The UK Chocolate AI Visibility Index 2026

Why this study matters

AI systems are increasingly used as recommendation environments. A consumer can ask which chocolate to buy as a gift, which brands have stronger ethical sourcing credentials, which products suit a dietary requirement, or which brands are worth trying in a particular market, and the answer is not a ranked list of webpages. The model constructs its own consideration set, compares brands, recommends products and often selects one or more preferred options.

That creates questions conventional search metrics cannot answer. A brand can be known to a model without being regularly recommended, mentioned frequently without becoming the first choice, strong for one consumer need and absent from another, and highly visible in one model while being marginal in the next. A globally recognised brand can also behave differently once the consumer names a specific market such as the UK.

This study was designed to measure those differences at scale rather than rely on isolated AI answers. It does not try to identify a single universally "best" chocolate brand. It analyses how brands compete for inclusion and selection inside model-native AI recommendation environments, and it is the second release in the UK AI Search Visibility Index series, following The UK Accounting Software AI Visibility Index 2026.

Study design

Netsleek ran 100 consumer-oriented chocolate prompts through the APIs of the three AI systems included in this study, ChatGPT, Gemini and Claude, across seven repeated observation rounds, producing 2,100 responses. Every attempted observation was captured successfully, giving a 100% valid capture rate, and each platform contributed 700 responses. The seven rounds were collected between 15 and 20 September 2026, with two rounds captured on 19 September, so the dataset covers seven rounds across six calendar dates.

Model-native capture

All observations used standard model completions with no live web search, browsing or retrieval grounding enabled. The study therefore measures what the models recall, associate and select from their learned knowledge, rather than which live webpages they would retrieve. Citation and source analysis sits outside this part and will be covered in a separate search-grounded part of the research.

Two evidence layers

During analysis the prompt set was split by whether the wording explicitly established a UK or British context. The 38 UK-grounded prompts produced 798 observations (38 prompts × 3 platforms × 7 runs) and form the evidence base for every UK-specific claim in this report. The 62 open-context prompts produced 1,302 observations and show which brands emerge when the consumer has not yet said where they intend to buy.

The two layers are never pooled. Open-context answers are excluded from UK rankings, and differences between the layers are described as observed context differences rather than causal effects, because the two groups contain different mixes of consumer questions and were not built as matched A/B pairs.

Tracked brand cohort

Thirteen brands were tracked consistently across the study: Lindt, Cadbury, Green & Black's, Hotel Chocolat, Tony's Chocolonely, Divine Chocolate, Montezuma's, Galaxy, Ferrero Rocher, Thorntons, KitKat, Terry's and Booja-Booja. The full response text was also audited for brands outside that cohort, which is how emergent competitors such as Ghirardelli, Godiva and Charbonnel et Walker were identified.

No single metric is treated as a complete measure of performance. A high mention rate does not mean a brand is frequently preferred, and a brand with lower overall visibility can lead within a specific consumer need, so the results are read through each of these lenses in turn.

¹ Responses were collected through each provider's API using fixed model versions available within Netsleek's measurement system, selected to support consistent, repeatable measurement at scale, rather than through the consumer apps, whose default models may differ. Model versions used during the observation period were OpenAI GPT‑4o, Google Gemini 2.5 Flash and Anthropic Claude Haiku 4.5. Results apply specifically to the tested model versions and the September 2026 observation period. AI platforms and underlying models change over time, so these findings should be interpreted as a time-bounded measurement rather than a permanent representation of current or future model behaviour.

01Four brands occupied the broadest UK consideration space

When an AI model is told the consumer is looking for chocolate in the UK, a clear upper group emerges within the tracked cohort. Across the 798 UK-grounded observations, Lindt appeared in 71.7% of responses, Green & Black's in 69.9%, Cadbury in 67.5% and Hotel Chocolat in 63.0%, a spread of only 8.7 percentage points between first and fourth.

