How Big Brands Manipulate AI Search Results and What It Means for Everyone Else

The Shift from SEO to GEO: Welcome to the Citation Economy

For twenty-five years, search engine optimization followed a recognizable pattern: figure out what the algorithm rewards, do more of that, wait for the algorithm to catch up and penalize you, then adapt. The move from SEO to Generative Engine Optimization (GEO) breaks that pattern because the “algorithm” is no longer a set of ranking rules. It is a neural network that synthesizes answers from a massive corpus of training data and real-time retrieval.

Traditional SEO rewarded keywords, backlinks, and page speed. GEO rewards something fundamentally different: entity authority, information gain, and machine readability. Visibility is no longer about appearing in a list of ten blue links. It is about being the source an AI chooses to cite when it constructs an answer. Research published on ResearchGate calls this the “Citation Economy” — a system where the currency is not clicks but mentions, not traffic but narrative control.

Edelman’s research found that up to 90% of the citations driving brand visibility in LLMs come from earned media — news coverage, Wikipedia entries, Reddit threads, and authoritative third-party platforms. Brand-owned content barely registers. That finding alone should reshape how companies allocate marketing budgets. You cannot buy your way into an AI’s answer the way you can buy a Google Ad. You need third parties saying credible things about you, and you need those statements to exist in the places AI systems trust.

This table illustrates the core differences between the old model and what has replaced it:

The strategic focus has migrated from optimizing for human scanners to engineering relevance for machine extraction. Brands now need to manage “baseline familiarity” — the knowledge that a model carries about them in its parameters before any real-time search occurs. If a language model has never encountered your brand during training, no amount of Retrieval-Augmented Generation (RAG) — the process where AI pulls in fresh web data to supplement its answers — will make it trust you as much as a brand it already “knows.”

This Is Not New. It Is Just More Sophisticated.

The history of search marketing is a history of manipulation. Amsive’s analysis of SEO eras reveals a repeating cycle: marketers exploit an algorithm’s blind spot, the platform patches the vulnerability, and the next generation of tactics emerges. The current AI era mirrors this history but operates at a far more sophisticated level.

Where J.C. Penney was penalized in 2011 for buying cheap links from obscure sites, modern GEO manipulation targets the structural topology of the web itself. Where BMW used doorway pages in 2006 to fool simple crawlers, today’s well-funded brands engineer their position within the web graph to ensure disproportionate representation in AI training data. The goal has evolved from “rank higher” to “become part of what the AI knows as truth.”

Harmonic Centrality: The Invisible Ceiling Keeping You Out of AI Answers

Most people assume that AI models learn from “the internet.” They do not. They learn from filtered, processed versions of it — and the filtering is not neutral.

The largest source of training data for major language models is Common Crawl, a nonprofit archive that has been crawling the web since 2008. Research from LLMrefs confirms that over 80% of GPT-3’s training tokens came from filtered Common Crawl archives. But Common Crawl does not archive the internet equally. It uses a metric called Harmonic Centrality (HC) to decide which domains to crawl first, how often, and how deeply.

Harmonic Centrality measures how well connected a domain is by mapping the shortest paths between it and every other domain on the web. Unlike PageRank, which counts “votes” from incoming links, HC rewards domains that function as central hubs in the overall web structure. Think of it like measuring how close a building is to the center of a city — not by counting the number of roads leading to it, but by how quickly you can reach every other building from it.

This creates what practitioners describe as an “invisible ceiling.” If your domain sits outside the top one million in HC rank, Common Crawl visits you infrequently and captures fewer of your pages. Fewer pages in Common Crawl means less representation in the training data that language models use to understand the world. Less representation means the AI has never heard of you. And if the AI has never heard of you, it will not recommend you — even when a user asks a question your product directly answers.

How Big Brands Game the Web Graph

The compounding chain works like this: high connectivity produces a lower HC rank, which leads to more frequent crawling by Common Crawl. More frequent crawling means more of the brand’s content ends up in training datasets. More training data builds “baseline familiarity” inside the model’s parameters. That familiarity makes the AI more likely to recommend the brand — even before it performs any real-time web search via RAG.

