On this page
Research summary
The Index at a glance
Each Protocol v1.0 study is a controlled observational snapshot: one exact sector prompt is tested consecutively across five named AI platforms, the visible responses are preserved as evidence, and the combined result is assessed against 15 fixed metrics.
In simple terms
The Index records what five AI systems returned for the same question on a stated date, then measures that evidence consistently. It does not claim that the same answer will appear forever.
Research objective and scope
The parent page is the permanent explainer for the Index. It sets out the research purpose, protocol, metric definitions, evidence rules and limitations; individual child pages document sector-specific tests.
The NeuralAdX UK Business AI Visibility Index is an ongoing observational research programme that records how named UK-sector businesses, organisations and publications are recommended, positioned and cited in live generative AI answers across five major AI platforms.
Each published sector page is a dated retrieval snapshot. It records what the tested systems returned during that specific session. It is not presented as a permanent market ranking, a guarantee of future AI visibility or a claim to know the platforms’ proprietary ranking logic.
In simple terms
We ask the same sector-specific commercial question across five major AI platforms, preserve what they return, and measure the visible results using the same 15 measurement rules every time.
Why the Index exists
AI answer engines can recommend brands, cite websites and present competitive choices directly inside generated responses. The Index exists to make those outputs observable, comparable and evidence-led rather than relying on unsupported visibility claims.
A business can be mentioned without its own website being cited. A company can perform strongly on one AI platform and weakly on another. A citation may be visible while the underlying destination URL remains hidden. Those are materially different outcomes, so the Index measures them separately.
By applying one published framework across different sectors, each child study becomes a comparable evidence record rather than an isolated claim.
Why five platforms?
One engine does not represent the whole AI answer environment. Testing Google AI Mode, ChatGPT, Microsoft Copilot, Claude and Google Gemini allows platform coverage, cross-engine reciprocal-rank prominence and multi-platform retrieval consistency to be measured from the same dated evidence set.
How every Index study works
Protocol v1.0 follows the same seven-stage path from sector selection to a published evidence record.
01
Define the UK sector
Choose a clearly bounded business, publication or organisational category and define which returned entities qualify for inclusion in that sector study.
02
Set the exact prompt
Use one sector-specific buyer-style prompt and preserve its exact wording.
03
Test five AI platforms
Submit the same prompt consecutively across the five fixed platforms.
04
Capture primary evidence
Preserve one screenshot per platform plus the live screen-recorded testing session.
05
Build the evidence ledger
Record eligible brands, complete recommendation order, mentions, visible own-domain citations, identifiable pages and visible supporting domains. The complete ordered brand lists are retained so reciprocal-rank, Jaccard and Rank-Biased Overlap calculations can be reproduced.
06
Calculate 15 metrics
Apply the locked definitions and denominator rules published on this parent page.
07
Publish the study record
Present the results, evidence, test date, methodology and limitations on a child page.
Scope rule
Unless a child study explicitly states otherwise, Protocol v1.0 records one dated retrieval run per platform for that study. This is a controlled observational snapshot, not an estimate of persistent or run-to-run visibility. Academic AI-search research has shown substantial variation across repeated runs, prompts and time, so any cross-engine consistency result in the Index describes only the five captured responses from that session. Future repeated-run studies may be added separately.
The five AI platforms
Every Protocol v1.0 study uses the same five named retrieval environments. The Index records the visible answers returned in the tested interfaces; it does not claim access to hidden model internals. All cross-engine calculations use only the ordered recommendations and citations observable in those captured responses.
01
Google AI Mode
Google’s AI search experience used as one retrieval surface in the Index.
02
ChatGPT
OpenAI’s assistant response environment used as one retrieval surface in the Index.
03
Microsoft Copilot
Microsoft’s assistant/search response environment used as one retrieval surface in the Index.
04
Claude
Anthropic’s assistant response environment used as one retrieval surface in the Index.
05
Google Gemini
Google’s Gemini assistant environment used as one retrieval surface in the Index.
