The UK Chocolate AI Visibility Index 2026
Which chocolate brands ChatGPT, Gemini and Claude mention when their answers carry source citations, which websites they cite, and why those two things tell different stories. Part 2 adds an evidence layer to the model-native baseline established in Part 1.
AI brand visibility and source visibility are different outcomes
Across 798 scheduled UK-explicit observations, Cadbury was mentioned in 404 (50.6%), while Cadbury's own websites accounted for only three of the 2,258 recorded UK citation occurrences. Hotel Chocolat appeared in 380 (47.6%), and its own website accounted for 82 recorded citations, so both brands were visible but their owned-source footprints looked very different.
The dataset also shows that apparently strong brand-publisher associations can shrink or disappear when the exact prompt and platform are held constant. The practical lesson is to measure who appears, which source is cited, and what consumer intent connects the two, rather than treating citation counts as a proxy for recommendations.
Updated 18 November 2026: brand figures were updated following a review of the brand-matching rules, and no brand ranking or finding changed. Methodology disclosures were also clarified, and a new analysis of cited source types by consumer need was added. See How to reproduce and audit this study for validation status.
Suggested citation: Netsleek Research (2026). The UK Chocolate AI Visibility Index 2026, Part 2: What AI Cites, Which Brands Appear and Why the Difference Matters. Netsleek. Available at https://www.netsleek.com/the-uk-chocolate-ai-visibility-index-2026-part-2/
Netsleek analysed 2,100 scheduled observations from ChatGPT, Gemini and Claude to 100 consumer chocolate prompts over seven rounds between 1 and 7 November 2026, capturing 5,119 source citations alongside the answer text. In the 798 UK-explicit observations, Cadbury was mentioned in 50.6% of answers, Lindt in 49.7% and Hotel Chocolat in 47.6%, yet citations to brands' own websites were distributed very differently, with three for Cadbury's own websites against 82 for Hotel Chocolat's. UK citations were led by Tesco (180), Which? (87), Ocado (84), Hotel Chocolat (82) and Ethical Consumer (63), and the three major supermarkets together accounted for 13.8% of UK citations. Tony's Chocolonely appeared in 77.2% of ChatGPT and Claude UK answers citing Ethical Consumer against a 28.4% overall baseline for those platforms, but no inclusion advantage appeared (−1.0 percentage points) when cited and uncited rounds of the same prompt and platform were compared, suggesting that the underlying consumer questions account for much of the raw association rather than establishing an independent benefit from citing the publisher. Captured citations appeared in 99.7% of ChatGPT's scheduled observations, 78.0% of Claude's and 4.3% of Gemini's; these are recorded citation-capture rates rather than verified search-invocation rates, so ChatGPT and Claude are treated as the main citation samples. The mix of cited sources changes with the question: brand-controlled sites supply 60.6% of citations in premium and gifting answers, while 95.9% of citation-bearing ethics and sourcing answers cite an editorial or consumer-information source.
Seven findings that matter
1. Brand visibility is not citation visibility
Cadbury leads tracked-brand inclusion in UK answers, yet Cadbury's own websites account for only three recorded citations, while Hotel Chocolat has slightly lower inclusion but 82 citations to its own website. These describe different evidence footprints, not the quality or desirability of either brand.
2. Three major retailers account for 13.8% of UK citations
Tesco, Ocado and Sainsbury's together account for 311 of 2,258 UK citation occurrences (13.8%). This is the share of three named retailers, not of all retail sources. That makes product listing accuracy commercially relevant, without proving that retail citations cause a brand to be recommended.
3. Citations are concentrated and diverse at once
The top 15 hostnames account for 41.4% of UK citations, yet the UK subset includes 322 distinct hostnames and the full export 716. A short list is worth auditing, but no single website captures the whole evidence environment.
4. Different platforms expose different sources
ChatGPT's leading UK sources are mainly retail and brand-owned sites, while Claude's include reference and aggregation sites. Because capture rates differ, these comparisons concern the exported citations rather than fully comparable retrieval behaviour.
5. Consumer intent helps account for source-brand links
Tony's Chocolonely appears in 77.2% of ChatGPT and Claude UK answers citing Ethical Consumer, against a 28.4% overall baseline, but the matched within-prompt difference is −1.0 points. The raw association is therefore not sufficient evidence of extra inclusion attributable to that source, although it does not show that citations never matter.
6. Specialist brands surface under specialist questions
Tony's Chocolonely appears in 90.5% of UK ethics and sourcing answers and Divine Chocolate in 76.2%, while Hotel Chocolat appears in 74.6% of premium and gifting answers. These describe associations with question families, not verified ethical or product claims.
7. One-time snapshots conceal variability
Across 114 UK prompt-platform combinations observed seven times, Lindt is present in all seven rounds in 38, Cadbury in 37 and Hotel Chocolat in 35. The study supports repeated, intent-controlled measurement rather than treating one answer as a stable ranking.
AI discoverability needs three distinct measures: brand inclusion, position within an answer, and the recorded source environment.
These dimensions should be read together but not conflated. The report is organised around them, moving from which brands appear, to what AI cites, to where brands and sources meet.
Why this study matters
A consumer asking an AI assistant for chocolate advice may receive several named brands, product recommendations, explanatory comparisons and source citations in a single answer. The presence of sources creates opportunities for corroboration, but it also makes the evidence chain more complex. A brand may be mentioned without being preferred, a source may be cited without being independently editorial, and a brand-owned product page may accurately establish availability or ingredients without verifying a quality claim. An engine may also provide a detailed answer while recording no citations at all.
Part 1 of this research programme investigated model-native brand recall and selection with live search disabled. Part 2 examines a separate November dataset capturing answer text alongside recorded source citations, adding an evidence layer: what AI answers say about chocolate brands and which external domains appear in their cited references.
This study investigates those distinctions in the UK chocolate category. It does not rank chocolate products by sensory quality, corporate ethics, retail sales or consumer satisfaction, and it does not directly measure actual UK consumer purchasing behaviour.