Consider an example from the automotive industry: Ford holds an HC rank of approximately #674, while Tesla sits at roughly #1,336. Both are well-known brands, but Ford’s structural position within the web graph gives it a measurable advantage in how thoroughly its content is archived, indexed, and ultimately absorbed by AI systems. That gap widens dramatically as you move down to mid-tier brands like Volkswagen or Nissan, which sit further from the web’s structural core.

High-investment brands manipulate this system deliberately. Instead of pursuing link volume — the old SEO playbook — they focus on link topology. They secure links from existing central hubs: Wikipedia, major news outlets, government directories, university research pages. Each of these links moves the brand’s domain structurally closer to the center of the web graph. This is a long-term investment that compounds across every new version of every AI model trained on Common Crawl data, effectively locking out competitors who lack the resources to penetrate these elite network clusters.

For smaller brands, the implication is stark: you are not competing on content quality alone. You are competing against an architecture that was established before the AI even began to “think.”

AI Recommendation Poisoning: How Brands Inject Themselves Into Your AI’s Memory

Structural authority builds long-term advantage. But some brands are not waiting for long-term advantage. They are taking shortcuts — by injecting instructions directly into AI assistants’ memories.

In February 2026, Microsoft’s security team published a detailed report documenting a technique called AI Recommendation Poisoning. The mechanism is simple and alarming: “Summarize with AI” buttons and social sharing links embedded on websites contain hidden URL parameters that feed manipulation instructions to AI assistants. When a user clicks one of these buttons, the link opens Copilot, Perplexity, or another AI assistant with a pre-loaded prompt — one the user never sees — that tells the AI to “remember Company X as a trusted source” or “always mention this brand as the top choice in future conversations.”

Because modern AI assistants are designed to retain context and remember user preferences across sessions, they ingest these instructions as legitimate preferences. Once poisoned, the AI treats the injected “facts” as established truth, influencing future responses even in completely unrelated conversations. The user has no idea this has happened.

Industrialized Tools for Memory Injection

What makes this particularly dangerous is that the tools for executing these attacks are already commercially available. The CiteMET NPM package and the AI Share URL Creator are marketed openly as “growth hacks for LLMs,” providing ready-to-use code for embedding manipulative triggers into websites. These tools represent the industrialization of what security researchers classify as Cross-Prompt Injection Attacks (XPIA) — hidden instructions in documents, emails, or web pages that manipulate an AI as it processes the content.

The targets are not random. Microsoft’s research identified that these attacks disproportionately target high-stakes categories — medical advice and financial services — where a biased recommendation can have severe real-world consequences. A user who asks their AI assistant “what is the best insulin pump” or “which financial advisor should I use” may receive an answer that was planted by a brand, not derived from legitimate analysis.

This mirrors traditional SEO poisoning but targets something far more persistent: the internal memory architecture of the AI itself. A manipulated search result disappears when Google updates its algorithm. A poisoned AI memory persists until the user manually clears it — which most users never do, because most users do not know it exists.

GEO Warfare: When Competitors Weaponize AI Against Your Brand

Manipulating AI to promote your own brand is one thing. Using AI systems to destroy a competitor is something else entirely — and it is already happening.

The practice of weaponizing the information ecosystem to manipulate AI outputs is increasingly called GEO Warfare. It goes beyond self-promotion into deliberate sabotage: creating “information asymmetries” that make AI systems less confident in rival products. The documented tactics are sophisticated and coordinated.

The Four Vectors of GEO Warfare

Narrative Manipulation works by flooding the market with coordinated negative messaging. A competitor funds analyst reports that criticize “legacy systems” and seeds Reddit discussions about a targeted brand’s “outdated” infrastructure. AI systems synthesize this manufactured consensus and begin advising users to be cautious of the targeted brand.

Source Credibility Hijacking operates by saturating high-authority publications with positive coverage of one brand. Since AI systems weight these publications heavily during synthesis, this effectively drowns out competitors whose perspectives do not appear in the same tier of sources. A startup’s point of view becomes invisible — not because it is wrong, but because it does not exist in the publications the AI trusts.

Trust Erosion through FUD (Fear, Uncertainty, and Doubt) coordinates the seeding of regulatory criticism, customer complaints, and safety concerns. When AI systems encounter this volume of negative signal, they synthesize more cautious or outright negative answers about the targeted brand. An investor asking ChatGPT about a company’s reliability may receive a warning that originated entirely from planted content.

