

• Industry: 11 sectors, from Automotive to Travel & Tourism. Every sector pools studies from at least two organisations.
• AI engine: …. The OpenAI API is tracked separately from the ChatGPT app because it behaves differently.
• Market and language: the market each question was asked from (… markets) and the language of the question (… languages).
• Question theme: every question was classified into one of 12 themes:
• Question type: branded questions name the brand a study was built around, or one of its products; unbranded questions ask about a need or a category.
• Month: when the answer was collected.

How sources are counted and classified

A source is a web domain that an engine retrieved or cited for an answer. Sub-domains roll up to one site, so pubmed.ncbi.nlm.nih.gov counts as nih.gov. “Sources per answer” counts distinct sites among the answers that used sources at all. Each of the … domains is assigned to one type:

Classification combines the source platforms’ own labels, rules for public-sector and academic domains, and curated lists for social, reference and marketplace sites.

One data source does not record structured citations for DeepSeek and Meta Llama answers, so those engines show few or no sources here. Their answers still count towards tone and brand measures.

Brands and organisations named

Across all answers, AI engines named … distinct brands and organisations. Spelling variants are merged, so “Apple” and “Apple Inc.” count once. The count covers companies, products, public bodies and other organisations, as the source platforms recorded them.

• … were named in at least five answers.
• … were named in at least ten answers.

… studies record every company or organisation an answer names, and … record only the brands each study was set up to track, so most of the names come from the studies with full coverage.

Brand framing (tone)

For English-language answers, every sentence that names a brand is scored with VADER, a sentence-level lexicon model, extended with business terms such as “award-winning”, “overpriced” and “recall”. Brand names are masked before scoring, so a name like “Kia” or “Bright” cannot sway the result. An answer counts as positive when the average score of its brand sentences is above 0.3, and as negative when it is below −0.05.

We validated this in two ways.

• On 35,500 answers that Meltwater also labels, the method rates 42.6% positive, against 44.1% for Meltwater. Positive-labelled answers score higher than neutral ones.
• For 39 brands tracked in Peec AI over the same 90 days, our brand scores correlate with Peec’s sentiment at r = 0.65 (rank correlation 0.73). The average gap is 3.8 points on a 0–100 scale.

Read tone as a comparison between groups rather than as an absolute verdict on any brand.

Anonymity and minimum sizes

• No client, study or question text is published, and the data files behind this site contain no study identifiers.
• Results are hidden when a selection covers fewer than 200 answers.
• Lists of sites appear only where the selection draws on studies for at least 2 different organisations. Each listed site must also be cited in studies for 2 or more organisations. Company websites are listed only when no single organisation’s studies supply more than two thirds of their citations.
• Inside each study, the focus brand’s own websites, including product, campaign and market sites, are grouped as “Focus brands’ own websites”.
• Industry pages report patterns across all the studies in that sector. A sector is shown only when it draws on studies for at least two organisations.

Things to bear in mind

• Studies ran at different times and for different lengths, so the mix of industries and engines varies by month. The Trends page follows a fixed set of long-running studies.
• Not every study tracked every engine. Compare engines within the same filters.
• Model-version profiles on the Trends page compare whichever studies used each version. The engine insights quote within-study comparisons.
• AI answers vary from one run to the next. Figures describe tendencies across many answers, not a guaranteed result for any single question.

Updates

New studies are added through the same pipeline. The site is regenerated from the refreshed data, and this page shows the current data version.