The 15 locked AI visibility metrics
The 15 definitions are the measurement contract for Protocol v1.0. The framework combines transparent NeuralAdX evidence rules, industry-aligned AI-search KPIs and established or explicitly adapted information-retrieval measures. Every metric must be calculable from the captured five-platform evidence; the Index does not infer hidden retrieval logic.
Metric level: metrics 01–13 are reported at brand level where evidence permits. Metrics 14–15 are study-level diagnostics describing the five-platform result set as a whole and are not used as individual-brand tie-breakers.
Provenance labels identify the methodological origin of each measure. External citation chips point to supporting research or industry definitions; they do not imply that those sources endorse or validate the NeuralAdX Index.
Visibility & competitive presence
How often a brand appears, where it appears, how consistently it is positioned and how strongly it competes across the five platforms.
Overall AI Visibility Rank
The final competitive order across the combined five-platform test. Protocol v1.0 ranks brands first by AI Platform Coverage, then Cross-Engine Mean Reciprocal Rank (MRR), then Domain Citations, then Brand Mentions. The next criterion is used only when the preceding criterion is tied; if all four are identical, the brands remain tied.
Brand Mentions
The total number of eligible times the brand is named in the five generated answer bodies. Citation chips, source cards, reference lists and interface labels are not counted as Brand Mentions. This is a NeuralAdX counting rule designed for the captured answer format; it should not be assumed to be identical to a third-party software vendor’s execution-level mention logic.
Share of Voice
The brand’s percentage share of all eligible Brand Mentions recorded in the complete five-platform study. It is calculated as the brand’s eligible mentions divided by total eligible mentions across all included brands, multiplied by 100. The concept is industry-aligned with competitive AI-search Share of Voice measurement.
AI Platform Coverage
How many of the five AI platforms present the brand as an eligible answer or recommendation to the tested sector prompt. Incidental narrative mentions, citation chips, source cards, reference lists and interface labels do not create platform coverage. The result is shown as both a number out of five and a percentage. The denominator is always five.
Average Brand Position
The brand’s average recommendation position across platforms that actually present it as an answer or recommendation. An explicit numbered rank is used where available; otherwise visible top-to-bottom recommendation order is used. Incidental mentions receive no position. Lower is better. This measure intentionally describes position conditional on appearance, so it is interpreted alongside Platform Coverage and MRR.
Position Consistency
Measures variation in the brand’s observed recommendation positions across platforms on which it is actually ranked. Population standard deviation is used: Very High = 0.00–0.49; High = 0.50–0.99; Moderate = 1.00–1.99; Low = 2.00 or higher. At least two ranked platform appearances are required; a brand ranked on only one platform is reported as N/A rather than being labelled artificially consistent.
Cross-Engine Mean Reciprocal Rank (MRR)
An adapted information-retrieval prominence measure that rewards both broad retrieval and high recommendation position. For each brand, that brand is treated as the target item in each engine’s eligible recommendation list: it receives 1/r where r is its recommendation position, while absence receives 0. Cross-Engine MRR is the mean of those five reciprocal-rank values and ranges from 0 to 1. Reciprocal Rank and MRR are established information-retrieval measures; Protocol v1.0 transparently adapts the averaging unit from queries to the five tested retrieval environments.
Citation & source ownership
Whether the brand’s own website is visibly used as a source, how often, where those citations appear and how broadly that citation-backed presence extends across the five platforms.
Domain Citations
How many clearly visible citations or source links point to the brand’s own domain across the five captured responses. Hidden, collapsed or unidentified sources are not guessed.
Citation Share
The brand’s percentage share of all visible eligible own-domain citations recorded in the complete study. It is calculated as the brand’s Domain Citations divided by total visible eligible own-domain citations across all included brands, multiplied by 100.
Average Citation Position
The average visible position of the brand’s own-domain citation where a citation order can actually be identified. Unordered, collapsed or hidden source sets are not given artificial positions. Lower is better.
Cited Pages
How many distinct pages on the brand’s own website can be confirmed from the evidence. If at least one cited page is confirmed but the visible URL detail cannot show whether additional citations lead to different pages, the result is reported as “At least 1 page” rather than giving unsupported precision.
