We analyzed 33,000 AI evaluations across ChatGPT, Claude, Perplexity, and Gemini for 47 B2B SaaS companies. 72% had at least one verifiably wrong factual claim about their own company in an AI response. Not mildly unfavorable framing. Not a missing nuance. A wrong fact stated with confidence.

The errors weren’t distributed randomly. When we pulled the full error set and categorized every misrepresentation, five patterns accounted for the vast majority of them. Each pattern has a different cause. Each has a different fix. And each one is showing up in AI responses that 51% of B2B buyers now use to start their vendor research.

That’s what this study covers. Where the errors come from, which companies are most exposed, and what the content structure is that makes a company more or less vulnerable to each one.

Stale pricing is the most common error

Pricing errors accounted for 30% of all factual errors in our dataset, making them the single most frequent category by a wide margin.

The mechanism is straightforward. A company restructures pricing in Q1. The change goes live on their website. The AI models that don’t ground answers in live web search, primarily ChatGPT and Claude, continue quoting the pre-restructure numbers for months. When we benchmarked one compliance automation company that had dropped its entry price by 40% six months before the run, all four models were quoting the old number. The free tier it had eliminated a year earlier was still being described as an option.

This matters more than most errors because pricing is binary in a buyer’s decision process. A misquoted feature creates ambiguity. A misquoted price creates a hard filter. The buyer is working with a budget. The AI says you’re $2,000 per month. The budget is $1,500 per month. The buyer moves on without knowing you restructured pricing specifically to win that segment.

Companies that have recently restructured pricing, launched new tiers, or sunset free plans are most exposed. The AI’s internal knowledge is a snapshot, and the gap between a company’s current pricing and the AI’s cached version tends to grow quietly until it’s actively tested. Perplexity, because it grounds answers in live search, catches pricing changes faster — but only when the pricing page is indexed and structured clearly enough to be scraped and interpreted accurately.

The fix is specific: a pricing page that is clear, machine-readable, and updated with enough context that it outcompetes cached descriptions. Statistics and structured data in content improve AI visibility by 41% according to Princeton’s GEO research. A pricing page that explains what changed and why is harder for an AI to override with stale information than one that just lists the current numbers without context.

Feature misattribution affects 43% of companies

The second error pattern was features attributed to the wrong company. 43% of the companies we benchmarked had at least one evaluation where the AI credited their capabilities to a competitor, or credited a competitor’s capabilities to them.

This happens when companies use identical marketing language. “API-first,” “real-time sync,” “enterprise-grade security,” “native integrations” — these are the phrases that cause misattribution. When every company in a category uses the same four adjectives to describe functionally different architectures, the AI has no basis for distinguishing them. It builds a composite description that blends capabilities across companies, and the resulting feature comparison table looks coherent while the attributions underneath it are wrong.

We benchmarked two competing developer tools that both described their core architecture as “event-driven.” In three separate evaluations, Claude attributed Company A’s webhook-based architecture to Company B and vice versa. The AI had no way to distinguish them from their descriptions alone.

The resulting feature comparison tables in AI responses are the most damaging version of this error. A buyer reads a side-by-side comparison. Every row is specific. Every cell has a claim. The AI stated it confidently. None of the attributions are flagged as uncertain. But the columns are effectively swapped.

Feature misattribution concentrates in categories where the vocabulary has converged. Security tools, data platforms, developer infrastructure, and compliance software are the most affected sectors in our dataset. The fix requires more than rewriting marketing copy. It requires differentiated, specific descriptions of how your architecture works — not what category it belongs to, but the mechanism that makes it distinct. Feature-specific documentation pages that AI can extract from without ambiguity are far more protective than a homepage that uses the same adjectives every competitor uses.

Competitive framing from opponent content

28% of the companies we benchmarked had their competitive framing built primarily from competitor-owned content.

This is the hardest error to detect because it doesn’t look like an error. The AI doesn’t say anything technically false. It frames the comparison using the dimensions the competitor established on their “Us vs. Them” page, and those dimensions happen to favor the competitor. The losing company’s differentiators aren’t wrong — they’re absent, because those differentiators weren’t in the competitor’s framing.

Zero-click rates for AI queries run at 83%, which means the framing an AI response establishes is often the only framing a buyer encounters. When the competitor has a structured comparison page optimized for AI extraction, and the evaluated company doesn’t, the competitor wins the framing before the buyer ever visits either site.

Citations function as ownership claims. A competitor who publishes one highly structured comparison page can exert more influence on an AI recommendation than a company with ten times the content volume, if that content volume doesn’t directly address the buyer’s comparative question.

The implication is blunt. If you haven’t published a comparison page that addresses “Company A vs. Company B” from your frame, the AI will build that comparison from whatever source exists — and the most likely existing source is your competitor’s page. This pattern is concentrated in smaller companies. Companies with fewer than 50 indexed pages have a 3.1x higher competitive framing error rate than companies with more comprehensive content coverage.

