Methodology
What the index measures
The UK AI Visibility Index is a precise public reference for UK brand visibility across the leading generative AI assistants.
We measure how often brands are recommended by consumer AI models when users ask for product or service recommendations. By submitting thousands of naturally phrased questions across various categories, we quantify the precise visibility of brands.
The question panels
The index currently tracks 20 categories. Each category contains between 3 and 4 specific topics. We use a panel of 50 prompts per category, resulting in a total of 1000 questions across the index.
These questions simulate natural UK search demand, encompassing product subtypes, price ranges, size modifiers, costs, life stages, and specific places. All prompts are formulated naturally and never contain brand names. The question sets are versioned and effective from a stated week, never changing mid-week.
| Category | Topics | Prompts |
|---|---|---|
| Banking and Current Accounts | Current accounts, Savings, Business banking | 50 |
| Investing and Wealth | Investment platforms, Fund managers, Investment trusts | 50 |
| Mortgages and Personal Finance | Mortgage brokers, Personal loans, Retail finance | 50 |
| Business Finance and Payments | Asset finance, Spend management, Payment providers | 50 |
| Cars and Motoring | Used car retailers, Car buying platforms, New car brands | 50 |
| Fashion and Clothing | Womenswear, Menswear, Occasionwear | 50 |
| Sports and Outdoor | Sports retailers, Running, Fitness tech | 50 |
| Beauty and Personal Care | Beauty devices, Haircare, Oral care | 50 |
| Home and Furniture | Bedding, Sofas and furniture, Beds and mattresses | 50 |
| Home Improvement and DIY | Bathrooms, Radiators, Flooring, Paint and decorating | 50 |
| Heating and Energy | Boilers, Heat pumps, Commercial heating, Energy suppliers | 50 |
| Household Appliances and Audio | Floorcare, Kitchen appliances, Hi-fi | 50 |
| Travel and Holidays | Cruise, Adventure travel, Ski, UK holiday parks | 50 |
| Eating Out, Hotels and Gifts | London hotels, Restaurant groups, Gifts and experiences | 50 |
| Food and Drink | Confectionery, Dairy, Tea and coffee | 50 |
| Telecoms and Broadband | Mobile networks, Broadband providers, Business mobile | 50 |
| Business Software and IT | IT resellers, Workforce software, E-commerce platforms | 50 |
| Construction and Building Materials | Construction chemicals, Aggregates, Steel, Plasterboard and insulation | 50 |
| Industrial Supply and Manufacturing | Chemical distribution, Industrial materials, Safety and PPE | 50 |
| Professional and Business Services | Construction consultancy, Recruitment, Facilities and catering | 50 |
The assistants and their models
Measurements are made through the assistants' APIs using their consumer default models with web search on. We track the models exactly as they are exposed to the public in the UK, without using consumer app automation or scraping.
Default models are checked monthly. When an assistant upgrades its default model, we require a 2-week overlap period before transitioning.
| Assistant | Current Model | Adopted From |
|---|---|---|
| openai/gpt-5.6-luna | 2026-W40 | |
| google/gemini-3.6-flash | 2026-W44 | |
| perplexity/sonar | Not supplied by source | |
| anthropic/claude-sonnet-5 | 2026-W40 |
Model Changes Log
| Week | Engine | From | To |
|---|---|---|---|
| 2026-W44 | google/gemini-3.7-flash | google/gemini-3.6-flash | |
| 2026-W40 | openai/gpt-5.2 | openai/gpt-5.6-luna | |
| 2026-W40 | anthropic/claude-opus-4.6 | anthropic/claude-sonnet-5 | |
| 2026-W40 | google/gemini-2.5-flash | google/gemini-3.7-flash |
How answers are collected
Every panel question is sent to the measured assistant APIs. The runs are grounded in the UK locale. Collection and publication dates are shown from the stored edition; retries, model overlap and retrospective collection can differ from the usual weekly schedule.
If an assistant fails to answer a question, we retry exactly 1 time. Completion is recorded per assistant by LLMLens. The public snapshot exposes aggregate completion totals for each category.
How brands are detected
No brand is pre-selected. Brands are detected dynamically from the raw text answers using our proprietary LLMLens technology. The detection identifies direct mentions, position in the list and domain citations.
Brand aliases are merged, accounting for legal suffixes (e.g., "Ltd", "Inc") and punctuation. Domains and favicons are inferred dynamically based on the citations provided in the assistant's response. Nothing is pre-registered.
"A brand is in the index because an assistant said its name."
Metrics formulas & worked example
Our core metrics provide a holistic view of brand presence:
- Visibility: Answers mentioning the brand divided by completed answers within the rolling window (4 weeks). Each answer counts once, even if it names the brand repeatedly.
- Visibility (This Week): The same formula applied exclusively to the latest week's data.