After those four the distribution widens. Montezuma's appeared in 45.7% of responses, a 17.3 point drop from Hotel Chocolat, followed by Thorntons and Galaxy at just under 30%. A lower position here does not necessarily mean a weaker recommendation, because some brands are tied to narrower consumer needs and perform extremely strongly when those needs are expressed, which Finding 03 tests directly.

The UK consideration set is not restricted to British brands

Explicit UK context did not produce a British-only answer. Lindt, a Swiss brand, had the highest presence of any tracked brand, and Tony's Chocolonely and Ferrero Rocher remained part of the landscape alongside brands more closely associated with the British market. Geographic context appears to shape relevance rather than instruct the model to return domestic brands, so the useful question for a global brand is whether the model still considers it relevant once the consumer says where they intend to buy.

02Presence is not priority: the brands AI mentions most are not the brands it surfaces first

Presence shows whether a brand entered an AI-generated consideration set, but not where it sat within the answer. Netsleek therefore recorded the first tracked chocolate brand surfaced in each of the 798 UK-grounded responses, and the competitive picture changed. Although Lindt led on presence and Green & Black's ranked second, Cadbury was the brand most often surfaced first, leading 269 responses (33.7%), ahead of Lindt at 28.4%, Hotel Chocolat at 10.2% and Green & Black's at 8.8%.

First when present shows how often a brand held the first tracked position among the responses in which it appeared at all. None of these measures describe market share or purchasing behaviour.

Cadbury moves from third in presence to first in position

Cadbury appeared in 67.5% of UK-grounded responses, behind Lindt and Green & Black's, yet it was the first tracked brand in 33.7% of all responses. When Cadbury appeared at all, it led the tracked set 49.9% of the time, which means it did not merely enter many consideration sets but frequently opened them.

Lindt combines breadth with prominence

Lindt had the broadest UK presence and was also first in 28.4% of responses, leading 39.6% of the answers it appeared in. It was the only tracked brand to combine more than 70% presence with more than 25% first-position share, so its significance comes from both breadth of inclusion and prominence within the answer.

Green & Black's shows the largest presence-position gap among the leaders

Green & Black's appeared in 69.9% of UK-grounded responses, only narrowly behind Lindt, yet it was first in just 8.8% and led only 12.5% of the answers that included it. Hotel Chocolat followed a similar pattern at 63.0% presence and 10.2% first position. Both brands have exceptionally broad relevance across the prompt set, but broad inclusion does not translate proportionally into leading position.

Smaller brands can punch above their presence, and some large ones never lead

Tony's Chocolonely appeared in 22.2% of responses but led nearly one in five of its appearances (19.8%), and Divine Chocolate showed a smaller version of the same pattern at 14.6%. Terry's (38.6%) and Booja-Booja (31.3%) posted high first-when-present rates, although these rest on only 44 and 32 appearances respectively and should be read with that sample size in mind.

The opposite pattern is just as clear. Montezuma's appeared in 45.7% of responses but led only 3.3%, Thorntons appeared in 29.7% but led 0.1%, Galaxy led 0.5% from 29.2% presence, and Ferrero Rocher was present in 11.5% of responses without leading a single one of the 798. These brands are clearly part of the models' UK chocolate knowledge, but they play a different role in the answer from Cadbury or Lindt.

03Different consumer needs produce different brand leaders

The overall landscape shows which brands appear broadly, but not what they are being associated with. The 38 UK-grounded prompts were grouped into eight broad consumer-need themes for reporting, and because each theme contains a different number of prompts, the analysis compares rates within each theme rather than raw mention counts. The result is a far more differentiated market, where some brands appear across almost every question and others become dominant only when a specific need is expressed.

Tony's Chocolonely and Divine own ethical sourcing

The sharpest shift in the study occurs around ethical cocoa and responsible sourcing. Tony's Chocolonely appeared in 59 of 63 responses in this group (93.7%) against 22.2% across the UK dataset, and Divine Chocolate rose from 24.8% overall to 90.5%, with Green & Black's following at 82.5%. Response order reinforced the association: Divine was first in 28 responses, Tony's in 26 and Green & Black's in 9, so either Divine or Tony's led 54 of the 63 answers, or 85.7% of the group.

