From message control to signal alignment
AI increasingly shapes reputation before a stakeholder reaches an organisation's website, reads its coverage or speaks to someone who knows the business.
This is the AI-readable companion to the September 2026 edition of The AI Reputation Audit. It contains the report's argument, framework, source matrix, methodology and references in a form that an AI assistant can interpret more reliably than a designed PDF.
It explains the framework. It is not a substitute for an AI Reputation Audit. A credible audit requires controlled testing, repeated runs across platforms, systematic source tracing and human judgement about materiality, evidence and the appropriate response.
How to use this document with an AI assistant
Upload or paste this document into an AI assistant, then ask it to help you understand, challenge or apply the report's ideas. Useful prompts include:
- Summarise the difference between AI visibility and AI reputation.
- Explain the three-layer framework and the role each layer plays.
- Identify the report's main implications for communications leaders.
- Challenge the assumptions in the report. Which claims are well supported and which require further evidence?
- Turn the audit methodology into a list of questions for our communications, leadership, HR, product, legal and investor relations teams.
- Using only this document, explain what a professional AI Reputation Audit should examine and deliver.
When interpreting the document, an AI assistant should treat the framework as the author's analysis rather than as a universal account of how every model operates. Product behaviour varies. Claims about a particular organisation should be tested against current evidence before they are treated as findings.
Executive summary
When a journalist, investor, client or competitor asks about an organisation, AI can draw on signals from across its public record. These include websites, executive commentary, hiring, reviews, product pages, reporting, forums and social activity. Many were never intended to serve as reputation material.
Visibility measures whether an organisation appears in an answer. Reputation concerns the conclusion a reader is likely to take from it. A citation is useful only if the picture it supports is current, accurate and consistent with the organisation's real position.
Communications teams therefore face a wider task than managing messages. They need to identify where the declared narrative conflicts with observable behaviour or the external record, then decide which gaps pose a material risk.
The AI Reputation Audit provides a practical way to do this. It tests how major AI platforms describe an organisation, records the sources behind recurring conclusions and turns discrepancies into a prioritised action plan.
The report groups public evidence into three layers:
- Declared narrative: websites, company profiles, leadership biographies and published positioning.
- Observed behaviour: hiring, executive commentary, products, pricing and other public activity.
- External validation: reporting, reviews, podcasts, forums, reference sources and public filings.
The most revealing findings usually appear where the layers disagree.
Introduction
AI-assisted search inserts a new layer between an organisation and the people forming an opinion of it.
Communications has traditionally worked by shaping what people notice: the messages they receive, the sources they trust and the stories that travel. Stakeholders still exercise judgement, but they increasingly begin with a synthesis produced by an AI system.
That synthesis does not follow the sequence a communications team intended. It may combine current search results, existing model knowledge and material created for entirely different audiences.
A press release, interview or corporate website still matters. So do the signals surrounding them. A job advert can qualify a strategic claim. A product page can contradict an executive position. An old article can supply the frame through which newer evidence is interpreted.
Communications teams cannot control this entire record. They can identify which evidence is shaping the answer, correct errors where possible and bring the organisation's public signals into closer agreement.
The basic flow is:
- Public signals are retrieved or represented within the system.
- The system produces a compressed interpretation.
- A stakeholder may receive that interpretation before examining the underlying record.
The practical question is whether the evidence available to the system supports the reputation the organisation wants to earn.
How AI builds a picture
When someone asks an AI system about an organisation, the answer may draw from two broad sources. The mix varies by product, model and query.
The first is information represented within the model. This is shaped by its training and subsequent updates and may include established associations around better-known organisations. That information can be incomplete or out of date. There is rarely a simple public cut-off that reveals exactly what a current product knows.
The second is retrieved material. Some AI products search the web or other connected sources when answering. They may find recent reporting, corporate pages, job listings, reviews or other public evidence. Other products rely more heavily on information already represented within the model. The same product may also behave differently depending on the question.
The resulting answer is a synthesis of whatever information the system uses. It can sound authoritative without being complete, current or correct.
An organisation can change faster than the picture an AI system has formed of it.
How reliable is this picture?
The user usually sees an answer, not a complete account of how it was produced.
Some AI products provide citations, but these do not necessarily reveal every influence on the response or explain why one source was selected over another. Without citations, it becomes harder to separate current evidence from older model associations.
