The one finding that reframes everything
Winning one engine tells you almost nothing about the others. In our family-law panels, a single firm was the runaway leader on ChatGPT and named zero times on Gemini and Perplexity across sixty combined runs. Another firm led Gemini and was invisible on ChatGPT. Published research on cross-model agreement lands in the same place: the engines name the same top business fewer than half the time. So the useful question isn’t “am I visible in AI?” It’s “which engine, for which question, from which source?”
ChatGPT: directories decide the shortlist
On ChatGPT, complete and well-reviewed directory profiles carry the recommendation. In legal panels, nearly every ChatGPT shortlist was built on directory profiles; a firm with a strong profile reached the top recommendation repeatedly, while firms with thin profiles didn’t appear — regardless of actual reputation. ChatGPT also runs a separate in-chat map widget, fed by different local data, whose firm set differs from its prose answer. The lever: own your category’s authoritative directories, and treat the map widget as its own hygiene task.
Gemini: your Google profile and its own memory
Gemini leans on Google Business Profile data and a strong parametric preference for a few entrenched names it already “knows” in a market. Its answers frequently carry no resolvable citations at all — the recommendation comes from the model’s memory more than from live retrieval. That makes Gemini a knowledge-and-entity problem: consistent business information, an authoritative Google profile, and third-party corroboration move it, not a single new web page.
Perplexity: it retrieves broadly, but only quotes what’s quotable
Perplexity cites more sources than any other engine, drawing on vertical directories and businesses’ own pages. But we repeatedly watched it cite a business’s website as a source and never name the business in the answer — the page was retrieved and found nothing extractable to quote. Getting selected is only half the job on Perplexity; the other half is having a standalone, quotable sentence on the page.
Claude: review mass in one mode, credentials in another
Claude alternates between two behaviors on the same question. In one mode it renders Google review cards and ranks almost purely by star score and review volume — a four-star firm with a few dozen reviews loses to a five-star wall. In the other mode it synthesizes verifiable credentials and openly editorializes against directories it considers low-quality. Neither mode overlaps with ChatGPT’s directory dependence, which is why a business can win ChatGPT and lose Claude on identical assets.
Google’s AI surfaces: your own content, if it’s concrete
Google AI Overviews and AI Mode ground answers in a mix of directories, review sites, and — crucially — businesses’ own educational and pricing pages. AI Mode in particular runs a deep multi-site crawl per answer and maps your content semantically, which lets a content-strong, directory-weak business win there first. The repeated pattern across all the content-reading surfaces: the business that publishes real numbers wins the cost questions.
What to do with this
Stop optimizing “for AI” as if it were one thing. Map each engine to its lever — directories for ChatGPT, your Google profile for Gemini, quotable pages for Perplexity, reviews and credentials for Claude, concrete-fact content for Google’s AI — and fix them in the order your customers actually search. That per-engine mapping, for your business and market, is exactly what a Cited panel produces.