How AI wants chocolate

Every brand, named first

Indexed against the average of the 55 indexed chocolate brands, so 100 is the category average and a brand is read against the rivals it actually competes with.Indexed against the average of every indexed brand in the whole index — 2,582 brands across 67 categories — so 100 is the average brand anywhere here and a number reads the same on every board. 55 chocolate brands rated by the assistants in the latest weekly run; Named in is the share of 60 AI answers that name each brand.

What this category is rated on

The brands are most divided on values & integrity: Tony's Chocolonely is rated 6.7 and Twix 3.2 on the 1–7 scale, a spread of 136 against the feature's usual spread (100 = as contested as usual). Against every category, chocolate is rated highest on emotional magnetism (126) and lowest on emotional bond (83).

A feature's category mean is the average of its brands' ratings, 1–7. Vs all categories indexes it against every indexed brand across every category (100 = the average brand anywhere). The contested feature is the one the category's brands are most spread on, measured against how spread that feature usually is.

Rated well is not the same as named

Across the 55 brands, how highly the models rate a brand when asked and how often their answers name it barely go together (rank correlation 0.29).

Rated well, rarely named: Desire above the category median, named in 5% of answers or fewer. Named, not desired: among the most-named quarter of the category, yet indifferent or below. A relationship inside one week's run, never a cause.

Where the models disagree

The models broadly agree on these brands (mean correlation 0.79 across 3 pairs: ChatGPT–Claude 0.77 · ChatGPT–Gemini 0.83 · Claude–Gemini 0.78). Still, 311 of 1,705 brand-statement ratings (18%) split by 2 points or more on the 1–7 scale, and the typical brand's Desire differs by 7 points between its most and least generous model.

The widest 5 of 311 splits.

Across the whole index, on the same brands: Gemini rates +4.5 points of Desire % against its peers; ChatGPT rates −1.7 points of Desire % against its peers; Claude rates −3.0 points of Desire % against its peers. A generous model lifts every brand it rates, which is why scores are combined across models and indexed within the category.

What sets the leading brands apart

For the category's leading named brands, the glossary statements each is rated furthest above and below the category's average, in points on the 1–7 scale.

What the answers say

The category's answers say enough to judge 47 of the 1,705 brand-statement ratings (2.8%). Of those, 72% back the rating, 2% lean against it and the rest are mixed. The answers were not written to describe every brand on every statement, so this is a check on the ratings, not a second score. Counts only; no answer text is published.

Method & limitations

Every week, ChatGPT, Claude and Gemini are each asked directly, the way a survey asks a respondent, to rate every brand in the category on the 31 statements of the Havas Science of Desire glossary, on the survey's own 1–7 agreement scale. Statements roll up to nine features and the three Drivers of Desire (a feature is the mean of its statements; a driver is the sum of its three features, out of 21; Desire is the sum of the drivers, out of 63). Each brand is indexed against the average brand in its own category — 100 is the category average, 115 is fifteen per cent above it — and the Vs all categories switch re-indexes the same scores against every indexed brand in the index. Desire % is Desire as a share of its maximum. A brand is above indifference when its Desire is above two-thirds of its maximum (42 of 63), below indifference at or under one-third (21), and indifferent in between. The summary beside each brand says where it stands on the three drivers: an index of 110 or more counts as ahead, 90 or less as behind.

These are the assistants' own ratings when asked. They are not a survey of people and not an endorsement, and they are not read from what the assistants happen to say in an answer: the answers to the category's unbranded consumer prompts are used only to count how often each brand is named, and as supporting evidence. Relationships between ratings and how often a brand is named are correlational. A brand is indexed only when all 31 statements have been rated in a run by at least two models, and a category needs at least 4 such brands to get a board. See the full methodology.

Read the category report

Which chocolate brands the models recommend, the sources they trust and how often a brand page is read but not cited: Chocolate · The Category Report.

Other categories