AI systems can also misstate evidence, confuse organisations, invent citations or repeat errors found in the underlying sources. A 2026 analysis by PAN Communications reported that 31% of more than 11,000 ChatGPT-generated links were misattributed or fabricated. This is one commercially produced study rather than a universal error rate, but it illustrates the risk.
A fabricated claim requires a different response from a genuine contradiction in the public record. The organisation may need to correct an underlying source, publish clearer evidence, make reliable information easier to retrieve or report an error to the platform. No single intervention guarantees that the next answer will change.
A note on variability
Outputs can vary by product, model version, prompt, date, location and access to live search. A useful audit records those conditions and distinguishes recurring conclusions from isolated errors.
Variation is something to measure, not a reason to trust whichever answer appears first.
The three layers
The audit groups an organisation's public evidence into three layers. Each serves a different reputational function. Their value lies in comparing them.
Layer 1: Declared narrative
What the organisation says.
Websites, company profiles, leadership biographies and published positioning establish the organisation's intended story.
Layer 2: Observed behaviour
What the organisation appears to do.
Hiring, executive commentary, products, pricing and other public activity indicate its priorities and direction.
Layer 3: External validation
What others record.
Reporting, reviews, podcasts, forums, reference sources and public filings provide evidence beyond the organisation's direct control.
Comparing the layers
An AI answer may combine evidence from more than one layer. Agreement can reinforce an organisation's position. Contradiction can weaken it. Silence can leave an important claim supported only by the organisation's own assertions.
When an answer appears surprising, trace the conclusion through the three layers:
- Which sources support it?
- Which sources contradict it?
- Is the underlying evidence current and accurate?
- Does the conclusion recur across platforms and prompts?
Use the three layers as a diagnostic rather than a fixed ranking of sources. Different AI products retrieve and use evidence in different ways. A source's influence can vary with the product, prompt, sector, geography, recency and available public record. The framework helps test recurring outputs and trace the evidence behind them. It does not claim to reproduce an AI system's internal reasoning.
The source matrix
Use the matrix to test whether the public evidence surrounding an organisation supports its declared position.
Begin with a claim the organisation considers central to its reputation. Then compare that claim with the sources across all three layers. Look for four conditions:
- evidence that supports the claim
- evidence that contradicts it
- silence where corroboration should exist
- outdated material that may distort the current picture
The matrix covers sixteen source types. For each one, it asks what the source may signal, what it cannot establish, how it may affect reputation and what communications should examine.
Source influence depends on context. Public availability does not guarantee that material will be retrieved, cited or given much weight. Its importance can vary by product, prompt, sector, geography, recency and the amount of competing evidence. The influence labels below are indicative rather than fixed rankings.
The framework works best for organisations with a substantial public footprint. Where the record is thin or new, the first task is to establish a credible evidence base.
A visibility checklist asks where you appear. The source matrix asks what those appearances add up to.
Layer 1: Declared narrative
Website: medium potential influence
- What it may signal: positioning, stated priorities, claims, proof points and current news.
- What it cannot establish: whether the claims are true, current or reflected in behaviour.
- Reputation implication: the organisation's official narrative can establish a frame, but unsupported claims may appear promotional or incomplete.
- What communications should examine: whether central claims are current, specific and supported by observable behaviour and independent evidence.
LinkedIn company page: low to medium potential influence
- What it may signal: official updates, visible activity, company voice and evidence of momentum.
- What it cannot establish: whether the organisation is genuinely active, successful or authoritative.
- Reputation implication: a company page can reinforce identity and continuity but usually has limited value on its own.
- What communications should examine: whether the page aligns with leadership and employee activity and whether effort is proportionate to its likely influence.
Leadership biographies: medium potential influence
- What they may signal: credentials, career history, current role, expertise and public positioning.
- What they cannot establish: whether expertise is current, applied or recognised by others.
- Reputation implication: biographies help shape perceptions of authority, especially when stronger independent signals are sparse.
- What communications should examine: whether biographies are consistent across channels and supported by visible evidence of current work.
Layer 2: Observed behaviour
LinkedIn activity: medium to high potential influence
- What it may signal: current priorities, themes, consistency of thought and leadership visibility.
- What it cannot establish: whether the organisation acts consistently with what its people publish.
- Reputation implication: sustained activity can shape associations around both individuals and the organisation.
- What communications should examine: whether recurring themes substantiate the declared position and whether highly coordinated activity risks appearing artificial.