Hotel Chocolat becomes near-universal once the purchase is a gift

Across the six premium, gifting and occasion prompts, Hotel Chocolat appeared in 122 of 126 responses (96.8%), up from 63.0% overall, and was the most frequent first brand at 37.3%, ahead of Lindt at 31.7% and Cadbury at 30.2%. Thorntons also became far more visible in this part of the market, rising from 29.7% overall to 60.3%, which shows how differently a brand's competitive position can look once the reason for buying becomes specific.

Lindt dominates taste and chocolate type

Across questions on taste, texture and chocolate type, including milk, dark and white chocolate, Lindt appeared in 145 of 147 responses (98.6%) and was first in 42.2%. Its visibility is not confined to one buying occasion, which helps explain why it performs so strongly across the UK landscape as a whole.

Green & Black's wins on breadth, Cadbury on the mainstream

Green & Black's appeared in 94.6% of taste and chocolate-type responses, 83.8% of general discovery responses, 82.5% of ethics responses, 74.6% of dietary responses and 58.3% of flavour and filling responses. That breadth across very different needs, combined with its low first-position rate, suggests its strength lies in being relevant almost everywhere rather than consistently opening the answer.

Cadbury was repeatedly prominent in mainstream discovery, flavour and value questions, leading 44.8% of general discovery responses, 40.5% of flavour and filling responses and 39.7% of value and retail responses. The value category was also the most fragmented in the study, since 27 of its 63 responses contained none of the 13 tracked brands at all, a sign that value-oriented discovery introduces a wider competitive set.

Small brands can hold strong specialist positions

Booja-Booja appeared in only 4.0% of UK-grounded responses overall but in 34.9% of dietary-requirement responses, more than eight times its overall rate, and it was the first tracked brand in 15.9% of that group. Terry's showed a comparable pattern, rising from 5.5% overall to 35.7% of flavour and filling responses and leading 20.2% of them.

04The same UK question produces different competitive sets across AI platforms

Each model received the same 38 UK-grounded prompts across seven rounds, giving 266 UK-grounded responses per platform, yet ChatGPT, Gemini and Claude frequently surfaced different brand combinations. For some brands the difference between models exceeded 40 or even 60 percentage points, as Fig. 1 at the top of this report shows.

Engine gap is the difference between a brand's highest and lowest presence rate across the three platforms. It is a simple observable measure of divergence, not a composite score.

Montezuma's and Thorntons move in opposite directions

No tracked brand diverged more than Montezuma's, which appeared in 70.3% of Gemini responses and 60.9% of ChatGPT responses but only 6.0% of Claude responses. The prompts and the UK context did not change, only the model generating the answer, so a brand measuring a single platform could conclude it was either highly prominent or marginal. Thorntons showed almost the reverse pattern at 56.0% on Claude against 22.9% on ChatGPT and 10.2% on Gemini, and Ferrero Rocher reached 29.7% on Claude against under 3% on both other models.

Each model has brands it favours

Lindt appeared in 93.2% of Claude's UK-grounded responses, against 66.9% on ChatGPT and 54.9% on Gemini. Green & Black's was strongest on ChatGPT at 80.1%, and Gemini gave Galaxy (46.6%) and KitKat (17.7%) far more visibility than either other model. Cadbury and Hotel Chocolat showed narrower ranges, which indicates broader cross-engine presence among the most visible brands.

Engine differences became even larger once response order was considered. ChatGPT most often surfaced Cadbury first (40.6%) and Gemini did the same (49.2%), while Claude surfaced Lindt first in 74.4% of its UK-grounded responses, against 4.5% on ChatGPT and 6.0% on Gemini. This does not make one model objectively better, but it shows that response structure and brand prominence are highly model-dependent.

Matched answers shared only about half their brands

Because every prompt was run on all three models in every round, the tracked-brand sets from matched responses could be compared directly. On average the intersection represented approximately 49.5% of the combined tracked-brand set, with ChatGPT and Gemini overlapping most and Claude diverging more substantially from both.