- Citation Share: Number of answers where the brand's domain was cited as a source, divided by total completed answers.
- Entity groups: Each named entity is grouped from its recorded type as a brand, comparison site, publisher, public body or other. Category leaderboards open on the brand view; every other group stays available from the filter, and no rows are removed from the underlying measurements.
- Brand visibility is the share of measured assistant answers naming a brand, using the published category answer window.
- Named first: Answers in which the brand was the first brand named, divided by all completed primary answers in the category for that week, including answers naming no brand. The supplied share, numerator and denominator must be present together; missing or ambiguous positions mean the measure is omitted, not zero. This weekly measure is distinct from rolling visibility and is never estimated from average position.
- Average Position: The mean 1-based index position of the brand within the answers that mention it.
- Movement: The rank difference compared to the previous published week.
- Assistants column: Coloured icons identify assistants that named the brand in the measured window; faded icons indicate no recorded mention. This is presence across 4 assistants, not a rank or an additional visibility score.
- Recorded mentions: A source-provided mention total, not a distinct-answer count. It may exceed the number of answers; visibility remains the published distinct-answer percentage.
Published sample and illustrative threshold
For Banking and Current Accounts, the baseline is 50 questions × 4 assistants × 4 weeks = 800 answers before model overlap and incomplete runs. That baseline assumes 200 answers per week. The published rolling-window total is 900 answers; this is the denominator, not the baseline estimate. The published total differs by 100 answers. Parallel model readings, including Gemini shadow-model collection around transitions, can add answers; the snapshot's aggregate does not separately attribute every extra answer to a model. Illustratively, 18 distinct answers naming a brand out of 900 gives 18 ÷ 900 = 2.0%.
Ranking thresholds
Ranks use a rolling methodology to smooth out week-to-week API volatility. Brands are ordered primarily by visibility, then by citation share, and finally by total mentions as a tie-breaker.
To be published on the leaderboard, a brand must achieve at least 2% visibility and meet our 90% detection confidence threshold. These thresholds apply uniformly across both category and topic boards.
Quality guards
Data integrity is paramount. If a week's run fails our automated quality guards, the category is placed on hold. When a category is held, we carry over the last known good data and display a public notice. If a category experiences 2 consecutive holds, it is escalated for manual review. Other categories continue to publish normally.
The most recent successful edition remains visible until a new one passes quality checks and is published.
Corrections & withdrawals
Once a week is published, its snapshot is cryptographically hashed and becomes immutable. If a critical flaw is discovered post-publication, we withdraw the affected week and publish a corrected snapshot.
This process generates a transparent audit trail. Corrections happen strictly via upstream processes; we do not perform silent local database overwrites.
A corrected edition shows its latest publication date, while retaining the original collection date when available.
Limitations
The index has several intentional limitations:
- Sample Size: It represents a bounded sample of prompts per week, not exhaustive real-world traffic.
- Non-Determinism: Generative models are inherently non-deterministic. The rolling window mitigates but does not eliminate noise.
- Locale: The current index is strictly UK English and evaluated from UK IP spaces.
- Organic Only: We do not track paid placements or brand-supplied input streams.
Archives, topics, assistants and comparisons
Quarterly reports are editorially approved readings of pinned stored evidence, not a new collection process. Quarter membership follows the weekly edition end date, not a later republish date. Overlapping rolling answer windows are never summed into a quarter-wide answer count. Findings require traceable metric references; quoted excerpts are verbatim and attributed. Missing evidence keeps an issue in private preview rather than filling seven findings with estimates.
Report source geography is a labelled TLD/known-domain heuristic, not verified publisher domicile: .uk domains and the configured BBC, Guardian, FT and Reuters domains are treated as UK; other valid domains are labelled non-UK heuristic. International publishers can be misclassified. Source shifts compare only domains recorded in both endpoint windows. A model-release step or source change does not establish causation.
The complete report PDF is available only by requested email delivery. This one-off request does not subscribe the reader to marketing without separate consent. Private editorial previews are unlisted, short-lived and noindex; no preview measurements enter public sitemaps or normal report pages.
Optional source research lists citation counts, the share of citations listed in that view, associated brands and the recorded publisher/brand classification. Unknown or conflicting classifications remain unknown. Sitewide source totals use published categories only, excluding held categories; category views identify the evidence edition. “Brands lifted” is a recorded association in cited answers, not proof that a source caused a recommendation.
Consensus is the supplied agreement score, with disagreements ordered by the spread of recorded assistant values where at least two exist. Concentration is the supplied top-three share; volatility is the supplied top-ten reshuffle proportion. Missing values are omitted, not treated as zero, and none replaces the weekly visibility score. Recorded model slugs are displayed only when supplied.