Citation-Backed Presence
The percentage of the five AI platforms that both surface the brand substantively and visibly cite its own domain. The denominator is always all five platforms: one platform = 20%, two = 40%, three = 60%, four = 80%, five = 100%.
Evidence breadth & cross-engine structure
How broad the visible third-party evidence is, how similar the five engines’ recommendation sets are and how concentrated visible citation ownership is within the study.
Source Diversity
How many different visible third-party domains can be confidently associated with evidence around the brand’s placement in the captured responses. The brand’s own domain and unidentified hidden sources are excluded. This is an evidence-breadth measure, not a judgement of source authority or quality.
Multi-Platform Retrieval Consistency
A study-level cross-engine similarity measure reported as two separate values, not collapsed into an invented composite score. Mean Pairwise Jaccard measures set overlap between the eligible recommendation lists; Mean Pairwise Rank-Biased Overlap (RBO) measures ranked-list similarity with greater weight near the top. Five platforms create 10 unique engine pairs. To align the calculation with recent AI-search measurement research, Protocol v1.0 fixes RBO at persistence p = 0.9 and uses the non-extrapolated minimum-bound form. The cross-engine use is an explicit adaptation: it compares the five engines within one dated session and does not estimate run-to-run, prompt or temporal stability.
Visible Source Citation Concentration (Gini)
A study-level measure of how unevenly all clearly visible, identifiable citation-source occurrences are distributed across the distinct cited domains in the five captured responses. It therefore describes the concentration of the visible source ecosystem, not only citations to recommended brands’ own domains. A Gini value of 0 indicates an even distribution; values closer to 1 indicate greater inequality among the observed cited domains. Under the uncorrected finite-sample formula used here, the theoretical maximum is (n − 1) / n for n observed domains. If fewer than two distinct identifiable cited domains are visible, the metric is reported as N/A. This follows the citation-concentration concept used in recent AI-search measurement research while retaining the Index rule that hidden or unidentified sources are never guessed.
Reproducible calculation notes
Protocol v1.0 publishes the implementation rules for its derived academic measures so the same evidence ledger should reproduce the same result.
Cross-Engine MRR
Formula: MRR = (RR1 + RR2 + RR3 + RR4 + RR5) / 5, where RR = 1 / recommendation position and RR = 0 when the brand is absent as an eligible recommendation. If a brand is repeated within one engine answer, only its first qualifying recommendation position is used for RR. Report to three decimal places.
Mean Pairwise Jaccard
Formula per engine pair: J(A,B) = |A intersection B| / |A union B| using the unique eligible recommendation sets. Duplicate appearances of the same brand within one answer do not create additional set items. Average the 10 unique platform-pair values. Order is ignored. If one set is empty and the other is not, Jaccard = 0; if both are empty, that pair is N/A.
Mean Pairwise RBO
Use the non-extrapolated minimum-bound RBO with fixed persistence p = 0.9. For each depth d, let Ad be the overlap proportion between the two ranked prefixes; then RBOmin = (1 − p) × Σ[pd−1 × Ad] from d = 1 to k, where k = min(|S|, |T|). Duplicate brands are removed from each ranked list before calculation. Average the 10 unique platform-pair values. If one list is empty and the other is not, RBO = 0; if both are empty, that pair is N/A. Where an engine explicitly presents tied ranks, preserve the tie and apply the cited tie-aware RBO treatment rather than breaking the tie arbitrarily. Because this is a conservative lower bound, even perfectly matching finite lists can score below 1; report it as a 0–1 coefficient, not as a percentage of agreement.
Visible Source Citation Concentration
Count every clearly visible, identifiable citation-source occurrence by domain across the five captured responses, then sort the resulting positive domain counts as y1 ≤ y2 ≤ … ≤ yn. Gini = [2 × Σ(i × yi)] / [n × Σyi] − (n + 1) / n. This is the rank-weighted computational form used in the cited 2026 AI-search measurement study. Report from 0 to 1 to three decimal places. Hidden or unidentified citations are excluded; if fewer than two distinct identifiable cited domains are available, report N/A.
