Reputation as a revenue factor: how LLM sentiment drift quietly devalues brands.

For decades, reputation was a PR discipline. SEO was a technology discipline. ORM was the unloved borderland between them — often pushed into legal teams, rarely strategically integrated. That organizational separation worked as long as the signals were processed in separate channels.

With LLMs spreading as information sources (ChatGPT: 500M+ weekly users, Gemini: ~350M, Perplexity: ~30M), the separation no longer holds. Generative models process brand mentions, sentiment data, structured trust signals and E-E-A-T indicators simultaneously — and communicate the result to the user in a single synthetic answer.

The sentiment-drift mechanism

The term sentiment drift describes a phenomenon we have systematically documented in client engagements since 2024: a brand's representation in LLM answers shifts over time — even when the objective data has not changed.

The causes lie in two mechanisms:

Training-layer drift

LLMs are regularly trained on new data. Between training cycles, information lingers in the model's "memory layer" — even when it has long been overtaken by current web reality. A viral negative thread on Reddit or a negative trade-media piece can embed itself in training data and stay present in the model's perception for months or years.

RAG-layer drift

Modern systems such as Perplexity, ChatGPT Search and Google AI Overviews use Retrieval Augmented Generation: a query triggers a live fetch from the web. Here, sentiment hits even faster. If a brand query such as "[brand] reviews" surfaces a top results page made up of Reddit criticism, Trustpilot complaints and a critical blog post, the LLM synthesizes those signals into its answer — even when 95% of actual customer experiences are positive.

Why classical ORM is no longer enough

The established ORM playbook — review management, suppression content, legal requests to Google — was designed for a world in which users read and evaluate search results themselves. In the generative world, the AI decides what is relevant and presents it as a consolidated answer. The user sees the verdict, not the sources.

Concretely, the following classical tactics fail:

• Suppression content: works for SERP positions, but not for LLMs that understand content semantically. A blog post titled "Company X does not have this problem" can be paraphrased as confirmation of the problem.
• Review gamification: LLMs detect conspicuous review patterns (sentiment bias, phrase repetition) and downweight such sources.
• Legal takedowns: remove content from the live index but not from training data of models built on older crawls.
• PR counter-narratives: without an entity strategy, they are not stored as linked to the brand.

The reputation-engineering model

Reputation engineering is the systematic build-up of trust signals that act across every layer of an AI answer: training corpus, RAG retrieval, on-page signals and structured data.

Layer 1: training-layer signals

Content must land in LLM training sources. The dominant sources (Common Crawl, Wikipedia, qualified news, curated datasets) determine how a brand is stored in future model iterations. Measures:

• Wikipedia articles with verified relevance (notability) and qualified sources
• Contributions to highly authoritative trade media with clear authorship
• Academic citations where possible (research partnerships)
• Podcast transcripts and video captions (scraped by many models)

Layer 2: RAG-layer signals

What gets retrieved at live query time. Here, classical SEO combines with citation-optimized content:

• Dominance of the brand SERP (positions 1–10 for "[brand]", "[brand] reviews", "[brand] ratings")
• Highly authoritative own-domain pages with a clear information architecture
• Updated FAQ schemas with answers to critical questions
• Structured review aggregation via the AggregateRating schema

Layer 3: sentiment monitoring inside generative engines

Classical brand monitoring (Brandwatch, Talkwalker, Meltwater) is no longer enough. What is needed is prompt-based sentiment tracking:

• Weekly audits against 100–300 relevant brand prompts
• Measurement: is the brand mentioned? In what context? With what sentiment? Which competitors are surfaced as alternatives?
• Tools: Profound, Otterly.ai, BrandRank.ai, custom setups with API access

"Reputation in the LLM era is no longer about what you say. It is about how models learn the world about you. And those learning processes run in dimensions no PR team knew three years ago."

Case pattern: how crises play out differently in LLMs

Across 40+ reputation crises affecting European brands (2023–2025), a consistent pattern emerges — one that is fundamentally different from classical PR crises:

Days 0–3: viral peak

The negative event goes viral on social media, Reddit, X. Classical PR response: statement, explanation, sometimes an apology.

Days 3–21: classical media coverage

Trade media pick up the topic. Ranking shifts appear on the SEO side. Search volume for critical brand queries rises.

Days 21–60: the invisible LLM lag

While the Google SERP is normalizing and PR cycles wind down, LLMs only begin to "understand" the crisis. Models learn the association brand ↔ crisis topic and replay it in their answers. The peak of LLM visibility of the crisis: days 30–45.

Days 60–180: sentiment persistence

Without active countermeasures, the crisis stays present in LLM answers. Especially robust: associations with specific topics (e.g., "Company X" ↔ "privacy issue"). These associations persist across retraining cycles because they are dominant in the training-data distribution.

The five principles of modern reputation engineering

1. Proactive over reactive: trust signals are built before the crisis. After the crisis, it is too late for training-layer impact.
2. Source before channel: a qualified trade-media article has more impact in LLMs than 100 social posts. Distribution strategy must target LLM-relevant sources.
3. Entity consistency: every positive signal must be linked to the brand entity. Loose content that does not clearly belong to the brand has no effect.
4. Long-term frequency: the memory of LLMs is additive. Constant, qualified signals over months beat short, intensive campaigns.
5. Measurement at the prompt level: reputation KPIs are measured directly in LLM answers — not in traditional brand trackers alone.

The reputation-impact formula: quantifying sentiment drift

Reputation is measurable. We use a composite score — the Reputation Drift Index (RDI) — to track the evolution of brand perception inside LLMs. The RDI ranges from −100 to +100 and is calculated weekly.

