AI hallucinates your pricing because your data isn't anchored. Three fixes: put pricing in HTML tables (not paragraphs), use explicit negation to deny features you don't have, and timestamp your truth so the AI prioritizes fresh data. PDFs are kryptonite. Mirror their content as HTML.
Imagine this. A potential customer opens ChatGPT and asks "Does [Your Brand] offer enterprise support?" The AI confidently replies: "No, [Your Brand] is strictly for small businesses and doesn't offer SLA support."
But you do offer enterprise support. It's your highest-margin product. You have a PDF brochure about it on your site. You have a blog post about it. The AI just told your customer to go to your competitor.
This isn't theoretical. As search shifts from links (Google) to answers (AI), the risk of brand hallucination has exploded. AI models aren't malicious. They're stochastic parrots. If they can't find a clear, structured fact to anchor their response, they statistically guess, and often guess wrong.
In our companion guide, Your About Us Page Is Useless: Why You Need an Entity Home, we covered how to build an Entity Home so the AI knows who you are. This guide teaches you how to protect what you say: the mechanics of data anchoring, the technical defense that ensures ChatGPT, Perplexity, and Gemini tell the truth about your pricing, features, and policies.
The Physics of the Lie: Why Hallucinations Happen
To fix the problem, understand the mechanics of failure. LLMs don't read your website like a human. They ingest it, tokenize it, turn it into vectors. When a user asks a question, the model looks for the most probable next word. The deeper mechanism (prediction vs retrieval) is in our breakdown of Hallucinations vs Reality.
Hallucinations about pricing and features usually stem from three structural failures on your website:
The adjacency failure is the same chunk-severance mechanism we documented in The Semantic Schism: when the price and its label land in different chunks, the price becomes orphaned data.
The Defense Strategy: Data Anchoring
You can't stop AI from guessing. You can force it to guess correctly by providing data anchors: content so structurally rigid the AI prefers it over its own probability model. Three anchors.
If you have a price, a feature list, or a return policy, it has to live in an HTML <table>. LLMs are trained heavily on code and structured data. They trust the horizontal relationship of a table row explicitly.
A paragraph saying "Our Basic plan is $10 and includes X, while Pro is $20 and includes Y."
A table where Row 1 is "Basic" and Column 2 is "$10."
Audit your pricing and features pages. If you're using fancy CSS grids or images to display prices, stop. Hard-code them into semantic HTML tables for the crawler, but render them cleanly to avoid the token tax that can break RAG parsing of bloated table markup.
AI often hallucinates features you don't have because it assumes you're like your competitors. If all your competitors offer a free trial, the AI might assume you do too. Stop it with explicit negation: sentences that clearly state what you do not offer.
"We offer paid plans starting at $50." (The AI might still wonder if there's a free tier.)
"We do not offer a free trial. All plans are paid subscriptions."
Place these negative anchors in your FAQ section. They give the RAG agent a direct "no" to retrieve when a user asks "Is there a free trial?"
The ghost data problem (AI citing old prices) is massive. Fix it by time-stamping your core data. Don't just write "$49/month." Write:
Current Pricing (Updated June 2025): $49/month.
When an AI compares conflicting data points (your 2019 blog vs your 2025 pricing page), the recent date acts as a freshness signal. Most modern models are biased to prioritize data associated with recent dates.
The PDF Jailbreak
Many B2B companies hide their best data (case studies, white papers, technical specs) inside PDFs. Disaster for GEO. If your enterprise security spec is a PDF, the AI likely won't read it deep enough to answer a specific question about your encryption standards.
The fix: adopt the HTML-First publishing model. Keep the PDF for humans who want to download it, but also publish the full text as a standard HTML web page (/specs/enterprise-security), and wrap that page in Article or TechArticle schema. This ensures the brains of your content are accessible to the eyes of the AI, the same rendering principle from The Invisible Website.
Find out what AI is saying about your pricing right now.
Free audit. Detects PDF-trapped data, unanchored prices, and ghost-data conflicts that trigger hallucinations.
Run a brand-safety audit free →Verification is the New Reputation Management
In the old world, reputation management meant checking Yelp reviews. In 2025, it means auditing AI answers.
You can't afford to let a robot guess your pricing. You have to feed it. Tables, explicit negation, and HTML-first content turn your website from a "probabilities source" into a "facts source."
Now that you've built your Entity Home and secured your brand safety, the final question is: is it working? Our companion guide on Share of Model vs Rank Tracking covers the new metrics of the AI age: why rank tracking is dead and how to measure whether you're truly winning the algorithm war.
References & Further Reading
- Anthropic Research: Constitutional AI and Hallucination. Insights into why models hallucinate and how grounded data reduces error rates. https://www.anthropic.com/research
- Google Search Central: Best Practices for E-commerce Structuring. Why tables and structured data are preferred for product data. https://developers.google.com/search/docs/specialty/ecommerce
- Website AI Score: The RAG Chunking Mismatch. Our technical guide on how formatting affects AI retrieval. https://websiteaiscore.com/blog/rag-chunking-mismatch-formatting-guide