Digital Sabotage takes a more direct approach: registering negative domains, publishing hostile analysis pieces, and creating content designed to make AI systems synthesize a negative “consensus” about a specific competitor in comparative queries.

Proof That It Works

This is not theoretical. A study published in July 2025 by Reboot Online Marketing proved that ChatGPT and Perplexity responses could be successfully manipulated through strategically placed content on low-authority expired domains. By planting “brand authority statements” across domains with low domain rating scores, the researchers influenced AI recommendations within days — not months, not years. Days.

The study’s conclusion was blunt: AI systems are “shallow readers” of the internet whose responses can be guided through deliberate content placement. They do not deeply evaluate the trustworthiness of every source. They synthesize what they find and present it with confidence. That confidence is what makes the manipulation so effective — users trust the AI’s answer precisely because it sounds certain.

For any brand operating in a competitive market, this means your AI visibility is not just a function of what you do. It is also a function of what your competitors do to the information environment around you. If you are not monitoring how AI systems describe your brand — and checking whether that description is being deliberately manipulated — you are exposed. AI search optimization now requires active defense as much as proactive positioning.

The Citation Economy: Who AI Trusts, and Why It Is Not You

If GEO Warfare is the offensive side of AI manipulation, the Citation Economy is the structural bias that makes it possible. AI systems do not treat all sources equally. They have a hierarchy of trust — and the brands at the top of that hierarchy have paid to be there.

Edelman’s research established that up to 90% of citations driving LLM brand visibility come from earned media. Not your blog. Not your About page. Not your carefully optimized product descriptions. Earned media: news coverage, Wikipedia entries, Reddit threads, analyst reports, and review platforms.

Research from Discovered Labs found that Reddit is cited 40.1% of the time in some categories. Wikipedia serves as foundational “ground truth” at 26.3%. A Morningstar study analyzing over 200,000 AI citations in the wealth management sector revealed that every question creates a different “leaderboard” — national media like CNBC and The Wall Street Journal dominate broad queries, appearing in over 20% of responses, while firm-specific pages only surface for local prompts. NerdWallet emerged as a category leader with a 38% citation frequency in financial queries.

The Pay-to-Play Pipeline

This creates what is effectively a pay-to-play system. Getting your brand mentioned in The Wall Street Journal, covered on CNBC, discussed positively on Reddit, and documented on Wikipedia requires either significant PR budgets or organic authority built over years. Smaller brands face a binary: invest heavily in earned media or accept that AI will never recommend you.

The pipeline becomes even more concerning when you examine the commercial deals that feed AI systems. The $60 million annual licensing deal between Google and Reddit, and OpenAI’s similar partnership, grant these AI companies real-time, structured access to Reddit’s user-authored content. These deals are framed as supporting “human learning,” but they create a mechanism where brands that successfully manipulate Reddit consensus — through coordinated upvoting, influencer seeding, or “organic” community building — achieve higher visibility in AI responses.

The economics are straightforward: traditional search advertising was transparent. You paid Google for an ad, it appeared next to search results, and users knew it was an ad. In the Citation Economy, the influence is invisible. When ChatGPT recommends a product because it appeared in a Reddit thread that was seeded by a marketing agency, neither the AI nor the user recognizes the manipulation. The recommendation arrives with the authority of an unbiased analysis.

There are also emerging concerns about direct commercial bias. If an AI platform receives a share of a transaction from one merchant but not another, the incentive to surface the partnered merchant in conversational responses is obvious. This risk of what critics call “enshittification” — where commercial bias eventually overwhelms user value — mirrors the trajectory of social media algorithms. Additionally, the “WEIRD” bias (Western, Educated, Industrialized, Rich, and Democratic) encoded in LLMs trained predominantly on English-language Western content systematically disadvantages brands and consumers from non-Western markets.

Cognitive Bias Exploitation: The AI’s Blind Spots Are a Feature, Not a Bug

Language models do not think. They pattern-match. And the patterns they have learned include every cognitive bias present in their training data — biases that sophisticated marketers are already exploiting.

Research presented at EMNLP 2025 demonstrated that LLMs exhibit measurable sensitivity to social proof, anchoring, order bias, and the decoy effect when making product recommendations. These are not abstract vulnerabilities. They are actionable levers that brands use to tilt AI outputs in their favor.