Study design
The raw export contains exactly 2,100 scheduled observations, representing 100 unique prompts × 3 platforms × 7 repeated rounds, collected from 1 to 7 November 2026. Of these, 2,094 have a successful outcome and six, all from Claude, returned no answer, while every scheduled prompt, platform and round combination is present. The dataset contains 5,119 recorded citation occurrences, 2,115 distinct literal URL strings and 716 distinct normalised hostnames.
| Design element | Captured scope |
|---|---|
| Questions | 100 fixed consumer questions |
| Platforms | ChatGPT, Gemini and Claude, accessed through the research export1 |
| Observation rounds | Seven, dated 1 to 7 November 2026 |
| Attempted observations | 2,100 (700 per platform) |
| Successful observations | 2,094 (700 ChatGPT, 700 Gemini, 694 Claude) |
| UK-explicit questions | 38, producing 798 attempts and 794 successful responses |
| Open-context questions | 62, producing 1,302 attempts and 1,300 successful responses |
| Captured citations | 5,119 overall; 2,258 UK-explicit; 2,861 open-context |
| Citation-bearing responses | 1,274 overall; 529 UK-explicit |
Two non-interchangeable geographic evidence layers
As in Part 1, 38 prompts explicitly mention the UK, Britain or British. They account for 798 attempted observations and are the primary evidence base for UK-specific findings. The remaining 62 prompts do not establish that market explicitly and are reported as open-context, and because the two groups contain different prompt content, their differences are not presented as matched causal effects of adding a geographic term.
Fixed tracked cohort
The consistent 13-brand cohort from Part 1 is Lindt, Cadbury, Green & Black's, Hotel Chocolat, Tony's Chocolonely, Divine Chocolate, Montezuma's, Galaxy, Ferrero Rocher, Thorntons, KitKat, Terry's and Booja-Booja. Retaining this cohort improves continuity across the research programme without silently enlarging the rankings.
How each measure is defined
Presence is the number of responses containing a tracked brand or explicitly assigned product alias, divided by all attempted observations in the relevant group. A response containing multiple brands contributes one occurrence to each present brand, and because the denominator includes unanswered attempts, which are separately disclosed, the reported rates reflect the scheduled study grid.
First tracked brand is the earliest textual reference in the answer to one of the 13 tracked brands. This is an order-of-appearance measure and is not described here as the first recommendation, the most preferred brand or the winner of a comparison.
Recorded citation is an individual URL reference in the export, and the same URL may be cited more than once, with each occurrence counted. A citation-bearing response is a response with one or more captured URLs, and URL and hostname counts are descriptive, since links have not all been checked for live accessibility or substantive claim support.
Cross-round persistence is measured at the prompt-platform level across the seven observations. A brand or source appearing all seven times is classified as persistently present for that question and platform, although persistence of mention is not persistence of endorsement.
1 Model labels recorded in the research export were gpt-6.1-terra (ChatGPT), gemini-4-argon (Gemini) and claude-haiku-5-5 (Claude). These are recorded export labels, taken verbatim; they have not been authenticated as provider model identifiers, and the export contains no provider response metadata or search-tool invocation records, so grounding configuration and search invocation are also unverified. Results apply to the outputs observed from 1 to 7 November 2026 and should be read as a time-bounded measurement rather than a permanent representation of current or future platform behaviour.
Before looking at sources, the first layer measures which tracked brands enter UK answers, where they sit, and how that changes by platform, consumer need, repetition and market context. These findings use all three platforms, since brand mentions were captured for every answer.
01UK brand inclusion is led by a tightly grouped set
Across 798 UK-explicit attempts, Cadbury is mentioned in 404 answers (50.6%), Lindt in 397 (49.7%), Hotel Chocolat in 380 (47.6%) and Tony's Chocolonely in 320 (40.1%). The top three are separated by only 3.0 percentage points, suggesting no overwhelming overall presence leader within this dataset.
| Rank | Tracked brand | UK answers containing brand | Presence |
|---|---|---|---|
| 1 | Cadbury | 404 / 798 | 50.6% |
| 2 | Lindt | 397 / 798 | 49.7% |
| 3 | Hotel Chocolat | 380 / 798 | 47.6% |
| 4 | Tony's Chocolonely | 320 / 798 | 40.1% |
| 5 | Montezuma's | 294 / 798 | 36.8% |
| 6 | Green & Black's | 292 / 798 | 36.6% |
| 7 | Galaxy | 244 / 798 | 30.6% |
| 8 | Divine Chocolate | 156 / 798 | 19.5% |
| 9 | Terry's | 124 / 798 | 15.5% |
| 10 | KitKat | 91 / 798 | 11.4% |
| 11 | Thorntons | 71 / 798 | 8.9% |
| 12 | Booja-Booja | 71 / 798 | 8.9% |
| 13 | Ferrero Rocher | 44 / 798 | 5.5% |
These are text-extraction counts using the brand-matching rules set out under How to reproduce and audit this study. They are not market shares and do not sum to 100%.
02Presence is not the same as being surfaced first
The first-occurrence ordering of the fixed 13-brand cohort produces a different picture. Cadbury is the earliest tracked brand in 215 of 798 UK answers (26.9%), Hotel Chocolat in 136 (17.0%), Lindt in 117 (14.7%) and Tony's Chocolonely in 99 (12.4%), while 61 answers contain no identifiable member of the tracked cohort.
Order of first textual appearance is sensitive to prose structure, introductory caveats and text placement, so it is not a positive recommendation measure. The correct next stage is to code whether each brand was affirmatively recommended, neutrally compared, explicitly discouraged or merely named.
