TL;DR
- AI assistants get discrete factual queries wrong at a high rate — the Tow Center for Digital Journalism found error rates above 60% across several AI search engines on verifiable facts, and JMIR measured GPT-4's hallucination rate on verifiable claims at roughly 28.6%.
- Pricing, founding year, HQ location, and integration lists are the four categories where brand facts break down most often, with pricing and integrations tied closely to gated or unstructured content.
- Smaller and lesser-known brands see structurally higher error rates — there's simply less corroborating text for a model to triangulate against.
- Structured entity pages (not just inline JSON-LD snippets) correlate with meaningfully better accuracy than pages lacking them.
- Google, Microsoft, and OpenAI have each confirmed that schema markup feeds their generative and product-citation features, which is a direct lever brands can pull.
AI hallucination of brand facts refers to AI assistants stating incorrect pricing, founding dates, headquarters, feature lists, or integration details about a company when answering buyer questions — not because the model is "lying," but because it's pattern-matching against incomplete or conflicting source material and producing a fluent, confident answer regardless of accuracy.
Where the errors actually come from
Large language models don't query a live database when asked "what does X cost" or "when was Y founded." They generate the most statistically plausible continuation based on what they were trained on, plus whatever retrieval context a tool call pulls in at answer time. When a brand's own site is the only source and that source is unstructured or paywalled, the model falls back on older training data, a stale blog post from three years ago, a user forum post, or its own best guess — and presents all of these with the same tone of certainty.
This is why the Tow Center research is a useful gut check for anyone assuming AI answers are closer to search results than to essays: discrete facts (a date, a dollar figure, a city) fail more often than general descriptive claims, because there's no single correct phrasing to converge on and because retrieval often surfaces outdated or third-party restatements instead of the canonical source.
The four categories where brand facts break down
Across the kinds of questions buyers actually ask — "how much does it cost," "when was it founded," "where is it based," "does it integrate with Salesforce" — four error types show up repeatedly:
- Pricing errors. Models cite outdated tiers, round numbers that don't match actual plans, or invent a "free forever" plan that doesn't exist. This tracks closely with pricing-page readability — if pricing lives behind a "contact sales" form or is rendered client-side, there is no clean text for a crawler to extract.
- Founding details. Year founded, founder names, and funding history get conflated with similarly named companies or confused across rebrands and acquisitions.
- Invented features. Models describe integrations or capabilities that sound plausible for the product category but don't exist for that specific brand — a byproduct of pattern-matching against competitors rather than the actual company.
- Competitor conflation. Smaller brands get attributed features, pricing, or history belonging to a better-known competitor in the same space, especially when the smaller brand's own content doesn't explicitly disambiguate itself.
This lines up with what comparison pages are good for — explicit, named contrast reduces the chance a model merges two brands into one answer.
Why smaller brands get hit harder
Accuracy on these questions tends to correlate with how much corroborating text exists about a brand across the web. A well-known SaaS company has analyst coverage, review-site profiles, press mentions, and a Wikipedia page all reinforcing the same facts, so a model has multiple consistent signals to draw from. A smaller or newer brand often has only its own marketing site, and if that site's facts live in a hero banner image, an FAQ accordion that requires a click, or a PDF pricing sheet, there's effectively no clean text for retrieval systems to pull from at all. The gap isn't really about brand size — it's about how many independent, structured, text-based sources exist. This is consistent with findings in why AI cites Reddit and Wikipedia more than your homepage: third-party, text-heavy, well-structured sources often out-compete a brand's own page for citation.
The schema correlation
Structured entity pages — not just a JSON-LD block bolted onto a homepage, but dedicated About and Organization pages with consistent, explicit facts stated in plain text — correlate with roughly a 30% accuracy gain in evaluations of AI-generated answers about brands. This isn't a Google ranking artifact; it reflects that structured, unambiguous facts are easier for retrieval and generation pipelines to extract and quote faithfully than prose buried in a narrative paragraph.
This is reinforced by public statements from both Google and Microsoft in March 2025 confirming that Schema Markup feeds their generative search features, and separately by OpenAI confirming ChatGPT uses structured data for product-related results. None of these confirmations guarantee citation or ranking, but they establish that schema is part of the input pipeline these systems draw from — which is a very different claim than "schema gets you cited."
Organization schema with accurate sameAs links, a dedicated About page written as an entity document, and plain-text pricing are the concrete levers here — see Organization schema + sameAs and Schema.org markup for AI search for implementation detail.
What to actually fix, in order
Start by finding out what AI assistants are currently saying about your pricing, founding details, HQ, and integrations — not what you assume they're saying. CiteFlow's free scan audits how AI assistants crawl, understand, and cite your site across seven dimensions and shows a sample of real AI answers citing (or not citing) your URLs, which is the fastest way to see where your own facts are getting garbled before you guess at a fix.
Once you know where the gaps are: unlock pricing from behind forms and render it as plain text, build or correct an Organization schema block with accurate sameAs links, and write an About page that states founding year, HQ, and leadership in plain declarative sentences rather than narrative copy. These are mechanical fixes, not ranking promises — they reduce the raw material available for a model to hallucinate from.
FAQ
Does fixing schema guarantee AI assistants will cite my facts correctly?
No. Schema and structured pages give models cleaner source material to draw from, which correlates with lower error rates, but no vendor or site owner can guarantee a specific model's output on any given query.
Which fact category is most commonly wrong — pricing or founding year?
Both show up frequently, but pricing errors are more directly tied to content accessibility (gated or JS-rendered pricing pages), while founding-year errors are more often a byproduct of conflation with similarly named or acquired companies.
Is this hallucination problem specific to smaller brands?
Smaller and lesser-known brands see it more often because there's less corroborating text across the web to reinforce a single consistent fact set, but well-known brands still get pricing and feature details wrong, especially after a rebrand or pricing change.
Sources
Drafted with AI assistance and published by CiteFlow. Sources are linked in the text.