Job listings: high potential influence
- What they may signal: where the organisation is investing, the capabilities it values, locations, growth areas and operating priorities.
- What they cannot establish: whether the strategy will succeed or whether all advertised roles will be filled.
- Reputation implication: hiring evidence can corroborate or contradict stated strategy because it reflects resource allocation.
- What communications should examine: whether current roles support published priorities or create a materially different picture of the organisation's direction.
Executive commentary: high potential influence
- What it may signal: strategic priorities, judgement, consistency, domain expertise and leadership conviction.
- What it cannot establish: whether the organisation is executing effectively or delivering results.
- Reputation implication: executive commentary often carries more weight than general corporate messaging because it attaches claims to accountable individuals.
- What communications should examine: whether leaders' positions are specific, consistent over time and supported by organisational decisions and outcomes.
Product and pricing pages: medium potential influence
- What they may signal: what the organisation actually sells, how it positions value, target audiences and commercial direction.
- What they cannot establish: customer satisfaction, product quality or competitive strength.
- Reputation implication: product evidence can reveal a gap between broad positioning and the organisation's real offer.
- What communications should examine: whether the commercial offer supports the strategic narrative and whether important changes are visible in the public record.
Employee activity: medium potential influence
- What it may signal: visible expertise, cultural cues, employee advocacy and the breadth of participation around a theme.
- What it cannot establish: the views or experience of the full workforce.
- Reputation implication: distributed employee evidence can corroborate organisational claims, while silence or contradiction can weaken them.
- What communications should examine: whether relevant expertise is visible naturally and whether employee activity is consistent with the organisation's stated culture and priorities.
Layer 3: External validation
News and editorial: high potential influence
- What it may signal: independent coverage, prominence, controversy, credibility and established narrative frames.
- What it cannot establish: the full picture or the organisation's current position when coverage is old or incomplete.
- Reputation implication: repeated editorial frames can persist after the facts or the organisation have changed.
- What communications should examine: which narratives recur, how old they are, whether they are accurate and where credible evidence could update the record.
Podcasts and interviews: high potential influence
- What they may signal: executive thinking, nuance, confidence, domain expertise and long-form consistency.
- What they cannot establish: wider organisational behaviour or commercial performance.
- Reputation implication: long-form conversations can provide rich evidence for leadership and expertise claims, particularly when transcripts or detailed summaries are available.
- What communications should examine: whether executive appearances reinforce a coherent position and whether useful material is accessible beyond the audio or video itself.
Independent and creator media: medium to high potential influence
- What it may signal: how sector specialists, analysts, creators and niche communities interpret the organisation.
- What it cannot establish: broad consensus or formal authority.
- Reputation implication: credible independent voices can carry disproportionate influence within a specialist audience.
- What communications should examine: which commentators matter to priority stakeholders and whether their assessments are informed, current and fair.
Reviews: high potential influence
- What they may signal: recurring customer or employee experiences, operational strengths and persistent complaints.
- What they cannot establish: a representative view in every case. Review populations may be self-selecting or manipulated.
- Reputation implication: repeated patterns in reviews can strongly affect trust and may be repeated in AI-generated answers.
- What communications should examine: the operational causes behind recurring themes, the credibility of the evidence and whether public responses are proportionate and factual.
Forums and communities: medium to high potential influence
- What they may signal: peer discussion, real-world problems, unfiltered questions, product workarounds and specialist sentiment.
- What they cannot establish: verified identity, representativeness or factual accuracy.
- Reputation implication: forums can reveal concerns that formal channels miss, but individual claims require careful verification.
- What communications should examine: recurring objections or misconceptions and the underlying product, service or information gap. Manufactured participation creates its own risk.
Wikipedia: high potential influence
- What it may signal: established facts, history, category, notable events and the summary view of an organisation.
- What it cannot establish: a complete or perfectly current picture.
- Reputation implication: concise reference material can supply the default frame used in later summaries.
- What communications should examine: factual accuracy, missing context and whether reliable independent sources exist to support legitimate corrections through Wikipedia's editorial process.
Directories and filings: medium potential influence
- What they may signal: company facts, ownership, location, status, executive appointments and official records.
- What they cannot establish: reputation, strategy or qualitative performance.
- Reputation implication: factual inconsistency across official or semi-official sources can undermine confidence and confuse identification.