03The same UK questions yield different brand sets by platform
Each platform received 38 UK-explicit prompts on seven occasions, or 266 attempts per platform. Yet their presence counts vary substantially, and for several brands the platform that answered mattered far more than the brand's overall rate suggests.
| Brand | ChatGPT / 266 | Gemini / 266 | Claude / 266 |
|---|---|---|---|
| Cadbury | 71 (26.7%) | 204 (76.7%) | 129 (48.5%) |
| Lindt | 118 (44.4%) | 164 (61.7%) | 115 (43.2%) |
| Hotel Chocolat | 112 (42.1%) | 202 (75.9%) | 66 (24.8%) |
| Tony's Chocolonely | 101 (38.0%) | 169 (63.5%) | 50 (18.8%) |
| Montezuma's | 31 (11.7%) | 184 (69.2%) | 79 (29.7%) |
| Green & Black's | 70 (26.3%) | 162 (60.9%) | 60 (22.6%) |
| Galaxy | 42 (15.8%) | 121 (45.5%) | 81 (30.5%) |
| Divine Chocolate | 45 (16.9%) | 69 (25.9%) | 42 (15.8%) |
Montezuma's has a 57.5 percentage-point difference between the highest and lowest platforms (69.2% on Gemini against 11.7% on ChatGPT), and Hotel Chocolat has a 51.1-point difference (75.9% on Gemini against 24.8% on Claude). These are observed contrasts, not evidence that a provider systematically prefers, promotes or penalises a brand.
Gemini's median response is 724 words, against 228 for Claude and 202 for ChatGPT, and longer answers create more opportunities for mentions. Raw mention-rate comparisons are therefore not length-normalised, and a per-answer or per-1,000-words sensitivity analysis would be needed before asserting that one model is genuinely more inclusive.
04Consumer need changes which brands surface
The 38 UK-explicit prompts were grouped into reporting themes based on their original order and wording. Theme sizes are uneven, so the figures below are within-theme rates rather than percentages of the whole dataset.
The eleven themes are a reporting taxonomy applied to the prompt set during analysis, because the export's prompt category field was not populated at collection. The prompt-to-theme mapping was reconstructed and validated against the raw export. The themes organise the questions for reporting and are not evidence of changing consumer preferences.
| UK-explicit consumer theme | Prompts | Responses | Most frequently included tracked brands |
|---|---|---|---|
| General discovery | 3 | 63 | Tony's Chocolonely 61.9%; Cadbury 58.7%; Hotel Chocolat 55.6% |
| Taste and quality | 4 | 84 | Lindt 75.0%; Cadbury 66.7%; Green & Black's 51.2% |
| Chocolate type | 4 | 84 | Lindt 92.9%; Green & Black's 70.2%; Montezuma's 52.4% |
| Flavours and fillings | 4 | 84 | Lindt 58.3%; Hotel Chocolat 57.1%; Terry's 53.6% |
| Value and retail | 3 | 63 | Cadbury 46.0%; Lindt 25.4%; Galaxy 19.0% |
| Premium and gifting | 3 | 63 | Hotel Chocolat 74.6%; Montezuma's 39.7%; Cadbury 19.0% |
| Occasions and practical use | 4 | 84 | Hotel Chocolat 76.2%; Cadbury 64.3%; Lindt 57.1% |
| Baking and use | 1 | 21 | Green & Black's 90.5%; Lindt 85.7%; Cadbury 23.8% |
| Dietary requirements | 3 | 63 | Galaxy 61.9%; Lindt 47.6%; Cadbury and Montezuma's 44.4% (joint) |
| Ethics and sourcing | 3 | 63 | Tony's Chocolonely 90.5%; Divine Chocolate 76.2%; Montezuma's 36.5% |
| Overall evaluation | 6 | 126 | Cadbury 77.8%; Hotel Chocolat 70.6%; Tony's Chocolonely 67.5% |
Tony's Chocolonely appears in 57 of 63 ethics and sourcing answers (90.5%) and Divine Chocolate in 48 of 63 (76.2%), while Hotel Chocolat appears in 47 of 63 premium and gifting answers (74.6%). These observations show how brands are associated with particular questions, and they do not independently validate sustainability claims, nutritional suitability or superior product quality.
The dietary theme warrants extra care, because some answers discuss mainstream brands as exclusions, comparisons or allergen cautions rather than positive dietary recommendations. Without a sentiment-and-eligibility coding pass, Galaxy's 61.9% in this theme should not be read as Galaxy being the best dietary option.
05Repeated observations reveal persistent brand associations
The seven-round design makes it possible to observe stability rather than relying on one answer per question. Among the 114 UK-explicit prompt-platform combinations (38 questions × 3 platforms), Lindt appears in all seven rounds for 38 combinations, Cadbury for 37, Hotel Chocolat for 35 and Tony's Chocolonely for 31.
Montezuma's shows the most intermittent pattern, appearing in some but not all rounds for 53 combinations, against only 17 where it appeared every time. High persistence means a brand repeatedly occurs in similar answer contexts, not that the model repeatedly recommends it or that the brand is a higher-quality purchase.
06Open-context questions produce a different landscape
Across the 1,302 attempts where Britain or the UK was not specified, Lindt appears in 707 responses (54.3%), Tony's Chocolonely in 444 (34.1%) and Cadbury in 247 (19.0%). Hotel Chocolat falls to 64 appearances (4.9%) and Montezuma's to eight (0.6%), against 47.6% and 36.8% respectively in the UK-explicit group.
This is an observed association between different prompt groups, not proof that simply inserting "UK" into a matched question would cause those changes. The question themes and their distribution differ across the groups, and Part 3 will retain this caveat when comparing equivalent prompts across study periods and systems.
The second layer turns to the 5,119 recorded citations, asking which websites appear, how concentrated they are, how consistently they recur and how far the platforms can be compared. ChatGPT and Claude are treated as the main citation samples, because Gemini's recorded citation capture was limited.
07The citation layer is substantial, but unevenly captured
The whole export contains 5,119 citation occurrences across 2,100 attempts. The 798 UK-explicit attempts account for 2,258 of them and the open-context attempts for 2,861, but the distribution across platforms is far from even.
| Evidence metric | ChatGPT | Gemini | Claude |
|---|---|---|---|
| Attempts, entire dataset | 700 | 700 | 700 |
| Successful responses | 700 | 700 | 694 |
| Recorded citations, entire dataset | 2,793 | 180 | 2,146 |
| Answers with recorded citations | 698 (99.7%) | 30 (4.3%) | 546 (78.0%) |
| Citations per scheduled response | 3.99 | 0.26 | 3.07 |
| Recorded citations, UK-explicit subset | 1,229 | 137 | 892 |
| Citation-bearing UK answers | 266 / 266 | 22 / 266 | 241 / 266 |
The difference cannot responsibly be interpreted as a simple "ChatGPT searches more than Gemini" finding. It may depend on differences in API tool configuration, search invocation criteria, provider citation formats and export logic, and the asymmetry is itself an operational lesson: citation reporting requires capture-method validation.