Response length does not explain the gaps. Gemini's answers were more than twice as long as ChatGPT's and more than three times as long as Claude's, yet it named only slightly more tracked brands per response. Length cannot explain why Montezuma's appears in 70.3% of Gemini responses and 6.0% of Claude responses, or why Ferrero Rocher appears in 29.7% of Claude responses and under 3% elsewhere.

05How stable are AI brand recommendations across repeated runs?

Generative AI does not behave like a search results page. Ask the same model the same question again and the wording, examples, ordering and brands can all change, which raises a fundamental measurement problem: how much of an observed brand landscape is persistent signal, and how much is ordinary generative variation? The repeated structure of this study was designed to answer that.

All 798 UK-grounded responses were textually unique, and not one answer was reproduced word for word. At the level of the individual response, variation was universal. The competitive structure beneath those responses was considerably more persistent, which is the central distinction of this section: changing answers do not necessarily mean a changing brand landscape.

The relative order of the leading four shifted slightly from round to round, with Lindt leading several rounds and Cadbury and Green & Black's each moving to the front at points, but the same brands repeatedly occupied the top of the landscape. A brand moving from 70% in one round to 67% in the next does not automatically mean its AI position has weakened, because some movement is present even when the overall structure is intact.

Aggregate rankings were remarkably consistent

Comparing the tracked-brand ranking from every round against every other round gave an average rank correlation of approximately 0.98, and even the least similar pair of rounds correlated at approximately 0.96. Two forms of variability therefore exist at the same time: response-level variability in wording, examples and composition, and aggregate stability in the broader ordering of brands.

Different brands carry different levels of variability

Montezuma's moved across a 14.9 point range, far wider than the roughly 6 to 10 points seen for the other leaders. A single measurement taken in Round 5 would have put it at 36.0%, while one taken in Round 3 would have shown 50.9%, and although both observations are real, neither represents the full pattern. Variability is not uniform across brands, so a single universal adjustment would conceal real differences in how stable each brand's representation is.

First position was often more persistent than the wording suggested

The 38 UK prompts across three models created 114 prompt-model combinations, each observed seven times. In 61 of them (53.5%) the same tracked brand held first position in all seven runs, and a further 38 produced only two different first-position brands, so 86.8% of combinations produced no more than two leading brands across the whole series. Where a leading pattern formed, the dominant first-position brand held that spot in an average of 85.7% of the seven observations.

Presence persistence and candidate-set overlap

The share of the 114 combinations in which a brand appeared in all seven runs was 52.6% for Lindt, 49.1% for Green & Black's, 45.6% for Cadbury, 40.4% for Hotel Chocolat and 28.1% for Montezuma's. This separates frequency from persistence, since two brands can share the same overall rate while one appears sporadically across many questions and the other appears almost every time within a narrower set.

Repeated responses to the same prompt on the same model shared approximately 70.7% of their combined tracked-brand set on average, with a median of 75.0%. AI recommendations are neither fixed nor random: they typically contain a persistent core competitive set surrounded by a more variable outer set of alternatives, and a competitor that appears once may belong to that changing edge rather than the stable core.

Model choice produced far more variation than repetition

Across the tracked brands, median variation between the seven rounds was approximately 6.1 percentage points, while the median difference between the three platforms was approximately 15.0 points. For Montezuma's the run-to-run range was 14.9 points against a model range of 64.3, and for Thorntons it was 3.5 points against 45.8. An organisation measuring only one AI platform can therefore obtain a highly repeatable answer while still holding an incomplete picture of the wider recommendation environment.

This is not a universal noise floor

The repeated design shows the ordinary variation observed within this specific experiment, not a fixed variability threshold for AI search in general. Observed variation depends on the models, the category, the prompts, the metric, the observation period and the type of consumer intent, so the more cautious and defensible conclusion is that different AI visibility signals can exhibit different levels of natural variability. Future research can test whether that pattern holds across industries, models and measures such as citation occurrence and citation-domain selection.

