Why the four engines behave differently
All four LLMs are trained on overlapping web data and grounded through similar retrieval pipelines, but the citation choice is a downstream product decision, not a training-data property. Each engine's designers made different calls about how many sources to cite, how to attribute them, whether to expose citation numbers in-line or in a source list, and how strongly to weight authorship and freshness. Those product choices produce visibly different citation patterns in tracked healthcare queries, and a brand aiming to earn citation across all four has to work with those differences, not against them.
The good news is that the underlying content pattern that satisfies all four is highly consistent: reference-style prose, named-expert authorship, tight extractable answers, sourced numbers, clean canonical structure. A page engineered for any one of these engines is meaningfully better positioned for the other three than a page engineered for classical SEO alone. The per-engine work is a set of targeted overlays on top of that base pattern, not four separate content programmes.
ChatGPT · the named-expert engine
ChatGPT's search and browsing modes typically ground on one to three sources per answer. That is a tight citation set; the bar for making it is higher than for a wider-net engine like Google AI Overview. In tracked healthcare queries, the single most reliable lever for earning ChatGPT citation is named-expert authorship — the same fix that lifts AIO citation lifts ChatGPT citation more, because ChatGPT's smaller citation set is more selective on trust signals.
The three-move ChatGPT playbook, in order of leverage:
- Deploy Physician schema and credentialled bylines on every clinical page. Not just the money pages — every article, every treatment explainer, every FAQ.
- Make the primary-topic H1 read as a single unambiguous entity. "LASIK surgery cost in Bangalore — procedure, recovery, and pricing" beats "Everything you need to know about LASIK" because the H1 tells the retrieval step exactly which entity the page is anchored on.
- Keep canonical URLs clean of tracking parameters. ChatGPT's retrieval step lands on the canonical; a canonical URL polluted with UTM parameters costs citation trust.
Perplexity · the numbered-fact engine
Perplexity cites three to five sources per answer with numbered footnotes tied to specific sentences. That structure means Perplexity's grounding step reads for extractable factual claims that can be attributed to a specific paragraph — a stat, a percentage, a comparative claim, a sourced number. Pages that surface those cleanly earn more Perplexity citation than pages that make the same claim in dense narrative prose.
The three-move Perplexity playbook:
- Insert two to four sourced numeric facts into every decision-stage page. Recovery time in days, per-cycle success rate, prevalence data, comparative cost range. Each fact tied to a source citation on the page itself.
- Add a mid-body outbound citation to a peer-reviewed source. Perplexity weighs source-page citation quality when it selects grounding sources; a page that itself cites well is a stronger candidate.
- Keep FAQ answers tight and extractable. Perplexity extracts sentence-level; long answers get truncated. 200 characters per answer is a working ceiling.
Claude · the reference-prose engine
Claude, when accessed through products that give it search grounding, cites a small set of sources per answer and biases toward reference-style, even-handed prose. Promotional language, superlative claims and pitch tone all downweight a page in Claude's retrieval. This is the engine where "we are the best" pages lose to "here is how X actually works" pages, on the same underlying topic.
The three-move Claude playbook:
- Rewrite promotional pages into reference-style articles for the top query cluster. Keep the same substantive information; strip the pitch. Claude is not looking for a sales page.
- Add credentialled author bylines. Claude reads authorship as an authority signal similar to how ChatGPT does — missing bylines cost citation.
- Structure content into clear headings and short paragraphs. Claude's grounding prefers content it can retrieve in coherent extractable blocks; wall-of-text prose loses even when the substance is strong.
Gemini AI mode · the Google-retrieval-shape engine
Gemini AI mode leans heavily on Google's retrieval stack. Its citation behaviour overlaps significantly with Google AI Overview and, to a lesser extent, with the classical SERP. This means most Gemini optimisation is downstream of AIO optimisation — the same six-schema-type architecture, the same freshness discipline, the same entity work. Where Gemini differs is in multi-modal grounding: it will readily cite video and image content alongside text, more visibly than the classical AIO panel does.
The two-move Gemini playbook:
- Run AIO optimisation first. Anything that lifts Google AIO citation typically lifts Gemini AI mode citation as a downstream effect.
- Add a video asset (via YODA) targeting the same AIO-triggering query cluster. Gemini's multi-modal grounding gives video assets more citation surface than AIO alone; a YouTube asset aligned to the cluster earns Gemini citation slots the text-only pages cannot.
The base pattern all four engines reward
Underneath the per-engine overlays, the base pattern is the same. A healthcare page engineered for LLM citation looks like this:
- Direct-answer paragraph in the first 100 words, written to stand alone as a self-contained answer.
- FAQPage schema with 6–10 tight answers, each under 300 characters.
- Person / Physician schema with credentials and specialty on every clinical author.
- MedicalBusiness / MedicalProcedure / MedicalCondition entities cross-referenced across the site.
- Two to four sourced numeric facts in the body, each with a citation on the page.
- Speakable specification pointing at the FAQ and direct-answer selectors.
- dateModified within 90 days for evolving topics.
- Reference-style prose — no superlatives, no pitch language, no hollow "leading" or "best" claims.
Get the base pattern right and citation shows up across all four engines within one to two months. Add the per-engine overlays and citation share compounds meaningfully faster.