The figures in this section are therefore recorded citation-capture rates, not verified search-invocation rates. The export does not contain provider tool-use records, so it cannot distinguish answers where search was not invoked, citations that were returned but not extracted, and answers that were legitimately given without citations.
08UK answers cite retailers, brands and independent publishers
Of the 2,258 recorded citations associated with UK-explicit prompts, the most frequently cited hostnames span supermarkets, brand-owned websites, consumer-information publishers and reference sites. Tesco leads with 180 occurrences, followed by Which? with 87, Ocado with 84, Hotel Chocolat with 82 and Ethical Consumer with 63.
1. Retail and availability evidence
Tesco, Ocado, Sainsbury's and other merchants can supply products, prices, formats and UK availability. These three named retailers account for 311 UK citation occurrences (13.8%), a figure that does not cover all retail sources, and a retail listing is not automatically a third-party quality endorsement.
2. Brand-controlled primary evidence
Hotel Chocolat, Pump Street and Tony's Chocolonely-owned web properties can document ranges, ingredients and corporate claims. They should not be counted as independent corroboration of those claims.
3. Editorial and consumer-information evidence
Which?, Ethical Consumer, Saga and other publications provide comparison and interpretation. Their weight depends on the specific article's methods and editorial independence.
4. Aggregation and reference sources
Wikipedia, shopping aggregators and republishers may provide contextual information. They should not be treated as equivalent to original reporting or primary scientific evidence.
The leading hostname alone accounts for 180 of 2,258 UK citations (8.0%). Repeated references to the same site do not mean there were 180 independent sources or 180 independent endorsements.
How citation source roles change across consumer needs
To move beyond which individual websites are cited, every hostname cited in UK-explicit answers was classified into one of the four source roles above, with a fifth "unresolved" category where a site's function could not be established with confidence. The classification was applied after collection to all 322 UK-cited hostnames, using the rules below, and each of the 2,258 UK citation occurrences inherits its hostname's role.
Classification rules
Retail and commerce: sells or lists products for sale, including supermarkets, department stores, online shops, marketplaces and price-comparison sites. Brand-controlled: the website of a food or confectionery brand, maker or its parent company. Editorial and consumer information: publishers of articles, reviews or rankings, consumer organisations, NGOs and campaign scorecards, certifiers, awards bodies and trade press. Reference, aggregation and community: encyclopaedias, data and survey providers, government, health and academic sources, content republishers and community forums. Unresolved: function not determinable from the hostname and cited URL.
Roles describe what a site does, not whether it is editorially independent, accurate or claim-supporting. The classification was made by a single analyst and has not been independently double-coded. Consumer themes are the reconstructed reporting taxonomy from Finding 04.
| Source role | UK citation occurrences | Share of 2,258 | Citation-bearing UK answers citing this role | Share of 529 |
|---|---|---|---|---|
| Editorial and consumer information | 808 | 35.8% | 400 | 75.6% |
| Retail and commerce | 566 | 25.1% | 256 | 48.4% |
| Brand-controlled | 474 | 21.0% | 189 | 35.7% |
| Reference, aggregation and community | 285 | 12.6% | 194 | 36.7% |
| Unresolved | 125 | 5.5% | 106 | 20.0% |
An answer can cite more than one role, so the answer-level shares do not sum to 100%. Across all retail and commerce sources, retail accounts for 25.1% of UK citation occurrences, compared with 13.8% for the three named supermarkets alone.
| Theme | Prompts | Cited answers | Retail | Brand | Editorial | Reference | Most cited sources |
|---|---|---|---|---|---|---|---|
| Premium and gifting | 3 | 40 | 5.0% | 67.5% | 42.5% | 30.0% | Charbonnel et Walker, Hotel Chocolat, Pump Street |
| Ethics and sourcing | 3 | 49 | 46.9% | 34.7% | 95.9% | 34.7% | Ethical Consumer, Which?, PAN UK |
| Taste and quality | 4 | 54 | 25.9% | 16.7% | 88.9% | 18.5% | Saga, Tesco, Great British Food Awards |
| Dietary requirements | 3 | 45 | 55.6% | 42.2% | 77.8% | 42.2% | Tesco, Sainsbury's, Ocado |
| Baking and use | 1 | 13 | 84.6% | 23.1% | 76.9% | 15.4% | Tesco, Waitrose, Trolley |
| Flavours and fillings | 4 | 55 | 76.4% | 69.1% | 54.5% | 32.7% | Tesco, Hotel Chocolat, Joybuy |
| Value and retail | 3 | 48 | 52.1% | 2.1% | 60.4% | 41.7% | Tesco, Saga, The Independent |
Retail to Reference are answer-level shares: the percentage of citation-bearing answers in each theme citing at least one source of that role. Seven of the eleven themes are shown; the full table for all eleven is in the validation register.
The evidence mix changes markedly with the question. Premium and gifting answers draw 60.6% of their citations from brand-controlled sites, led by Charbonnel et Walker, Hotel Chocolat and Pump Street, while value and retail answers draw just 0.6% from brand sites. Ethics and sourcing answers lean on editorial and consumer-information sources, with 95.9% of citation-bearing answers citing at least one, led by Ethical Consumer, Which? and PAN UK, and taste and quality answers do the same at 88.9%.
Questions about practical product choice behave differently. Retail and commerce sources account for 54.5% of citations in the single baking prompt, 43.5% for occasions and practical use, and 41.9% for dietary requirements, where Tesco, Sainsbury's and Ocado product pages lead. The platforms also differ in their mix: retail and brand-controlled sites account for 63.4% of ChatGPT's UK citations, while editorial and reference sources account for 66.0% of Claude's.