One response is an observation. A repeated pattern is evidence.

06The open-context landscape: what AI recommends when the consumer does not say where they are

Sixty-two of the 100 prompts asked about products, quality, gifting, dietary needs, value and other chocolate topics without naming a country, producing 1,302 open-context responses. These are not a global market ranking, since no controlled set of countries was tested. They show which brands AI models introduce before geography constrains the answer, which mirrors the early stage of many real consumer journeys where research starts before the consumer says where they will buy.

* Not part of the original 13-brand tracked cohort. Prompts reached counts the distinct open-context questions in which the brand appeared in at least one of the 21 observations per prompt.

Lindt held its lead, but the field around it changed

Lindt appeared in 892 of 1,302 open-context responses (68.5%) and reached 59 of the 62 prompts, or 95.2% of the distinct questions, so its presence was not tied to one narrow characteristic. The field around it, however, became far more international once geography was removed.

Ghirardelli is the clearest example. It appeared in 53.3% of open-context responses and reached 59 of 62 prompts, matching Lindt's coverage, despite not being part of the tracked cohort. Godiva (32.3%, 46 prompts) and Valrhona (27.8%, 43 prompts) were also broad participants, and Toblerone appeared in only 19.0% of responses but surfaced across 43 prompts, a reminder that frequency and breadth are different forms of AI visibility.

British brands still travel

Cadbury appeared in 28.0% of open-context responses across 42 prompts, and Green & Black's in 22.0% across 45, so neither needed UK context to be introduced. The recurring international consideration set also included Tony's Chocolonely, Theo Chocolate, Amedei, Alter Eco, Endangered Species Chocolate, Michel Cluizel, Dandelion Chocolate, Divine Chocolate, Milka, Raaka, Hu and Ritter Sport.

The models still disagreed

Model differences continued without geography. Valrhona appeared in 53.2% of Gemini's open-context responses and never once in Claude's, while Toblerone reached 42.4% on Claude against 5.5% on Gemini. Tony's Chocolonely was the notable exception, with near-identical presence across all three platforms, which suggests some brand associations transfer across models much more consistently than others.

07What changes when the consumer specifies the UK

Moving from open discovery to a UK-grounded question materially changes which brands AI systems surface most often. The shift is not simply a matter of swapping international brands for British ones: some global brands stay highly visible, some brands become dramatically more prominent, and others remain recognised but appear far less consistently once UK relevance is established.

Only Lindt and Cadbury appear in the top five of both environments, which is a substantial change in the composition of the leading consideration set. The two groups contain different mixes of consumer questions, so the differences below are observed context differences rather than a controlled estimate of adding "UK" to an otherwise identical prompt.

Market amplification: Hotel Chocolat, Montezuma's and Green & Black's

Hotel Chocolat showed the largest observed difference, rising from 5.7% of open-context responses to 63.0% of UK-grounded responses and from 18 of 62 open prompts to 37 of 38 UK prompts. Montezuma's moved from almost absent, at 0.2% and just two prompts, to the fifth most surfaced tracked brand at 45.7% across 34 of 38 prompts. Both are strong examples of the difference between general model recall and market-grounded relevance.

Green & Black's followed a different path. It was already known across 45 open-context prompts, but its frequency rose from 22.0% to 69.9%, turning it from an occasional candidate across many questions into one that appeared repeatedly. Cadbury combined both characteristics, with meaningful international recall at 28.0% and much stronger UK relevance at 67.5% across 37 of 38 prompts.

Cross-context persistence: Lindt

Lindt changed very little, at 68.5% in open context and 71.7% in UK-grounded discovery, and it reached every one of the 38 UK prompts. It does not depend on UK context to enter consideration, and unlike several international brands it does not recede once the market is specified, which makes it the clearest example of cross-context persistence in the study.

