Why ChatGPT and Perplexity disagree about your business
We track LLM visibility for hundreds of businesses. The four major models almost never agree. The disagreement is the point — and it's how you know what to fix.
We've tracked LLM visibility across roughly 200 businesses for the better part of a year through the AI Mention Tracker. The single most consistent finding: the four major models — ChatGPT, Claude, Gemini, Perplexity — almost never agree about who to recommend.
A typical Visibility Report from this week, anonymized:
Business: regional landscaping company, 8 locations
Visibility score: 41% across 32 keywords
ChatGPT: 62% mentioned (20/32)
Perplexity: 0% mentioned (0/32)
Gemini: 56% mentioned (18/32)
Claude: 47% mentioned (15/32)
That Perplexity zero is real. We see it constantly. It is not a bug.
Why the disagreement happens
The four models are not asking the same questions of the same data. They differ on five axes:
Training data freshness. Claude and ChatGPT have meaningfully different cutoffs and refresh cadences. A business that launched in late 2024 will appear in some models 6+ months earlier than others.
Retrieval architecture. Perplexity is heavily retrieval-augmented — it's pulling live web results and grounding its answers in them. ChatGPT (especially with browsing enabled) does some of this. Gemini and Claude use different retrieval pipelines depending on the surface. Your visibility in Perplexity is partly a function of whether your domain ranks in real-time web results, not just whether you exist in training data.
Source weighting. Different models weight different sources differently. Yelp matters a lot to Gemini in our testing. It matters less to Claude. Reddit references matter more to ChatGPT than to Perplexity. Local news domain mentions move Gemini in a way they don't move ChatGPT.
Schema preference. This is the one we obsess over. Claude consistently does a better job parsing nested JSON-LD with Service and Person schema. Gemini gives more weight to LocalBusiness and aggregateRating. ChatGPT seems to use schema as a reliability signal more than a primary content source. Perplexity uses it as a tiebreaker between otherwise-similar candidates.
Refusal posture. Some models will refuse to recommend specific businesses ("I can't endorse one over another") in some categories, especially regulated ones — legal, medical, financial. Refusal posture varies by category and by model.
Why the disagreement matters
Two reasons.
First, your customers are not all using the same model. Right now, ChatGPT has the largest user base, but Perplexity is growing fast among researchers and B2B buyers, Claude is increasingly the default in technical and professional services, and Gemini is integrated into Google's surface in ways the others aren't. The customer asking "best electrician near me" on each of those is a different customer.
Second, the disagreement tells you what to fix. A business that's strong on Gemini and weak on Perplexity has a real-time-web-presence problem. A business that's strong on Claude and weak on Gemini probably has a Yelp / Google Business Profile problem. A business that's weak on all four has a structured data and entity-establishment problem.
The Mention Tracker is built to surface exactly this — not as a single number, but as a per-model decomposition that tells you what to do this week.
Three patterns we keep seeing
After thousands of scans:
- Perplexity zero is most often a domain authority problem, not a content problem. Newer domains with good schema still get filtered out of Perplexity's retrieval set. Time and link-building fix it; nothing else does, fast.
- Gemini disagrees most often with ChatGPT on businesses without strong reviews. If you have schema and a website but few or no Google reviews, Gemini is much more skeptical than ChatGPT. ChatGPT will recommend you off training data alone. Gemini wants the social proof.
- Claude consistently rewards specificity over breadth. A practice with a few highly-detailed service pages outperforms a practice with many shallow ones. This holds even controlling for word count.
What to do this week
If you've never measured your LLM visibility, you can do it for free at ai.relevantdomain.co. It takes about a minute to set up and your first scan is free.
What you'll likely find: a number lower than you expected, and a per-model breakdown that disagrees with itself in interesting ways. That disagreement is your roadmap.