What this could mean for brand evidence planning
These are observed associations between question types and the kinds of sources cited, not evidence that any source role causes a brand to be included or recommended. They do suggest that the evidence a brand needs to keep accurate depends on the questions it wants to be relevant for. For dietary and practical questions, accurate retail product information, including ingredients, allergens and formats, sits directly in the cited evidence, while for ethical questions the cited evidence is largely independent assessment, so clear and independently assessable sourcing claims matter more than a brand's own statements. Whether improving those sources changes AI answers would need a controlled before-and-after test.
09Citation concentration is not source consensus
The UK-explicit sample contains 2,258 recorded citation occurrences from 322 hostnames. The 15 most frequently cited hosts contribute 934 occurrences (41.4%) and the leading 50 contribute 1,505 (66.7%), while at the other end 125 hostnames are cited only once. Across the full study, 716 hostnames are present and 286 appear only once.
A relatively compact group of websites accounts for a substantial fraction of observed references, but the evidence environment is not limited to those sites. A concentrated citation record is also not the same as a consensus of independent publishers, because repeated URLs, syndication, model-specific citation formatting and brand or retailer ownership all affect what the totals mean.
Why a source-role audit is more useful than a backlink count
A source's value is conditional on the consumer question. For a request about vegan ingredients, an accurately maintained allergen or product page can be more directly relevant than a generic "best chocolate" listicle, while for an ethical-sourcing query, independent evidence about sourcing practices may carry more interpretive value than a brand's own promise. Citation counts alone do not resolve whether a source is current, accurate, independent, UK-relevant or supportive of the specific claim.
10Frequently cited sources are not necessarily stable from round to round
Across the 114 UK-explicit prompt-platform combinations, a source hostname can recur in every round, some rounds or no rounds. Even the most frequently cited websites were cited in all seven rounds for only a handful of combinations.
| Hostname | UK responses citing host | Groups cited all 7 rounds | Groups cited intermittently |
|---|---|---|---|
| tesco.com | 107 | 6 | 19 |
| which.co.uk | 80 | 2 | 25 |
| ethicalconsumer.org | 58 | 3 | 13 |
| hotelchocolat.com | 77 | 5 | 14 |
| en.wikipedia.org | 53 | 0 | 25 |
| ocado.com | 46 | 1 | 14 |
These are counts of citation-bearing answers, not citation occurrences, since an answer may reference the same hostname more than once. The distinction matters because headline occurrence totals can be inflated by multiple citations from one site within a single response.
For monitoring, source persistence should be reported alongside source share. A campaign cannot infer durable "citation ownership" from one successful answer.
11Platform citation differences are both substantive and a measurement warning
ChatGPT records citations in all 266 UK answers, Claude in 241 and Gemini in only 22. Among recorded UK citations, the platforms also lean on visibly different kinds of source.
| Platform | Most frequently recorded hostnames in UK-explicit answers |
|---|---|
| ChatGPT | Tesco (177), Ocado (84), Hotel Chocolat (80) |
| Claude | Wikipedia (54), Joybuy (48), TasteAtlas (40) |
| Gemini | Reddit (13), BBC Good Food (8), YouGov (7) |
These differences can reflect different search and answer pathways or different export and citation behaviour, and the raw data cannot separate the two. ChatGPT and Claude are therefore treated as the main observed citation samples, while Gemini's brand findings are reported in full and its citation capture is explicitly labelled as limited.
A check of the response text clarifies part of the picture. None of Gemini's 700 answers contains a URL or numbered citation marker, so no in-text citations were lost, but 54 of its uncited answers name sources such as Which? or YouGov without links, and whether Gemini returned sources in metadata that the export did not capture cannot be determined. Of ChatGPT's 2,793 cited URLs, 2,538 carry the utm_source=openai tag that OpenAI's web search adds to cited links, consistent with that tool having been used, and 7 Claude answers (3 UK-explicit) contain URLs in their text that were not captured as citations.
No cross-platform claim about citation efficiency, source preference or grounding quality should be made until provider tool-use logs and citation mapping are checked. Long Gemini responses with no recorded URLs are not evidence that they were ungrounded, and equally the data cannot establish that they were grounded.
The third layer joins the two records together, asking how a brand's own website relates to its visibility and whether citing a particular publisher is linked to a brand appearing. It contains the study's most important methodological result.
12Brand-owned evidence and brand inclusion have different distributions
The counts below compare how often each brand is mentioned with how often its identified brand-associated hostnames are cited in UK-explicit answers, as shown in Fig. 1 at the top of this report. They are deliberately not called exhaustive corporate-ownership totals, because subsidiaries, parent-company pages, alternative country domains and retailer-hosted official shops would need a full ownership registry to be included.
| Tracked brand | Responses mentioning brand / 798 | Identified brand-associated citation occurrences | Responses with both brand and owned-domain citation |
|---|---|---|---|
| Cadbury | 404 | 3 | 3 |
| Lindt | 397 | 26 | 24 |
| Hotel Chocolat | 380 | 82 | 77 |
| Tony's Chocolonely | 320 | 34 | 32 |
| Divine Chocolate | 156 | 20 | 20 |
The sharpest contrast is between Cadbury and Hotel Chocolat, and it survives a full audit of the cited hostnames. Every UK-cited host was checked for direct brand domains, parent-company domains and brand-named subdomains: Cadbury's own sites account for exactly three citations (cadbury.co.uk twice and shop.cadbury.co.uk once), no domain belonging to its parent company Mondelez International was cited anywhere in the dataset, and Hotel Chocolat's 82 citations all point to hotelchocolat.com. A high number of brand mentions does not require an equally high number of citations to that brand's site, and while citation of a brand's official website and mention of that brand naturally occur together, neither condition proves that the website caused a recommendation.
Brand-owned source visibility is not inherently weaker evidence. Official product pages may be the best primary source for ingredients, sizes or terms of sale, but they are not independent endorsements of subjective quality, ethical standing or superiority, so brands should treat accurate owned properties and reliable external coverage as complementary evidence rather than interchangeable signals.
| Ownership tier (UK citations) | Cadbury | Hotel Chocolat |
|---|---|---|
| Direct brand-owned domains | 3 (cadbury.co.uk 2, shop.cadbury.co.uk 1) | 82 (hotelchocolat.com) |
| Parent-company domains | 0 (no Mondelez domain cited) | 0 |
| Retailer or comparison pages for the brand's products | 9 (Ocado 4, Sainsbury's 2, Tesco 2, Trolley 1) | Not separately counted |
| Independent third-party pages naming the brand | Not counted as owned evidence | Not counted as owned evidence |
Retailer pages are not counted as brand-owned evidence: they are controlled by the retailer even when they describe the brand's products. Ownership was checked at domain level; individual page authorship was not inspected.