Global brands that recede once the UK is explicit

Ghirardelli fell from 53.3% to 14.3%, although it still appeared at least once in 32 of the 38 UK prompts, so the models associate it with many UK chocolate needs but surface it far less consistently. Godiva fell from 32.3% to 8.1% and Valrhona from 27.8% to 3.4%, which shows that strong representation in general model knowledge does not guarantee prominence once a market-specific set is being built.

Two competitive moments

For a global brand, AI-assisted discovery creates at least two moments that need to be won separately. In open discovery the question is whether the brand enters the initial consideration set before any geographic constraint is applied, and in market-grounded discovery it is whether the brand stays competitive once the consumer says where the product must be relevant or available. The competitors a brand faces also change with context: around Lindt, the open-context field is led by Ghirardelli, Godiva, Cadbury and Valrhona, while the UK field is led by Green & Black's, Cadbury, Hotel Chocolat and Montezuma's.

08The competitors AI introduced beyond the tracked cohort

The study tracked 13 brands consistently, but the models were never restricted to them. Across the 2,100 responses they repeatedly introduced additional chocolate companies, some occasionally and others often enough that excluding them would give an incomplete picture of the landscape the models were constructing.

These brands were identified from the generated response text rather than supplied as the fixed cohort. The list highlights the most prominent examples and is not a complete census of every brand named in the dataset.

Ghirardelli would have been one of the largest brands in the study

Ghirardelli appeared in 808 of 2,100 responses (38.5%) and reached 91 of the 100 prompts at least once. Had it been tracked, that presence would have placed it between Green & Black's (40.2%) and Hotel Chocolat (27.5%) on overall frequency, and in open-context discovery it was second only to Lindt. Godiva (23.1%) and Valrhona (18.5%) also exceeded several tracked brands across the full dataset, including Montezuma's (17.5%) and Tony's Chocolonely (17.0%).

UK context introduces a different set of challengers

Once the UK was explicit, the models introduced another group of relevant competitors. Charbonnel et Walker rose from 1.6% to 22.1% across 20 of the 38 UK prompts, Rococo Chocolates from 0.4% to 20.2% across 21 prompts, and Prestat from 0.2% to 17.4% across 17 prompts. Willie's Cacao did not appear in the identified open-context responses at all, yet reached 8.1% of UK-grounded responses across 15 prompts.

The reverse also held. Theo Chocolate (13.5%), Dandelion Chocolate (11.1%), Endangered Species Chocolate (8.6%), Raaka (6.0%) and Hu (6.0%) appeared only in open-context discovery and had no UK-grounded presence. AI did not simply enlarge the competitor list; it reconstructed it according to the recommendation environment.

Frequency and breadth are different forms of visibility

Milka appeared in only 7.5% of responses but surfaced in 42 of the 100 prompts, and Toblerone appeared in 13.8% of responses across 60 prompts. A brand can be high frequency and high breadth, high frequency within a narrow part of the journey, low frequency but broad, or a narrow specialist, and a single total mention count cannot distinguish between them.

09Which brands AI places together

A brand rarely appears alone in an AI recommendation. Most responses contain several alternatives, which creates a temporary consideration set around the consumer's question and makes it possible to measure which brands repeatedly compete inside the same answers. Co-occurrence here means only that both brands appeared somewhere in the same response, so it shows shared consideration space rather than proving that consumers substitute one brand for the other.

A core four-brand consideration set

Every pairing within Lindt, Green & Black's, Cadbury and Hotel Chocolat appeared together in at least 46% of UK-grounded responses, led by Lindt with Green & Black's at 56.4%. The four most visible brands are therefore competing for attention inside the same answers rather than dominating separate types of question.

Competitive neighbourhoods

Galaxy appeared in 233 UK-grounded responses and Cadbury was present in 228 of them (97.9%), one of the strongest conditional relationships in the dataset. Hotel Chocolat sits at the centre of a premium British cluster: it appeared alongside Charbonnel et Walker in 97.7% of that brand's appearances, alongside Prestat in 97.1% and alongside Rococo in 91.9%, which reinforces the premium and gifting association found in Finding 03.