Measurement contract
A child study must not change a metric’s meaning, denominator or scoring rule. Academic measures that are adapted for the Index are labelled as adaptations and their Index-specific implementation is published here. Child pages may use clearer plain-English display wording provided the underlying calculation remains unchanged. Any future substantive methodological change must be published as a new protocol version.
How the Overall AI Visibility Rank is calculated
The Index deliberately avoids an undisclosed composite score. Protocol v1.0 uses a published lexicographic ordering rule: the next criterion is considered only when the preceding criterion is tied. This keeps the competitive ranking auditable from the evidence table.
AI Platform Coverage
Brands presented as eligible answers or recommendations by more of the five engines rank ahead. This keeps broad multi-platform retrieval as the primary visibility criterion.
Cross-Engine Mean Reciprocal Rank (MRR)
If Platform Coverage is tied, higher MRR ranks ahead. Reciprocal rank gives greater weight to top positions while absence contributes 0, making it a transparent prominence measure across the five engines.
Domain Citations
If still tied, the brand with more visibly attributable own-domain citations ranks ahead.
Brand Mentions
If still tied, total eligible Brand Mentions is the final tie-breaker.
No hidden weighting
The Index does not blend the 15 metrics into an undisclosed percentage score. Multi-Platform Retrieval Consistency and Visible Source Citation Concentration are study-level diagnostics and do not change the brand ranking. If brands remain exactly tied after all four published ranking criteria, the tie is disclosed.
Evidence and transparency standards
The Index quantifies only what the captured material can support. This conservative rule is central to the methodology.
Exact prompt preserved
Every child study displays the exact prompt used in the test.
Position order rule
Use the explicit rank where an AI answer numbers its recommendations. If the recommendations are unnumbered, use their visible top-to-bottom order. Incidental mentions that are not presented as recommendations do not receive a position. If the same brand appears more than once as a recommendation in one answer, its first qualifying recommendation position is used for position metrics and ranked-list construction; later eligible textual occurrences may still count under Brand Mentions. If an engine explicitly assigns the same rank to multiple recommendations, preserve that displayed rank for brand-position and MRR calculations; for RBO, use the published tie-aware treatment rather than imposing an arbitrary order.
Test date preserved
Every study records the date of the live retrieval session.
Five original screenshots
The visible result from each of the five AI platforms is retained as primary evidence.
Live video evidence
The testing session is screen-recorded so the retrieval sequence can be reviewed alongside the screenshots.
No hidden-source guessing
Collapsed +1/+2 controls, unreadable URLs and unidentified citation targets are not attributed to a brand.
No false URL precision
If an own-domain citation is visible but the destination page cannot be distinguished, Cited Pages is reported conservatively.
All eligible brands counted
Share of Voice and Citation Share denominators include all eligible brands in the captured test, not only the eventual top five.
Sector eligibility rule
An entity is eligible when the captured AI answer presents it as belonging to the defined sector or category being tested. Clearly out-of-category entities, institutional or government resources where the category calls for commercial businesses or publications, individuals shown only as commentators, and incidental references are excluded. Material exclusions should be disclosed on the child study page.
Derived metrics reproducible
Ordered recommendation lists and visible citation counts used for MRR, Jaccard, RBO and Gini calculations must be recoverable from the published evidence ledger. RBO uses p = 0.9; the five platforms create 10 unique pairwise comparisons.
Limitations published
Each child study states that the result is a dated retrieval snapshot that may vary in future runs.
Evidence rule
If the screenshot does not expose enough detail to identify a citation, URL or source confidently, NeuralAdX Ltd does not guess it.
How to interpret an Index study
A live AI retrieval study observes a dynamic system. The aim is to make that observation transparent without pretending it is permanent.
What a study can show
Which eligible brands were surfaced in that session, how they were positioned, how often they were mentioned, which own domains were visibly cited, how strongly each brand performed on cross-engine reciprocal rank, how similar the five recommendation lists were, and how concentrated the visible citation-source ecosystem was within the study.
What a study cannot prove
It cannot guarantee the same answer indefinitely, reveal proprietary model ranking logic, establish causal reasons for a platform’s selection, or convert an unidentified hidden source into a known citation.