- What communications should examine: whether core records are accurate, current and consistent across authoritative listings.
Institutional and authority sources: high potential influence
- What they may signal: regulator records, research evidence, professional standing, awards, standards and official recognition.
- What they cannot establish: the organisation's full strategic or commercial position.
- Reputation implication: authoritative sources can strongly support or challenge claims involving compliance, expertise, quality or legitimacy.
- What communications should examine: whether relevant public records are accurate and whether the organisation's claims match the evidence held by credible institutions.
What this changes for communications
AI-generated answers may draw on evidence controlled by different parts of an organisation. Communications cannot manage every source, but it can coordinate the overall picture and act when those signals diverge.
1. Communications coordinates the whole view
Relevant signals sit across PR, search, HR, investor relations, legal, product and leadership activity. Each function owns part of the evidence, while communications is well placed to assess what the combined record says about the organisation. Its role is to maintain the overall view, bring the relevant teams together and escalate contradictions that could affect reputation.
2. Measurement expands beyond exposure
Reach, share of voice and sentiment remain useful. They do not show whether the evidence available across the three layers supports the organisation's intended position. An AI Reputation Audit adds a different set of measures: recurring conclusions, sources shaping those conclusions, material contradictions, outdated evidence and the time taken to address important gaps.
3. Executive visibility requires proof
Greater visibility can increase awareness of a leader's position. Credibility depends on whether the public record supports it. Executive communications should connect claims with decisions, outcomes and independent evidence. The audit should also test whether the leader's stated position is consistent with the organisation's observable behaviour.
A necessary limit
The purpose is to make accurate evidence easier to find and interpret. Where the organisation's behaviour creates the contradiction, communications should surface the problem rather than attempt to repair it through messaging alone.
Manage AI reputation as a recurring operating discipline, with clear owners and a regular review cycle.
The framing of communications as the function whose object is coherence draws on Elif Güvençer's Two Clocks (April 2026).
The AI Reputation Audit
The audit is a controlled investigation of how AI systems describe an organisation, which evidence shapes those descriptions and where the resulting picture creates reputational risk or opportunity.
1. Define the decision context
Set the priority audiences, markets and questions. The audit focuses on situations in which reputation could affect judgement, consideration or commercial decisions.
2. Build a controlled evidence set
Run structured query sets across major AI platforms and repeat the most important tests. Record the model, date, market, settings, answer and cited sources.
3. Trace the evidence
Work backwards from each recurring conclusion. Identify the sources contributing to it and classify them as declared narrative, observed behaviour or external validation.
4. Diagnose the gaps
Distinguish genuine contradictions from stale information, retrieval differences and model fabrication. Assess each finding for recurrence, credibility and potential reputational effect.
5. Prioritise the response
Rank the findings by materiality, confidence and ability to influence. Identify the right internal owner and the evidence, communication or operational change required.
The result is a documented baseline, a source map and a prioritised action plan against which change can be measured.
The audit cycle is therefore:
Define the decision context → Build a controlled evidence set → Trace the evidence → Diagnose the gaps → Prioritise the response → Repeat as the public record and AI systems change.
A rigorous audit turns fragmented outputs into a structured diagnosis and a prioritised plan for action.
The prompt is the easy part
A handful of searches can show what one AI system said on one day. They cannot reliably show whether the conclusion is stable, why it formed or what should be done about it.
A credible AI Reputation Audit requires stakeholder-specific query design, repeated testing across platforms and systematic source tracing. It must distinguish persistent patterns from isolated outputs, genuine contradictions from model errors and material risks from findings that merely look interesting.
The hardest part is interpretation. An organisation needs to know which conclusions could affect trust or commercial decisions, which evidence is producing them and whether the response belongs in communications, leadership, HR, product, legal or investor relations.
What the audit delivers
- A tested view of how major AI platforms describe the organisation
- A source map showing which evidence shapes recurring conclusions
- An assessment of alignment across declared narrative, observed behaviour and external validation
- A prioritised account of the reputational risks and opportunities
- A practical action plan with owners, sequencing and measures for reassessment
Adelbury's AI Reputation Audit turns fragmented AI outputs into a structured diagnosis of how an organisation is currently understood. It shows where the public record supports the intended reputation, where it works against it and which interventions deserve attention first.
Running prompts produces observations. A proper audit produces a basis for action.