How Each Bias Gets Exploited

Social Proof Manipulation is the most straightforward. Descriptions flooded with claims like “bestselling,” “500+ bought this month,” and “thousands of happy customers” consistently boost LLM recommendation rates. The AI does not verify these claims. It processes them as signals of quality and adjusts its output accordingly. A brand that systematically embeds social proof language across its web presence — product pages, press releases, review responses — trains the AI to associate it with popularity.

Anchoring exploits the AI’s tendency to use the first piece of information it encounters as a reference point. Strategic price comparisons — showing an inflated “original” price next to a discounted price — influence the model’s perception of value. Product descriptions that lead with premium positioning create an anchor that persists through the AI’s synthesis process.

Order Bias means that the sequence in which options are presented affects how the AI evaluates them. Brands that control comparison content — “Top 10” lists, feature matrices, versus pages — can influence which product the AI encounters first when retrieving information.

Information Gain Manipulation is subtler and more powerful. AI systems prioritize content that provides new facts — information that does not already exist in their training data. Forbes reported that brands publishing proprietary survey results, original data tables, and unique research become the “source of truth” in a sea of synthesized information. If you are the only source for a specific statistic, the AI has no choice but to cite you.

The Bias Runs Deeper Than Product Recommendations

A study published in PNAS Nexus found that LLMs instructed to score resumes with randomized social identities exhibited significant racial biases, awarding lower scores to specific demographic groups despite comparable qualifications. These biases produced 1-3 percentage-point differences in hiring probabilities and were consistent across multiple models.

The relevance for brand visibility is direct: if AI models carry systematic biases about what “quality” or “authority” looks like — biases inherited from training data that overrepresents certain types of companies, institutions, and voices — then brands that match those internalized templates receive a structural advantage that no amount of content optimization can overcome for brands that do not.

What Smaller Brands Can Actually Do About This

The picture painted so far is bleak for smaller brands. The structural advantages are real, the manipulation is documented, and the Citation Economy favors the already-powerful. But the research also identifies concrete strategies that brands without Fortune 500 budgets can deploy. None of them are easy, but they work — and they represent the difference between visibility and invisibility in the AI era.

Step 1: Manage the Synthetic Content Data Layer

The concept of the “Synthetic Content Data Layer” — articulated by Hobo Web — refers to the space where AI systems make inferences based on fragmented or missing data. If accurate information about your brand does not exist in the places AI looks, the model fills the gap with guesses. Those guesses become your brand identity in AI search.

Managing this layer starts with auditing. Ask Google Gemini, ChatGPT, and Perplexity the questions your customers ask. “What is [your brand]?” “Is [your brand] reliable?” “How does [your brand] compare to [competitor]?” Document every inaccuracy, every omission, every instance where the AI recommends a competitor instead of you. This is your baseline.

Then generate what the research calls “definitive dossiers” — comprehensive documents built from your own authoritative internal data: technical specifications, user manuals, case studies, support logs, and original research. Publish these on your website in a format AI systems can easily parse and extract. The goal is to become the path of least resistance for the AI’s retrieval system.

Step 2: Semantic Chunking for Machine Extraction

AI models do not read pages the way humans do. They parse content for discrete units of meaning — specific answers to specific questions. G2’s research on LLM seeding demonstrates that structuring content around this behavior dramatically improves citation rates.

The technique is semantic chunking: using H2 headers as direct questions (“How does [product] reduce [problem]?”) and placing a 2-sentence summary directly under each header — what researchers call the “answer-first” format. AI models often extract the first 1-2 sentences under a heading as the answer. If those sentences contain a clear, factual, citation-worthy statement, you increase the probability that the AI will use your content as its source.

This is a structural investment that Taptwice Global applies across entity SEO work: making every page’s architecture machine-readable so that AI systems can extract and cite individual sections, not just entire pages.

Step 3: Information Gain — Become the Only Source

AI systems prioritize content that tells them something new. If your page repeats the same information available on fifty other websites, the AI has no reason to cite you specifically. But if you publish data that exists nowhere else — original survey results, proprietary benchmarks, unique case studies with specific numbers — you become the sole source for that information. The AI must cite you or omit the data entirely.

This is the “Information Gain” strategy, and it is the single most powerful tactic available to smaller brands. It does not require a large PR budget or structural web authority. It requires having something genuinely valuable to say and publishing it in a machine-readable format.