Interpretation boundary
What this Index does not claim
Not permanent: a dated study does not prove that future AI answers will reproduce the same ordering.
Not predictive: the Index does not forecast future visibility, traffic, revenue or commercial performance.
No hidden inference: unseen citations, collapsed sources and internal model reasoning are not treated as observed evidence.
Not official statistics: this is a proprietary NeuralAdX Ltd observational research index, not a government statistical publication or peer-reviewed academic paper.
Retesting
Individual sectors may be tested again in future to examine change over time. Protocol v1.0 does not promise a monthly, quarterly or annual retest schedule; unless a child page says otherwise, its figures refer only to the dated session shown there. Repeated-run or longitudinal research should be reported separately from the single-session Index snapshot so stochastic stability is not confused with cross-engine consistency.
Methodology governance and research principles
The methodology is version-controlled. Plain-English clarifications may be made within the same protocol version where they do not change a metric’s meaning, calculation, denominator, eligibility outcome or ranking logic. Any substantive change to those rules should be published as a new protocol version rather than silently rewriting earlier studies.
Current protocol
Version 1.0
Effective from 19 August 2026. Five AI platforms, 15 locked metrics, evidence-led counting rules and transparent rank ordering. Protocol v1.0 is the launch methodology for the NeuralAdX UK Business AI Visibility Index.
Status
Proprietary observational index
Produced by NeuralAdX Ltd. It is not official UK statistics, not a peer-reviewed academic publication, and is not endorsed by the tested AI platforms or the external methodological bodies referenced below.
Launch methodology
Protocol v1.0 is the launch methodology for the Index. It uses 15 locked metrics, including Cross-Engine MRR, Multi-Platform Retrieval Consistency and Visible Source Citation Concentration. The same published definitions, denominator rules and evidence standards should be applied consistently to every child study using this protocol.
Academic and industry measurement foundations
Protocol v1.0 deliberately separates three things: metrics adopted from established information-retrieval practice, metrics adapted transparently for a five-engine observational study, and NeuralAdX-specific evidence rules. Recent AI-search research uses Jaccard similarity, Rank-Biased Overlap and Gini concentration to analyse brand/source stability and citation inequality; for RBO, the Index fixes p = 0.9 and the same non-extrapolated minimum-bound form used in the cited 2026 AI-search study, while explicitly adapting the comparison from repeated runs to five engines in one dated session. MRR is adapted from established information-retrieval reciprocal-rank evaluation. Industry-facing concepts such as Brand Coverage, Share of Voice, Average Brand Position and Domain Citations are aligned with widely used AI-search reporting, including OtterlyAI, while NeuralAdX retains its own explicitly published counting definitions for this Index.
Academic-use boundary
The cited research does not validate or endorse the NeuralAdX Index. Where an academic metric is used differently from its original experiment, the adaptation is stated explicitly. In particular, Protocol v1.0 uses Jaccard and RBO to compare five different engines in one dated session; this is not the same as estimating repeated-run, prompt or temporal stability. The Index remains a proprietary observational study, not a peer-reviewed paper.
Don’t Measure Once · arXiv:2604.07585
Rank-Biased Overlap & ties · arXiv:2406.07121
The Case for Reciprocal Rank · arXiv:2312.12672
Critical GEO survey 2023–2026 · arXiv:2607.14035
OtterlyAI Brand Report KPI definitions
NIST AI Risk Management Framework
UK Statistics Authority: Trustworthiness, Quality and Value
ACM Artifact Review and Badging
Research archive
Published Index studies
Each child page is a separate study record beneath this methodology page. Its retrieval date identifies when the AI answers were captured; its protocol version identifies the rules used to analyse that evidence. New records can be added without rewriting the core methodology unless the measurement contract itself changes.
| Study ID | Sector | Prompt type | Test date | Evidence | Record |
|---|---|---|---|---|---|
| NADX-UKBAIVI-001 | UK Cyber Security Trade Publications | Best / Market Leadership | 19 August 2026 | 5 screenshots + live video | Study 001 |
Mobile users: scroll horizontally. The exact tested prompt and complete results are preserved on the individual study record.