13Source-brand co-occurrence needs intent controls
Because each exported observation contains both the answer text and its citation list, the study can ask a direct question: when a specified domain is cited, how often does a specified brand appear anywhere in that answer? This is more informative than separate brand and domain rankings, but it is not proof that the cited page supports that brand, and it is not proof of causality.
In the UK-explicit ChatGPT and Claude subset, Tony's Chocolonely appears in 44 of 57 answers citing Ethical Consumer (77.2%), against a 28.4% baseline for the same two platforms. Divine Chocolate appears in 61.4% of those answers against a 16.4% baseline, so the raw contrasts are large.
Those baselines are the overall inclusion rates across all 532 scheduled ChatGPT and Claude UK observations, not the rate in answers that did not cite Ethical Consumer. The raw contrast is therefore a comparison with the overall platform baseline rather than a direct cited-versus-uncited comparison.
Holding the question constant
The comparison was therefore repeated within identical prompt-and-platform groups, keeping only groups where some rounds cited the source and others did not. This reduces confounding by prompt wording and model identity, although unobserved differences can remain.
How the matched comparison works
Matched group: one UK-explicit prompt on one platform, observed across seven rounds. Qualifying group: a group in which the source was cited in at least one round and not cited in at least one other. Within each qualifying group, brand inclusion in the rounds that cited the source is compared with inclusion in the rounds that did not, and the group-level differences are combined into one weighted figure.
Reproduced from the raw export: every value in the table below is calculated directly from it. The analysis covers ChatGPT and Claude, uses all scheduled observations, and combines groups with the weight ncited × nuncited ÷ (ncited + nuncited), the standard Mantel-Haenszel weight for a difference in proportions. No qualifying groups were excluded.
What sits behind the −1.0: Ethical Consumer was cited in some but not all rounds of 12 prompt-platform groups (36 cited observations). In 7 of those 12, Tony's Chocolonely's inclusion did not vary at all, appearing in every round or none whether or not the site was cited, so those groups add no evidence either way. Excluding no-answer rounds leaves every value unchanged except the two Which? pairs, which move by under one point. The matched figure is an observational comparison of 12 groups, not a randomised estimate of a citation effect.
| Cited host | Brand measured | Matched prompt-platform groups | Weighted within-group inclusion difference |
|---|---|---|---|
| ethicalconsumer.org | Tony's Chocolonely | 12 | −1.0 pp |
| ethicalconsumer.org | Divine Chocolate | 12 | +12.7 pp |
| tesco.com | Cadbury | 18 | −7.9 pp |
| tesco.com | Lindt | 18 | +3.0 pp |
| which.co.uk | Tony's Chocolonely | 22 | +6.4 pp |
| which.co.uk | Divine Chocolate | 22 | +9.6 pp |
| hotelchocolat.com | Hotel Chocolat | 14 | +88.5 pp* |
* Structurally unsurprising, because mentioning a brand and citing its own website are often part of the same answer-construction step. These are descriptive, weighted differences, not randomised estimates, p-values or independent observations.
The most important result is that the apparent association between Ethical Consumer citations and Tony's Chocolonely inclusion was not reproduced as an inclusion advantage in the matched prompt-platform comparison. This suggests that the underlying consumer questions account for much of the raw association, rather than establishing an independent benefit attributable to citing the publisher.
It does not show that Ethical Consumer had no influence, and a −1.0-point difference across 12 groups is not evidence that citations never affect brand inclusion. Divine Chocolate's matched difference of +12.7 points shows that a smaller association can remain after matching, so prompt intent should be controlled before attaching strategic meaning to any citation-brand pair.
A more precise model of citation value
Part 2 treats answer-side and source-side measurements as related but separate dimensions, not as steps in a causal pipeline. Some are measured in this study, while others require validation that has not yet been performed, so assigning an "authority score" purely from citation frequency would be misleading.
Answer-side
- Brand inclusion (measured)
- Earliest tracked brand (measured, automated)
- Positive recommendation (not yet validated)
Source-side
- Recorded citation exposure (measured)
- Source ownership and role (classified, single analyst)
- Citation recurrence (measured)
- Claim-level support (not yet audited)
Relationship
- Co-occurrence of brands and sources (measured)
- Matched prompt-platform comparisons (measured, reproduced)
- Observational associations, not causation
What marketing directors and brand teams should do differently
The findings do not provide an experimental optimisation recipe. They do provide a more defensible way to prioritise a brand's evidence environment and to test whether changes are reflected in AI answers, and each action below sets out what the data showed, what to do and what the study has not proven.
1. Treat retailer product pages as part of the factual evidence layer
2. Make official brand properties definitive for the facts they control
3. Build external evidence around specific consumer questions
4. Measure each AI platform as a separate selection and evidence environment
5. Track stability and validate whether evidence is actually used
A practical measurement scorecard
The study's methods translate into a scorecard any brand team can apply to its own category. Each metric answers a different question, and each needs a minimum standard of method before its results can be relied on.
| Metric | Question answered | Minimum defensible method |
|---|---|---|
| Brand inclusion rate | Did the brand enter the answer? | Validated alias matching with a fixed prompt and platform denominator |
| First-mentioned tracked brand | Who appeared earliest? | Position coding, explicitly not endorsement |
| Positive recommendation rate | Was the brand actually suggested? | Human-coded or validated classifier with polarity rules |
| Citation-bearing response rate | Was any URL captured? | Verified provider citation extraction and tool metadata |
| Host and URL frequency | Which sources recur? | Normalised hostnames, de-duplicated at a chosen level |
| Source ownership and role | Who controls the evidence? | Page-level ownership and editorial review |
| Brand-source co-occurrence | Which brands appear with which cited sources? | Response-level join, stratified by prompt and platform |
| Claim-level citation support | Does the linked page substantiate the claim? | Review of cited passage, date, region and context |
| Seven-round persistence | Is the signal repeatable? | Identical prompt and model grouping with variance tracking |
Measure selection and evidence separately, then study their relationship at the level of the actual consumer question.