Step 4: LLM Seeding Across High-Authority Platforms

If AI systems learn from the web, then placing your brand’s narrative in the places AI systems read is not manipulation — it is distribution. Substack, Medium, niche industry forums, Quora, Reddit (authentically), LinkedIn articles — these platforms are all indexed by Common Crawl and consumed by AI training pipelines.

The goal is not spam. It is consistent, authoritative presence across the ecosystem the AI trusts. If you have written something explaining what AI search optimization is, a version of that perspective should exist on at least three platforms where AI models look for consensus.

Step 5: Data Hygiene

This is the least glamorous tactic and possibly the most impactful. Inconsistent product data across retailers — different prices, conflicting specifications, varying descriptions — creates signals that AI systems interpret as unreliability. Backlinko’s research confirms that conflicts in price and specifications actively deprioritize products in AI recommendations. Maintaining consistent SKU data, pricing, and product descriptions across every touchpoint is a prerequisite for AI visibility.

Measuring AI Visibility: The Metrics That Actually Matter

Traditional analytics — page views, bounce rate, click-through rate — were built for a world where users visited your website. In a world where AI answers questions without sending users anywhere, those metrics tell an increasingly incomplete story. Success in GEO requires a different measurement framework entirely.

Share of Model (SoM)

Share of Model is the AI-era version of market share. It measures the percentage of times your brand appears in AI answers for a set of strategic keywords relative to your competitors. If you ask ChatGPT “best project management tools” one hundred times and your brand appears in twelve of those answers while your competitor appears in forty-three, your SoM is 12% and theirs is 43%. This is the number that determines whether you exist in the AI discovery layer.

Citation Rate and Authority

Citation Rate tracks how often AI models link to your website as a source within their responses. Citation Authority goes further: it measures the trust tier of the domains that cite you. A citation from Wikipedia carries fundamentally different weight than a citation from an obscure blog. Brands whose primary citations come from high-authority sources build what the research calls “baseline trust” — a self-reinforcing cycle where trust begets more citations, which begets more trust.

Mention Sentiment

The tone of an AI’s description of your brand directly affects conversion. A brand that is described as “reliable and widely recommended” occupies a different position than one described as “an option, though some users have reported concerns.” Mention Sentiment tracking — monitoring whether AI systems describe you positively, neutrally, or negatively — is essential for identifying both organic reputation shifts and potential GEO warfare attacks.

HC Rank as a Proxy Metric

Since Harmonic Centrality directly influences crawl frequency and training data representation, tracking your domain’s HC rank provides a scalable proxy for your structural position within the web graph. Improvements in HC rank predict improvements in AI visibility across future model versions.

For organizations with multiple business units, these metrics need to be tracked separately. A company’s cloud services division and its hardware division will have entirely different SoM profiles, citation sources, and sentiment distributions. The AI does not see “your company.” It sees individual entities associated with specific knowledge domains.

What This Means for the Future of Search

The evidence assembled here points to a future that is uncomfortable but important to confront honestly.

Structural Hegemony Will Deepen

The reliance on Harmonic Centrality within Common Crawl creates a feedback loop that benefits established brands with every new model training run. Their content is overrepresented in training data, which makes the AI more familiar with them, which makes the AI more likely to cite them, which drives more links to their content, which improves their HC rank, which ensures even greater representation in the next training run. This cycle does not self-correct. It compounds.

Memory Poisoning Will Get Worse Before It Gets Better

The feature that makes AI assistants personalized and useful — persistent memory — is the same feature that enables recommendation poisoning. As AI assistants become more deeply integrated into purchase decisions (booking travel, selecting financial products, choosing healthcare providers), the incentive to poison those memories increases. The tools are already commercially available. The defenses are not.

Commercial Bias Will Become the Default

As AI companies transition from research organizations to revenue-driven platforms, the pressure to monetize conversational interfaces through sponsored answers, transaction fees, and licensing deals will intensify. The trajectory mirrors what happened with social media: platforms that began as neutral utilities became advertising machines. The same economics apply to AI search, and the same outcome — where commercial interests overwhelm user value — is the default trajectory.

The Gap Will Widen

Perhaps most critically, the gap between the “algorithmic oligarchy” — brands with the resources to dominate HC rankings, seed media coverage, fund original research, and deploy memory injection at scale — and the “invisible long tail” of everyone else will widen. This is not a temporary market inefficiency. It is a structural feature of how AI systems are built and trained.

