Why AI recommends some companies — and ignores others
There's a comforting story where you optimize your website and the machines reward you. It's mostly wrong — and the truth about what they actually reward is stranger, and more humbling, than the story.
AI engines don't rank you. They assemble an answer, then reach for sources they can find, read, and corroborate. Whether your name makes it into the sentence depends less on your website than on whether the rest of the internet independently backs you up — and the engines disagree with each other about that, constantly.
Five AIs, five different answers
Start here, because it's the part nobody warns you about: there is no single "AI opinion" of your company to win. Ask the same buyer question across ChatGPT, Claude, Gemini, Perplexity, and Grok and you'll routinely get five different shortlists — different names, different sources, the occasional refusal to answer at all.
This isn't anecdotal. In one analysis of roughly 680 million citations, only about 11% of the domains ChatGPT cited were also cited by Perplexity.1 The engines don't share a worldview; each has its own taste in who counts as a source. So a single screenshot of ChatGPT naming your competitor tells you almost nothing — you might be winning Perplexity and invisible in Gemini, for reasons that have nothing to do with each other. "Visible in AI" isn't one state. It's five, and they don't agree.
Unlearn ranking
The next thing to drop is the mental model of a ranked list. Classic search hands you ten links and lets you choose. The model skips the list — it reads, decides, and recommends, in two steps people constantly blur together. First it retrieves a set of pages it could use. Then it cites the few it actually quotes and names.2 "I rank #1 on Google" buys a ticket to step one. It does nothing for step two. That's the whole game, and it's a different game.
The part that stings: it's mostly not your website
Here's the finding most marketing teams aren't braced for. When Ahrefs studied 75,000 brands, the strongest correlate of getting cited by AI wasn't anything on the brand's own site. Unlinked brand mentions — other people talking about you, no link required — correlated with citations at 0.664. Backlinks, the thing the SEO industry has ground on for twenty years, came in at 0.218. Roughly a 3× gap, in favor of conversations you don't control.3
And domain authority — the number your agency has lovingly fattened for a decade? A separate analysis pegs it as predicting under 4% of AI citations.4 Sit with that for a second before you renew the retainer.
This isn't the engines being mystical. A model answering "who's good at X?" is doing something closer to asking around than ranking. It wants agreement across independent sources before it'll commit your name to a sentence it has to stand behind. Your homepage is one source — and the one most obviously paid to say nice things about you.
The three questions behind every answer
Underneath all of it, a company has to clear three gates to get named. We score them formally as Content, Retrieval, and Trust, but in plain language they're three questions, and they're sequential — fail an early one and the rest don't matter.
Can it find and read you? retrieval
The least glamorous gate, and the one that silently disqualifies the most companies. AI crawlers generally don't run JavaScript — they see raw HTML. If your best content renders client-side or lives in a PDF, it effectively doesn't exist to the engine. You can be the sharpest firm in the category and be illegible. More →
Is there an answer to lift? content
The engine needs a clean, quotable passage that answers the question. Self-contained answers of roughly 134–167 words, with the point up top, are the unit that gets cited — about 4× more often than pages that bury it.5 A 3,000-word think-piece with the insight in paragraph nineteen is, to a model, unquotable. More →
Does it trust you enough to say your name? trust
The gate from the chart above, and the one that decides the close calls. Can the engine resolve you to a real, corroborated entity — credentialed authors, a presence in the knowledge graphs, third-party coverage, consistent facts across the web? When two firms are otherwise equal, the one the rest of the internet independently agrees exists gets named. Slowest to build, most decisive. More →
Back to the disagreement — and why it's the real story
Now the divergence makes sense. Each engine weighs those three questions against a different library of sources, so they reach different verdicts:
| Engine | Leans on | So it rewards |
|---|---|---|
| ChatGPT | Wikipedia, encyclopedic and major-news sources; fresh pages | Resolvable entities, recency, review-platform presence |
| Perplexity | Reddit and community discussion, the real-time web | Lived experience, being talked about by actual people |
| Google AI Overviews | Existing top-ranked organic results | Classic SEO strength, carried into the answer layer |
Directional, drawn from published source-mix analyses; exact proportions shift constantly. The point isn't the decimals — it's that "AI visibility" is five different problems wearing one trench coat.
Which is the uncomfortable, and oddly freeing, conclusion: there's no single number to chase. There's a per-engine reality, and the only way to know yours is to ask all five — with your real buyer questions — and read what comes back.
A model answering "who's good at this?" isn't ranking. It's asking around — and your homepage is the one voice in the room most obviously paid to lie.
So how do you actually win?
You don't game this; you earn it. Make the content liftable, make sure the machines can read it, and — the hard part — give the rest of the internet enough true, consistent, corroborated signal that an engine feels safe naming you. There's no schema tag for "trustworthy." It accrues.
What you can do quickly is find out where you actually stand: which questions you're invisible for, which competitors get named instead, which engine is wrong about you, and which of the three gates is closed. That's the difference between fixing the real bottleneck and pouring content into a wall the crawlers never reach.
See which gate is closed — and which engine is wrong about you
The Diagnostic runs your real buyer questions across all five engines, captures what they say, and scores every cited URL on whether AI can find, read, and trust you. One-time, $3,000, confidential under MNDA.
References
- Citation overlap between engines (analysis of ~680M citations): only ~11% of ChatGPT-cited domains also cited by Perplexity. See Averi, AI search visibility analysis (2026).
- On retrieval and citation as distinct steps, see "How LLMs Decide Whom to Cite: 2026 Research Analysis" and Beamtrace, "LLM Ranking Factors" (2026).
- Ahrefs study of 75,000 brands: unlinked brand mentions correlate with AI citations at 0.664 vs 0.218 for backlinks; YouTube mentions 0.737, branded anchors 0.527, branded search volume 0.392. Summarized in "Backlinks vs brand mentions: the 2026 AI visibility playbook".
- Domain authority as a weak predictor of AI citations: AuthorityTech, "Your Domain Authority Score Predicts Less Than 4% of AI Citations" (2026).
- Semantic completeness and citation likelihood (analysis of 15,847 AI Overview results), via "Google AI Overviews Ranking Factors" (2026).