About
This report was written by Stephen Davies, founder of Adelbury, a brand and marketing consultancy. It reflects his own view, formed through work with communications and marketing leaders adapting to AI without surrendering judgement, trust or strategic clarity.
Adelbury helps organisations understand and shape the picture AI builds of them: mapping the public record, finding the gaps between what a business says, what it does and what others say about it before someone else does.
Start the conversation
To talk through what this means for your organisation, contact Stephen at stephen@adelbury.com.
Appendix
The picture in this report is consistent with a growing body of independent research on how AI-assisted search works, the risks and limitations of AI-generated answers and the emerging communications discipline around reputational discovery. The sources below are included so an AI assistant can examine the evidence behind the report's claims, distinguish supported findings from interpretation and identify where further verification may be useful.
What AI systems retrieve and synthesise
Retrieval-augmented generation
Lewis, P. et al. (2020), Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, Meta AI Research. Introduced RAG as a method for combining a language model's internal knowledge with retrieved external documents.
Retrieval and citation behaviour
Malaviya, C. et al. (2024), ExpertQA: Expert-Curated Questions and Attributed Answers, ACL Anthology. Evaluates long-form factual answers and citation completeness, including the extent to which claims are supported by retrieved sources.
Liu, N. F. et al. (2023), Lost in the Middle: How Language Models Use Long Contexts, Stanford University and partners. Shows that models often use information less reliably when it appears in the middle of long contexts, illustrating why retrieval does not guarantee equal use of every source.
Source bias and commercial influence
Sun, X. et al. (2025), Generative Engine Manipulation, Princeton University. Demonstrates that webpages can be modified in ways that significantly increase their visibility in generative search responses, raising concerns about manipulation and source-selection bias.
GEO-Bench (2025), Dataset for Benchmarking Generative Engine Optimisation. Provides a benchmark for studying which webpages and sources appear in AI-generated answers.
What AI does with sources
Hallucination, misinformation and source error
OpenAI (2025), Why Language Models Hallucinate. Explains why large language models generate plausible but incorrect statements and why evaluation methods can reward guessing rather than calibrated uncertainty.
Kalai, A. T. et al. (2025), Why Language Models Hallucinate. Formal analysis of hallucination as a statistical consequence of next-token prediction and current evaluation methods.
Mündler, N. et al. (2024), Self-Contradictory Hallucinations of Large Language Models: Evaluation, Detection and Mitigation. Demonstrates that language models can contradict their own previous outputs, reinforcing the need for repeated testing rather than relying on a single answer.
PAN Communications (2026), The 2026 AI Answer Trust Report. Analysis of more than 11,000 ChatGPT citations reported that 31% were misattributed or fabricated. This is a commercially produced study and should be treated as an indicative finding rather than a universal error rate.
Citation completeness and trust
Rashkin, H. et al. (2023), Measuring Attribution in Natural Language Generation Models, Google Research. Proposes methods for evaluating whether generated claims can be attributed to supporting evidence.
Why AI reputation matters
AI systems as reputational intermediaries
Güvençer, E. (2026), The Two Clocks: Why AI Search Is Collapsing the Distance Between Narrative and Reality. Argues that communications teams increasingly face two simultaneous clocks: the narrative an organisation intends to present and the evidence AI systems can surface independently.
Lewandowsky, S. et al. (2020), Technology and Democracy: Understanding the Influence of Online Technologies on Political Behaviour and Decision-Making, European Commission Joint Research Centre. Examines how digital intermediaries shape public understanding and trust.
Machine-mediated reputation and visibility
Aggarwal, P. et al. (2024), Search Engines and Reputation Formation. Research on how ranked and summarised information environments affect trust and organisational perception.
Related industry research
Several industry reports have examined how organisations appear in AI-generated answers and how communicators are adapting:
- Muck Rack (2025), What Is Generative Engine Optimisation? https://muckrack.com/blog/what-is-generative-engine-optimization
- Semrush (2025), AI Search Visibility Study. https://www.semrush.com/blog/ai-search-visibility-study/
- Profound (2025), AI Answer Engine Insights. https://www.tryprofound.com/resources
- Authoritas (2025), AI Overviews and Search Visibility Research. https://www.authoritas.com/
These sources are useful for market context and emerging practice. Where possible, their claims should be checked against independent research and direct testing.
The AI Reputation Audit Stephen Davies, Adelbury September 2026 stephen@adelbury.com