Montezuma's belongs to more than one neighbourhood. When it appeared, Green & Black's was present 81.9% of the time, Hotel Chocolat 78.9%, Cadbury 63.8% and Lindt 61.4%, and Willie's Cacao appeared alongside Montezuma's in 64 of its 70 UK-grounded responses (91.4%). Tony's Chocolonely appeared with Green & Black's in 160 of its 177 responses (90.4%), placing the specialist ethical brand and the broad multi-need brand in the same quality-led consideration space.

Mainstream and premium sets overlap

The data does not form separate islands of mass-market, premium, ethical and specialist brands. Hotel Chocolat appeared alongside Cadbury in 46.0% of all UK-grounded responses, and Cadbury alongside Montezuma's in 29.2%, so AI answers frequently combine brands at different price points and market positions, with the consumer need in the prompt acting as the common denominator.

The neighbourhoods also move with geography. Lindt and Ghirardelli were the most frequent open-context pairing at 51.5% but fell to 14.3% in UK-grounded discovery, while UK-oriented pairings expanded sharply, so AI competition is better understood as a network of overlapping consideration sets than as a single leaderboard.

10Product and portfolio visibility: which products carry the brand

AI models frequently move beyond the master brand to name specific bars, ranges, boxed assortments and product families. A consumer may meet Cadbury through Dairy Milk, Lindt through Excellence or Lindor, Terry's through Chocolate Orange and Charbonnel et Walker through its Marc de Champagne truffles, while other brands are represented broadly without one recognisable product doing the work.

Named products were identified from the response text and linked back to their brand. The table highlights recurring products and is not an exhaustive catalogue.

Cadbury: a flagship plus a portfolio

Dairy Milk appeared in 41.5% of all UK-grounded responses and alongside Cadbury in 329 of the 539 Cadbury-containing responses (61.0%). Buttons (10.8%), Roses (7.4%), Heroes (5.0%), Wispa (4.1%), Flake (3.3%), Twirl (2.9%), Crunchie (1.8%) and Creme Egg (1.0%) also recurred, and around 68.6% of Cadbury-containing responses named at least one of these products, with nearly one in five naming two or more.

Lindt: two major product families

Lindt Excellence appeared in 40.6% of UK-grounded responses and in about 56.3% of those containing Lindt, while Lindor appeared in 32.8% overall and 45.6% of Lindt responses. At least one of the two appeared in 469 of the 572 Lindt responses (82.0%) and both appeared together in 114, giving the brand one route through bars and higher cocoa content and another through truffles, texture and gifting.

Terry's and Charbonnel et Walker: hero-product brands

Terry's was explicitly named in 44 UK-grounded responses, and Chocolate Orange appeared alongside it in 40 of them (90.9%), the clearest case of a brand whose AI representation is concentrated in one product. Charbonnel et Walker, which emerged outside the tracked cohort, had its Marc de Champagne products named in 76 of its 176 responses (43.2%), tying it to a recognisable luxury proposition.

Montezuma's and Hotel Chocolat: brand-led representation

Montezuma's Absolute Black appeared in only 38 responses, around 10.4% of Montezuma's appearances, with Like No Udder recurring at a lower rate, so its visibility is spread across the brand and range. Hotel Chocolat appeared in 63.0% of UK-grounded responses, yet its named lines (Sleekster, Selectors, H-Box and Velvetiser) together appeared in only around 7.4% of Hotel Chocolat responses, which indicates that the master-brand proposition carries most of its visibility.

Product representation matters because models often explain brands through examples, describing a brand as premium, ethical, giftable or mainstream on the strength of the products they associate with it. For companies with large portfolios, the question is not only whether AI knows and recommends the brand, but which products it uses to understand, explain and represent that brand.

What the findings mean for brands competing in AI recommendation environments

The clearest conclusion from the study is that a brand's position in AI recommendation environments cannot be captured by one visibility percentage. Across 2,100 responses, performance changed with the market the consumer named, what they were trying to achieve, which model answered, which other brands entered the set, and whether the analysis looked at presence or leading position.