Founding date and funding fabrication

Full fabrication, in the sense of the AI inventing details from nothing, was less common than the other patterns. It showed up in 15% of the companies we benchmarked, and it was heavily concentrated in smaller companies with sparse public presence.

When a model lacks sufficient training data about a company, it interpolates. The interpolation isn’t random — it draws from adjacent companies with similar names, similar sectors, or similar size profiles. One startup in our dataset was described by ChatGPT as “founded in 2018 and headquartered in Austin” when they were founded in 2021 in San Francisco. The AI had apparently merged their profile with a similarly-named company in an adjacent category.

Founding dates and funding amounts are the most commonly fabricated facts in our data. They’re also the hardest to correct because they don’t live on pages that AI models are grounding against in real time. They live in the model’s training data, which updates on a much longer cycle.

The defense is presence, not correction. Companies with thin public footprints are most vulnerable because there’s nothing to anchor the AI’s representation against. When the model has ten authoritative sources that all say the same founding date, interpolation becomes less likely. When it has one or zero, it fills gaps with its best guess.

The freshness gap: when do AI models update?

The five error patterns we’ve described all share a common underlying driver: the gap between when a fact is true and when an AI model reflects it.

Different models close this gap at very different speeds. Perplexity grounds answers in live web search by default, which means it can catch a pricing change within days if the pricing page is indexed clearly. ChatGPT and Claude have the longest lag — major knowledge updates take months, and the specific timing of when any individual fact propagates is opaque even to the model providers. Feature launches take three to six months to appear reliably in non-search-grounded models, which means a company can launch a significant product update and have it be absent from the AI responses that buyers encounter for an entire sales cycle.

AI search traffic grew 527% year-over-year according to Semrush. The evaluations buyers are running on AI today are based on facts that may be six to eighteen months out of date for the models without live search grounding. For companies in fast-moving categories, the content they published during a period of rapid change may be the worst representation of where they are now — and it’s the representation AI has locked in.

Tracking which models reflect which facts, and measuring how that changes after content updates, is now a measurable exercise rather than a guessing game. The per-model freshness dynamics vary enough that treating all AI responses as equivalent underestimates the problem.

What makes a company more vulnerable?

When we looked at which companies had the highest error rates, the variables that predicted vulnerability weren’t brand size or category competitiveness. They were content structure and content recency.

Companies with fewer than 50 indexed pages had a 2.3x higher error rate than companies with more comprehensive coverage. The AI doesn’t have enough authoritative source material to anchor its representation, so it relies on whatever it has — which is often partial, outdated, or borrowed from adjacent companies.

Companies that hadn’t published new content in 90 or more days had a 1.8x higher error rate. This one surprised us. We expected staleness to matter, but the relationship was stronger than anticipated. Fresh content signals to search-grounded models that the company is actively maintaining its public presence. For models that don’t use live search, new content still enters the training pipeline faster than old content gets refreshed.

Companies without comparison pages had a 3.1x higher competitive framing error rate specifically. The comparison page gap was the single strongest predictor of competitive framing errors in our dataset, and it’s the most addressable. A company that publishes one well-structured comparison page for each of its two or three main competitors closes the largest single vulnerability in their AI representation.

69% of B2B buyers changed their vendor selection based on AI guidance, and AI-referred traffic converts at 4.4x the rate of other channels. The buyers arriving from AI evaluations are high-intent and largely pre-qualified. The companies that lose those buyers to factual errors aren’t losing soft interest — they’re losing buyers who were already close to a decision.

The fix isn’t “more content” — it’s the right content

Adding content volume doesn’t close AI accuracy gaps. The 72% error rate isn’t a result of companies not publishing enough. It’s a result of companies not publishing the specific content types that anchor AI representation against the exact claims buyers encounter.

Pricing pages that clearly state current pricing, with enough context that an AI can’t confuse them with cached descriptions. Feature documentation pages that are specific enough that the AI can distinguish your capabilities from a competitor’s. Comparison pages that establish the evaluation frame from your perspective before a competitor does it for you.

This is what our benchmark surfaces — not a general AI presence score, but the specific errors in each model’s current representation of your company, the content gaps that are causing each error, and the priority order based on which errors are driving the most recommendation losses. The benchmark hands you misrepresentations with receipts: each wrong claim shown next to the source that contradicts it, so the fix is never a guess.

If you want to know which of the five patterns your company is currently exposed to, run a benchmark. The error rate in our dataset is 72%. The odds that your company’s AI representation is currently accurate across all four models are lower than most marketing teams expect.

For companies that want to work through what the fixes look like in practice, the AI presence improvement guide covers the content structure behind each error type.