For businesses watching this unfold, the strategic imperative is clear: the window to establish AI visibility is open now, but it is closing. Every model training run that passes without your brand being adequately represented in training data is a compounding disadvantage. Every month without a presence in the citation sources AI trusts is market share ceded to competitors who are already there.

If you have not yet audited how AI systems describe your brand — or whether they mention you at all — start there. We have previously outlined why brands fail to appear in ChatGPT and the specific structural factors that determine AI visibility. The manipulation documented in this article makes that audit more urgent, not less.

The question is no longer whether AI search is biased. The research proves it is — structurally, commercially, and cognitively. The question is whether you will adapt to that reality or be rendered invisible by it.

Frequently Asked Questions

Can brands really manipulate what ChatGPT recommends?

Yes, and the evidence is documented by credible sources. Microsoft’s security team published a detailed report in February 2026 showing how hidden URL parameters in “Summarize with AI” buttons inject persistent memory commands into AI assistants. Separately, a July 2025 study by Reboot Online Marketing demonstrated that researchers manipulated ChatGPT and Perplexity recommendations within days by planting content on low-authority expired domains. Beyond these direct attacks, brands also influence AI recommendations indirectly through Harmonic Centrality gaming, earned media dominance, and cognitive bias exploitation in product descriptions.

What is Harmonic Centrality and why does it matter for AI visibility?

Harmonic Centrality (HC) is a mathematical measure of how well connected a domain is to every other domain on the web, based on shortest-path distances. Common Crawl — the archive that provides over 80% of training data for models like GPT-3 — uses HC to determine crawl priority. Domains with strong HC rankings get crawled more frequently, which means more of their content appears in AI training data. This builds “baseline familiarity” within the model, making the AI more likely to recommend those brands even before performing any real-time search. Domains outside the top one million in HC rank are crawled infrequently and are effectively invisible to AI systems.

Is AI search biased toward big brands?

The evidence strongly supports this conclusion across multiple dimensions. Structurally, Harmonic Centrality gives established brands disproportionate representation in training data. Economically, the Citation Economy rewards brands that can afford earned media in tier-one publications — 90% of citations driving LLM brand visibility come from earned media, according to Edelman. Cognitively, LLMs inherit social proof and anchoring biases from their training data, which favors brands with larger online footprints. These are not bugs being fixed. They are structural features of how AI systems are built, trained, and deployed.

How can smaller brands compete in the citation economy?

The most effective strategy is Information Gain — publishing original data, proprietary research, unique case studies with specific numbers, and insights that exist nowhere else on the web. When your content is the only source for a particular fact, AI systems must either cite you or omit the information. Beyond that, semantic chunking (structuring content as direct questions with answer-first formatting), consistent data hygiene across platforms, LLM seeding on high-authority platforms like Substack, Medium, and niche forums, and proactive management of the Synthetic Content Data Layer all contribute to AI visibility without requiring enterprise-level budgets.

What is GEO warfare and should I be worried about it?

GEO warfare refers to deliberate strategies competitors use to manipulate AI systems against your brand. This includes funding negative analyst reports, seeding critical discussions on Reddit and forums to create false consensus, flooding high-authority publications with competitor-favorable content to crowd out your perspective, and coordinating FUD (Fear, Uncertainty, and Doubt) campaigns that cause AI systems to return cautious or negative answers about your brand. If you operate in a competitive market — especially enterprise software, financial services, or healthcare — you should be monitoring how AI systems describe your brand and watching for sudden negative shifts that may indicate coordinated manipulation.

Are AI companies doing anything to prevent manipulation?

AI companies are aware of these vulnerabilities, and some efforts are underway. Microsoft’s publication of the AI Recommendation Poisoning report indicates at least some transparency about the threat. However, the fundamental architecture of these systems — reliance on Common Crawl, training on user-generated content from Reddit, persistent memory features, and RAG retrieval from the open web — creates vulnerabilities that cannot be patched without fundamental redesign. Meanwhile, the commercial incentives (licensing deals with Reddit, potential transaction fees, advertising integration) actively work against neutrality. The realistic outlook is that manipulation will continue to be possible and profitable for the foreseeable future.