The better question is not "are we visible in AI?" but where, when, why and alongside whom AI considers our brand relevant.

Key findings at a glance

1. Lindt had the broadest UK-grounded presence at 71.7%, with the top four tracked brands within 8.7 points of each other.
2. Cadbury was surfaced first more often than any tracked brand, at 33.7% of UK-grounded responses.
3. Green & Black's appeared in 69.9% of UK answers but was first in only 8.8%.
4. Tony's Chocolonely and Divine rose to 93.7% and 90.5% presence on ethical sourcing questions and together led 85.7% of those answers.
5. Hotel Chocolat appeared in 96.8% of premium, gifting and occasion answers.
6. Montezuma's ranged from 6.0% to 70.3% depending on the platform, a 64.3 point engine gap.
7. Claude surfaced Lindt first in 74.4% of UK answers, while ChatGPT and Gemini most often led with Cadbury.
8. Rankings were stable across seven rounds, with an average rank correlation of about 0.98 despite every response being unique.
9. Platform choice produced roughly 2.5 times more variation than repeated runs (15.0 against 6.1 points at the median).
10. Ghirardelli, outside the tracked cohort, appeared in 53.3% of open-context answers and 38.5% overall.
11. Only Lindt and Cadbury ranked in the top five of both open-context and UK-grounded discovery.
12. Products carry brands: Dairy Milk appeared in 41.5% of UK answers, and Chocolate Orange accompanied Terry's 90.9% of the time.

What this study does not show, and what comes next

This part used non-search-grounded responses, so it does not establish which live webpages influence recommendations, which domains are retrieved or cited, whether brand-owned or third-party sources contribute more, or how behaviour changes when live retrieval is enabled. Those questions need a different capture condition, and mixing the two into one dataset would blur what each measures.

The comparison should reveal several distinct brand positions: brands strong in both environments, brands stronger in learned model knowledge than in the current retrievable web, and brands whose current digital and third-party footprint lifts them once live search is switched on. It will also separate recommendation stability from source stability, since the same brand can keep being recommended while the evidence behind it changes across models, runs, intents and markets.

Measurement definitions

Response presence

The share of relevant responses in which a brand appeared at least once, for example 500 of 798. It does not indicate that the brand was preferred or recommended first.

First tracked brand surfaced

The earliest brand from the tracked cohort to appear in a response. It measures response ordering, not purchase probability or explicit endorsement.

First when present

The share of a brand's appearances in which it was also the first tracked brand. It separates frequent presence from frequent leadership.

Prompt coverage

The number of distinct prompts where a brand appeared in at least one of up to 21 observations (3 platforms × 7 rounds). It measures breadth rather than frequency.

Engine gap

The percentage-point difference between a brand's highest and lowest presence across the three platforms. It is an observable measure, not a composite score.

Co-occurrence

Two brands appearing in the same response, with conditional co-occurrence asking how often brand B appears when brand A does. It shows shared consideration, not commercial substitution.

Product presence

The share of responses containing an identifiable named product or product family. It distinguishes master-brand visibility from product-driven visibility.

Percentage points (pp)

The arithmetic difference between two rates, so 5% to 15% is +10pp. It is not a relative percentage increase.

Rank consistency between rounds was assessed with rank correlation, since stability does not require identical wording. Netsleek publishes the tested models, sample sizes, observation period, capture condition, reporting categories, metric definitions and limitations, but does not disclose its proprietary prompt development, extraction architecture, entity resolution, classification logic, scoring, stability thresholds or diagnostic frameworks.

Limitations and responsible interpretation

Frequently asked questions

Research Authors

Ruan Masuret and Juanita Martinaglia are the co-founders of Netsleek, a research-led AI Search and Brand Discoverability agency. Their work examines how AI systems discover, interpret, evaluate and recommend brands across modern discovery environments.

Ruan and Juanita co-founded Netsleek, a research-led AI Search and Brand Discoverability agency. Explore further studies through Netsleek Research.