Key findings at a glance
- Cadbury was mentioned in 50.6% of UK observations, but Cadbury's own websites were cited only three times.
- Hotel Chocolat appeared in 47.6% of UK observations, and its own website was cited 82 times.
- Tony's Chocolonely appeared in 77.2% of answers citing Ethical Consumer against a 28.4% overall baseline, but showed no matched inclusion advantage (−1.0 points).
- Cadbury, Lindt and Hotel Chocolat led UK inclusion within 3.0 percentage points of each other.
- Cadbury was the earliest tracked brand in 26.9% of UK answers, against 14.7% for Lindt.
- Montezuma's ranged from 11.7% on ChatGPT to 69.2% on Gemini, a 57.5-point platform gap.
- Tony's Chocolonely appeared in 90.5% of UK ethics and sourcing answers.
- 5,119 citations were recorded, with ChatGPT and Claude carrying almost all of them.
- Tesco, Ocado and Sainsbury's together accounted for 13.8% of UK citations, and all retail and commerce sources for 25.1%.
- Brand-controlled sites supplied 60.6% of citations in premium and gifting answers, while 95.9% of citation-bearing ethics answers cited an editorial or consumer-information source.
- The top 15 hostnames accounted for 41.4% of UK citations, from 322 in total.
- Tesco was cited in all seven rounds for only 6 of 114 prompt-platform combinations.
- Hotel Chocolat fell from 47.6% in UK-explicit answers to 4.9% when no market was named.
Limitations and interpretation rules
Grounding comparability unverified
The export records citations but does not by itself prove that identical web search configurations were enabled and invoked for all three platforms on all prompts. Gemini's 4.3% citation-bearing rate is a validation priority.
Model-label provenance
The model labels are taken verbatim from the export. They are experimental metadata, not independently authenticated public product names.
Presence is not endorsement
Brand inclusion and earliest-position calculations use automated name and alias extraction, and can include neutral references, negatives and ambiguous product mentions. Final editorial rankings require manual validation.
No claim-level content verification
The 5,119 recorded references have not each been crawled to validate page content, publication date, independence or evidence quality. Nor has each been checked against the specific claim it supports.
Ownership checked at domain level
Brand and parent-company ownership was audited for every UK-cited hostname, but individual page authorship was not inspected. Retailer pages describing a brand's products are counted as retail, not as brand-owned evidence.
Occurrences are not independent observations
The same prompt, site and URL may recur across rounds. Citation counts are not a count of independent recommendations or endorsements.
Different response lengths
Gemini's much longer answers increase the opportunity to name brands. The report describes raw inclusion, not length-adjusted competitive preference.
Geographic groups differ by prompt mix
UK-explicit and open-context groups were not randomised matched A/B tests. Differences between them are observed, not causal.
Small samples in places
Some intent categories contain only one to four UK-explicit prompts, and the matched comparisons rest on 12 to 22 groups. These results are descriptive and should not be generalised.
Time-bounded and not causal
All findings describe outputs observed from 1 to 7 November 2026. The study cannot determine why a model included a brand or whether any specific webpage caused it.
Supported by this study
"Cadbury was mentioned in 50.6% of scheduled UK-explicit observations, but Cadbury's own websites were cited three times." "In the matched comparison, rounds citing Ethical Consumer showed no inclusion advantage for Tony's Chocolonely."
Not supported without further evidence
"AI recommends Cadbury more than any other brand" would need recommendation-polarity coding. "Hotel Chocolat's website caused its visibility" would need causal evidence the study does not collect.
How to reproduce and audit this study
The analysis starts from the exported observation table, in which each row is one scheduled observation. The recorded fields are batch label, day number, observation timestamp, engine, model label, prompt category, prompt text, capture outcome, citation count, cited URLs, response text and response identifier; the prompt category field was not populated. The 100 prompts are repeated for each of three platforms over seven rounds, and the analysis uses the fixed 13-brand cohort from Part 1.
Brand mentions are counted in the answer text, excluding web links and link labels. Named product lines count towards their brand only when they clearly belong to it (for example Dairy Milk, Wispa, Crunchie, Creme Egg, Double Decker, Twirl and Bournville for Cadbury, and Lindor for Lindt), while generic words such as "caramel" or "chocolate orange" are not counted. The earliest tracked brand is the first such match in the answer text.
A counted citation is a URL occurrence, and the normalised hostname is the lowercase URL host with www. removed, while other subdomains remain distinct. Brand-host co-occurrence is assessed within the same response, and the within-prompt comparison holds the prompt string and platform fixed across rounds.
Rates use scheduled attempts as denominators unless stated otherwise, and percentage-point differences are arithmetic differences between proportions. The matched citation contrasts are weighted within prompt-platform groups that contain both cited and uncited rounds, and they do not establish causal effects or statistical significance.
What was captured, calculated and reconstructed
| Element | Status |
|---|---|
| Response text, cited URLs, citation counts, capture outcome, model label | Captured directly during collection |
| Brand presence, earliest tracked brand, hostnames, citation counts, persistence | Calculated from the raw export |
| Observation, citation, hostname and co-occurrence totals | Recalculated from the raw export: all published totals reproduced exactly |
| Consumer themes | Reconstructed; validated against the raw export |
| Brand matching | Rules documented above; checked by reading every case where alternative rules disagreed |
| Brand-owned hostnames | Audited at domain level for every UK-cited host |
| Source roles | Classified after collection by a single analyst; 5.5% of UK citations unresolved |
| Matched comparisons | Reproduced exactly from the raw export |
| Actual model identifiers, search-tool configuration and invocation | Not verifiable from the export |
| Claim-level support of cited pages, recommendation polarity | Not yet undertaken |
Validation status
The checks above were completed against the raw export on 18 November 2026. The following quality-control steps remain open, and they are listed here rather than claimed as complete.