RDI = (Σ (s_i × w_i × c_i)) / (Σ w_i) × 100

where:
s_i  = sentiment score of the i-th prompt response (−1 negative, 0 neutral, +1 positive)
w_i  = prompt weight (business relevance × prompt frequency in real traffic)
c_i  = confidence score (how clearly the sentiment is expressed, 0-1)

Test set:  200 brand-relevant prompts across 4 LLMs = 800 queries/week
Threshold: RDI > +30 = healthy; 0-30 = neutral; −30-0 = warning; < −30 = critical

For sentiment classification we use two independent models (GPT-4o-mini + Claude Haiku) in cross-validation mode. Disagreements are reviewed manually. Inter-annotator agreement in our client setups sits at κ = 0.81 — operationally reliable.

Source weighting: which sources shape the model's memory

Not all sources have equal weight. We work with an internal source-weight index that classifies publications into four tiers. The tier assignment is based on (a) presence in Common Crawl, (b) domain-specific authority, (c) frequency as a citation source in LLM outputs.

The practical consequence: a single verified Wikipedia edit with a clean secondary source can generate more reputation impact than 50 corporate blog posts. Spending reputation budget on tier-3 distribution wastes budget.

Tutorial: reputation-defense setup in 30 days

Week 1 — audit & baseline

Run 200 brand prompts against all four LLMs, calculate RDI. In parallel: brand-SERP audit for the top 30 brand queries in Google, Bing, DuckDuckGo. Identify every source that appears in the top 10. Build a reputation inventory: each mentioning source with sentiment, tier and access status.

Week 2 — secure quick wins

Close the obvious gaps: claim the Knowledge Panel, update the Wikidata item, take the brand FAQ schema live, implement AggregateRating on the homepage, answer negative reviews with a verified counter-statement (do not have them deleted — that performs worse in LLMs than a visible, composed response).

Week 3 — tier-1 content sprint

At least three substantive pitches to tier-1 publications (expert contribution, data story, interview). In parallel: check Wikipedia notability; where it is given, draft a neutral article in the Draft namespace (own edits are COI — better an experienced external author).

Week 4 — infrastructure

Set up a weekly prompt-monitoring dashboard. Alert logic: RDI drop > 15 points in 7 days → automated ping to the comms lead. Second alert: new entity associations not present in the preceding weeks → possible emerging crisis.

Crisis playbook: the first 72 hours no longer decide alone

In classical crisis management, the rule was: "The first 72 hours decide everything." In the LLM era, that is only half true. The first 72 hours decide the short-term SERP and social impact. The next 30–90 days decide the LLM impact, which can preserve a crisis for years.

Our three-phase playbook:

Phase 1 (days 0–3) — containment. Classical crisis PR: statement, ownership, correction. In parallel: brief tier-1 sources with a validated fact check, so the first wave of coverage is balanced.

Phase 2 (days 3–30) — narrative reinforcement. Ten to fifteen qualified follow-up pieces with a constructive frame. Decisive: in every piece, the brand must appear in the same paragraph as the solution/response. That shifts the co-occurrence in LLM training data away from the problem and toward the correction.

Phase 3 (days 30–180) — entity reframing. Actively occupy new co-occurrence fields: the brand appears in contexts that have nothing to do with the crisis topic — innovation, civic engagement, operational excellence. The goal: the crisis becomes one of many dimensions of the brand's representation, not the dominant one.

Case study: reputation recovery of a B2B brand

A German mid-market company (mechanical engineering, ~EUR 180M revenue) was the target of a viral criticism campaign in Q1 2025 over alleged supply problems. No legally substantive core, but a well-linked Reddit thread with 2,400 upvotes, two trade-media pieces and a LinkedIn wave.

Initial RDI (two weeks after the viral peak): −47. The supply problem was mentioned in 73% of prompts about the brand. Classical ORM had tried to have the Reddit thread removed — unsuccessfully, because there was no legal violation.

Our approach: 14 weeks, 11 tier-1 publications (Handelsblatt, VDI Nachrichten, Produktion, three industry-association blogs, two podcasts, one scientific cooperation with a university of applied sciences including a paper). In parallel: Wikipedia neutralization through a clean fact check, a corporate-blog series on supply-chain transparency (8 articles), an interview series with customers as testimonial content.

RDI after 14 weeks: +18. After 26 weeks: +34. Mention frequency of the supply problem in LLM answers: from 73% to 9%. The critical factor: the crisis topic was not denied — the brand addressed it proactively in its own contributions, while installing a correction narrative. LLMs weight proactive, data-based communication more strongly than silence or defensiveness.

The connection to enterprise value: a worked example

Why is reputation engineering a CFO topic? We ran a simple attribution case for an e-commerce client:

Monthly organic sessions from brand queries:        180,000
Conversion rate on brand traffic:                     4.8%
Average order value:                                EUR 142
Monthly revenue from brand search:               EUR 1,228,000

RDI drop from +20 to −15 → empirically linked
conversion-rate reduction:                          −22%
(mechanism: LLM citations with hedging/negative framing
 reduce pre-click brand trust)

Monthly revenue loss:                            EUR 270,160
Annual loss with sustained effect:               EUR 3.24M

No PR budget would invest mid-six-figure EUR amounts in reputation — but an annual revenue loss of EUR 3.24M with rankings unchanged easily justifies EUR 300–500k in structural reputation engineering. The math only becomes visible when attribution is done correctly.

Conclusion

Reputation is no longer a soft goal in the AI era. It is a hard, measurable revenue factor that influences purchase decisions in every LLM answer. Organizations that continue to treat reputation as a residual PR item systematically underestimate how their market perceives them — and how fragile that perception is in the face of asymmetric events.

Reputation engineering moves the discipline where it belongs: into the center of strategic corporate communication, tightly interlocked with SEO, content and data.