Named and cited answer counts are shown separately. A named-but-not-cited count requires an explicit upstream measure: subtracting independent named and cited totals would assume an overlap that may not exist. Public brand analysis identifies topic leaders and other named leaders, not unmeasured question-level absence. Brand-focused source analysis is available only to verified owners; source citation totals remain labelled by category and assistant rather than being attributed solely to that brand.
Question pages read a separate weekly questions document. Search indexing requires at least three distinct brands with positive counts across at least two answering assistants, subject to a default-on category control. Other clusters remain noindex and appear within their topic page. Missing weekly observations remain gaps in the four-week history; absence of a document is not a zero result.
Question pages are free to read. Every brand each assistant named is shown, with no sign-up. A brand's position in an answer is shown only when it was recorded. Editorial framing is maintained in per-category configuration, with no prose generated at runtime; questions, variants, counts, sources and labelled excerpts come from the separate document.
Weekly archives read the currently stored published edition for each week. Movers compare ranks with the preceding stored published week; entrants and exits describe inclusion in those published leaderboards, not all mentions across the entire answer panel. No changes are inferred when either category is held, awaiting or absent. Explicit brand aliases are respected when comparing identities.
Displayed ranks count down the published list: 1, 2, 3 with no gaps. When reviewed name variants of one brand (for example M&S and Marks & Spencer) are shown as a single row, the duplicate row is removed, so the original rank numbers can skip. We show the position in the list instead; the underlying measurements are unchanged.
Topic pages use their published topic tables. Assistant pages rank recorded per-assistant visibility and preserve ties; a missing measurement is not zero. Held categories identify the edition supplying their carried-forward evidence. Source and consensus shifts appear only when both observations provide the supporting fields.
Comparison URLs use a single alphabetical slug order. Search indexing requires both brands to exceed 2% visibility in the same published category and positive named-together evidence there. Other valid pairs remain available as noindex views. Shared-category counts are labelled and are not combined into a new visibility score.
An unanswered or missing question is not treated as a measured absence.
Guides
Longer explanations of these measures, each with live examples from the current edition:
Quarterly memory research
Memory research is separate from the weekly retrieval-enabled Index. It samples API models, not consumer apps, for ChatGPT, Claude and Gemini only. Read the sample size (N), collection dates, model slugs, method notes and intervals alongside every result; intervals describe uncertainty, not a guaranteed ranking.
Retrospective readings retain their real collection date. Memory and retrieval are compared explicitly and are never blended into the weekly score. Memory results will appear only when the separate document is available.
Powered by LLMLens
The UK AI Visibility Index is powered by LLMLens, the AI search intelligence platform developed by Flaunt Digital. LLMLens allows brands to track their share of voice across generative AI models with unparalleled precision.
Our enterprise clients across retail, finance, B2B, and travel rely on the exact same detection algorithms and scoring mechanisms used to build this public index.
If you want to track bespoke keywords, competitors, or specific global locales, you can request a custom report.
Changelog
Removed brand claims and sign-in requirements for brand evidence. Question details are free when measured question data exists. Removed incomplete quote fragments, clarified repeated mentions, labelled assistant model changes, and added comparison and weekly archive summaries. Displayed ranking numbers now run consecutively after duplicate brand identities are combined; original measurements and weekly collection are unchanged.
Reworked discovery pages around measured changes and comparisons, removed visible citations and empty research routes, and made archive URLs lowercase. Search eligibility now requires substantive distinct evidence, restrained link density and natural internal discovery. Comparisons additionally require twenty shared answers and both brands in a shared category's top twenty. Weekly collection and stored measurements are unchanged.
Added the quarterly research report builder with pinned evidence, editorial publication checks, accessible inline charts, separate memory research, private previews and email-only PDF delivery. The first issue remains unpublished until its actual editorial and research inputs are approved.
Added optional cited-source views by category and assistant, consensus/disagreement research, explicit named-not-cited observations, concentration and reshuffle measures, and evidence-based brand topic analysis. These depend on separately supplied research; original scores and weekly collection are unchanged.
Added independent evidence-based question page consumers, three-brand/two-assistant indexing thresholds, explicit history gaps, category indexing controls and server-enforced report-capture access. Real measurements require reconciled questions documents imported separately.
Added read-only weekly archives, topic pages, per-assistant rankings and canonical brand comparisons. Comparison indexing is limited to named-together pairs exceeding 2% in a shared published category. Held/missing evidence is explicitly labelled; stored weekly documents and collection are unchanged.
Added work-email confirmation, manual review, private owner views and revocable access without changing public rankings or weekly collection. Complete-panel question evidence requires independent imports.
Added source-dated citations, structured metadata and complete directory links. Clarified recorded mentions versus distinct answers, published answer totals and assistant icons; retain model-change evidence from stored editions and label unknown adoption dates. Documented the separate quarterly memory method and its limitations.
Launch edition of the UK AI Visibility Index methodology.