Archive rule
The study identifier is permanent. If a sector is tested again in future, the new retrieval session should receive a new study record rather than overwriting the original dated evidence.
Citation guidance
How to cite the Index
A consistent citation makes it easier for readers, businesses and third parties to identify the methodology version being referenced. Individual child studies should additionally cite their permanent study ID and test date.
Suggested parent-page citation
NeuralAdX Ltd. NeuralAdX UK Business AI Visibility Index. Research Protocol v1.0. 2026.
Suggested citations are provided for identification and traceability. They do not imply academic peer review, institutional endorsement or official-statistics status.
Frequently asked questions
These answers explain how the parent Index and its sector child studies should be read.
What does the NeuralAdX UK Business AI Visibility Index measure?
It measures observable brand visibility and citation behaviour in a dated five-platform AI retrieval test. The 15 metrics cover recommendation presence, mentions, share of voice, position, reciprocal-rank prominence, citation ownership, source diversity, cross-engine similarity and visible source-citation concentration.
Is the Index claiming a permanent ranking of UK businesses?
No. Every child study is a dated retrieval snapshot. AI responses can change between runs and over time, so the result is evidence of what was returned during the recorded session, not a permanent guarantee.
Why does each study use five AI platforms?
Using five fixed platforms reduces the risk of treating one engine as representative of the whole AI answer environment and makes platform coverage, reciprocal-rank prominence and cross-engine recommendation similarity directly measurable within the dated session.
How is the overall ranking calculated?
Protocol v1.0 orders brands first by AI Platform Coverage, then Cross-Engine Mean Reciprocal Rank (MRR), then Domain Citations, then Brand Mentions as the final tie-breaker. Absence from an engine contributes 0 to MRR. Only visibly attributable own-domain citations count as Domain Citations. The Index does not use a hidden composite score.
Why are hidden or collapsed sources not counted?
Because the Index is evidence-led. If a screenshot does not reveal the underlying domain or URL, NeuralAdX Ltd does not guess which brand received the citation.
What is Multi-Platform Retrieval Consistency?
It is a study-level comparison of the five engines’ eligible recommendation lists. Mean Pairwise Jaccard measures whether the engines return the same brands, while Mean Pairwise Rank-Biased Overlap measures whether they also place overlapping brands similarly near the top. The two values are reported separately across the 10 unique platform pairs. They describe cross-engine agreement in that recorded session, not repeated-run stability.
How is sector eligibility decided?
An entity is included when the captured AI answer presents it as belonging to the defined sector or category being tested. Clearly out-of-category entities and incidental references are excluded, with material sector-specific exclusions disclosed on the relevant child study page.
Will every sector be retested on a fixed schedule?
No fixed retest schedule is promised under Protocol v1.0. NeuralAdX Ltd is initially expanding the Index across business sectors. Individual sectors may be retested later, but repeated-run or longitudinal measurements should be reported as a separate evidence layer rather than being inferred from a single Index snapshot.
Which Protocol v1.0 metrics use established academic or information-retrieval methods?
Mean Reciprocal Rank is an established information-retrieval concept. Jaccard similarity, Rank-Biased Overlap and Gini concentration are used in recent AI-search measurement research. Protocol v1.0 states where those measures are adapted to the Index’s five-engine, screenshot-based design rather than claiming that the NeuralAdX study reproduces the original academic experiments.
Are the results official statistics or peer-reviewed academic research?
No. The Index is a proprietary observational research programme produced by NeuralAdX Ltd. It uses published methodology and evidence standards, but it is not official UK statistics, a university study or a peer-reviewed journal publication.
Are Google, OpenAI, Microsoft or Anthropic involved in the Index?
No endorsement or affiliation is implied. The named platforms are the retrieval environments tested by NeuralAdX Ltd; the Index methodology and published analysis are produced independently by NeuralAdX Ltd.
NeuralAdX Ltd UK Business AI Visibility Index · Research Protocol v1.0 · Effective 19 August 2026 · 15-metric dated observational retrieval studies · Academic measures labelled as adopted or adapted · No affiliation with or endorsement by the tested AI platform providers is implied.