- Audit a sample of claim and URL pairs for source relevance, accuracy and UK applicability.
- Verify the ChatGPT, Gemini and Claude model identifiers and search-tool configuration from the collection system or provider logs.
- Inspect Gemini grounding metadata, if retained, to establish why 670 of its 700 answers carry no captured citation.
- Double-code the source-role classification and the brand coding on a stratified sample.
- Validate recommendation and polarity coding manually or independently before publishing first-recommendation measures.
- Retain the codebook and analysis script with version identifiers.
What comes next: Part 3
Part 3 will compare Part 1's September model-native baseline with this November search-grounded dataset. The comparison has to be explicit about three concurrent changes: live search and citation capture, tested model versions, and collection period. Unless the same models are separately tested with search switched on and off, the observed changes cannot be assigned uniquely to grounding.
Model-native recommendations
What models recall, associate and recommend from learned knowledge alone. Published September 2026.
Citations and brand visibility
Brand presence alongside recorded source citations. This is the report you are reading.
Model-native versus search-grounded
A comparison of the two datasets on the same 38 UK-explicit prompts. It will separate real change from normal variation.
Part 3 will concentrate on brand presence and first-position changes for the fixed 13-brand cohort on the same 38 UK-explicit prompts, followed by intent associations, ranking and stability. The citation analysis in Part 2 will serve as additional context rather than a metric imposed retroactively on Part 1.
Definitions
- Scheduled observation
- One planned prompt, platform and round combination, whether or not an answer was returned. There are 2,100 in total and 798 UK-explicit.
- Successful response
- A scheduled observation that returned an answer. There are 2,094 in total and 794 UK-explicit.
- UK-explicit
- A prompt containing the terms UK, Britain or British. These 38 prompts form the evidence base for UK-specific findings.
- Open-context
- A prompt with none of those explicit markers. These 62 prompts show which brands emerge before a market is named.
- Presence
- A tracked brand appears at least once in the answer under the stated matching rules. It does not indicate that the brand was recommended.
- First tracked brand
- The first named member of the fixed tracked cohort in the response text. It is not a recommendation indicator.
- Citation occurrence
- One captured URL reference, including repeated references. The same URL cited twice counts twice.
- Citation-bearing response
- A response with at least one captured citation URL. Responses with no captured URL are counted as non-citation-bearing.
- Distinct URL
- A unique literal URL string in the export. There are 2,115 across the full dataset.
- Normalised hostname
- The lowercase URL host with the www. prefix removed. Different subdomains remain separate.
- Response-level source reach
- The number of responses citing a hostname at least once, as in Finding 10. Unlike citation occurrences, repeated citations within one answer count once.
- Persistence
- Repeated inclusion of a brand or source across seven rounds for a fixed question and platform. Persistence of mention is not persistence of endorsement.
- Matched comparison
- A comparison made only within groups sharing the same prompt and platform, where some rounds cited a source and others did not. It reduces, but does not remove, confounding.
Data source: Netsleek Research's raw export for Part 2, containing observations dated 1 to 7 November 2026, with derived tables for UK brand coding, normalised citation hosts, brand-source co-occurrence and within-prompt associations. No third-party product claim or publication was treated as independently verified simply because a URL appeared in a model's output.
Frequently asked questions
Which chocolate brands appear most in UK AI answers in the Part 2 citation study?
In Part 2 of Netsleek's UK Chocolate AI Visibility Index 2026, Cadbury was mentioned in 50.6% of 798 scheduled UK-explicit observations, Lindt in 49.7% and Hotel Chocolat in 47.6%. These are mention rates rather than recommendation rates, and the three leaders were separated by only 3.0 percentage points.
Does a brand need its own website to be cited to appear in AI answers?
Not in this data. Cadbury was mentioned in 50.6% of UK observations while Cadbury's own websites were cited only three times, whereas Hotel Chocolat appeared in 47.6% and its own website was cited 82 times, so brand visibility and owned-source visibility behaved as separate outcomes.
Which websites do AI platforms cite for UK chocolate questions?
Of 2,258 recorded citations in UK-explicit answers, the most frequent hostnames were Tesco (180), Which? (87), Ocado (84), Hotel Chocolat (82) and Ethical Consumer (63). Tesco, Ocado and Sainsbury's together accounted for 13.8% of UK citations.
If a publisher is cited, does that make AI more likely to mention a particular brand?
Not necessarily. Tony's Chocolonely appeared in 77.2% of answers citing Ethical Consumer against a 28.4% baseline, but no inclusion advantage appeared when cited and uncited rounds of the same prompt and platform were compared. The underlying ethical-sourcing questions appear to account for much of the raw link, although this does not prove that citations never matter.
Do AI platforms cite different kinds of sources for different questions?
Yes, in this data the mix changed with the question. Brand-controlled websites supplied 60.6% of citations in UK premium and gifting answers, retail product pages led dietary, baking and occasion questions, and 95.9% of citation-bearing ethics and sourcing answers cited at least one editorial or consumer-information source such as Ethical Consumer or Which?.
Do ChatGPT, Gemini and Claude cite sources equally?
Not in the recorded data: captured citations appeared in 99.7% of ChatGPT's scheduled observations, 78.0% of Claude's and 4.3% of Gemini's. That gap may reflect differences in tool configuration and citation capture as well as search behaviour, so ChatGPT and Claude are treated as the main citation samples.
How is Part 2 different from Part 1?
Part 1 measured model-native answers in September 2026 with live search switched off, while Part 2 captured answers alongside recorded source citations in November 2026. The model versions and collection period also differ, so Part 3 will compare the two datasets while keeping those changes explicit.
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.
Netsleek Research is Netsleek's programme of original, first-party research into how AI search engines and AI models discover, evaluate and recommend brands. The UK Chocolate AI Visibility Index 2026 is part of the wider UK AI Search Visibility Index series, which began with The UK Accounting Software AI Visibility Index 2026 and continues from Part 1 of this index.
For methodology questions, media enquiries or to discuss what this dataset means for your brand, contact Netsleek. You can also explore the AI Search & Brand Discoverability Audit to see how AI models currently represent your own brand.