Per-LLM Optimisation

Every model reads differently.

ChatGPT, Claude, Perplexity, Google AI Overviews, and Microsoft Copilot do not return the same sources to the same question. They are trained on different data, retrieve in different ways, and cite by different rules. We tune for each — in concert, not in isolation — so your business shows up across the conversational layer rather than only in one corner of it.

Models we tune for · ChatGPT · Claude · Perplexity · Gemini · Copilot · Grok · Per-Model Tracking · Bilingual Visibility
01 — What’s Included

Six citation surfaces.
One coordinated strategy.

Per-LLM optimisation is the work that begins after the foundational pillars — visibility audit, citation build, schema — are in place. It is the model-by-model tuning that decides whether your business is the one cited across all of them or only in one or two.

Each row below is a real platform with its own retrieval mechanics, its own training cadence, and its own quirks. We treat them individually, then coordinate the tactics so the work compounds rather than fragments.

N° 01

ChatGPT (OpenAI)

Hybrid retrieval

ChatGPT cites from two distinct sources: its training corpus (knowledge baked in at training time) and live web browsing through SearchGPT and Bing. Older brands lean on training presence; younger or fast-moving ones lean on browsable authority. We optimise for both — establishing your entity in the kinds of sources that survive into training data while keeping the live-web profile crawlable, well-structured, and fresh enough that ChatGPT’s browsing layer pulls you in. We track what GPTBot crawls, where SearchGPT is sending traffic, and how ChatGPT’s citation surface treats your domain across personas and prompt variations.

N° 02

Claude (Anthropic)

Knowledge-first

Claude has historically leaned on a deep knowledge cutoff with measured live-web retrieval, and its citation behaviour rewards genuine authority over freshness. The pages Claude surfaces tend to be the considered, well-cited, well-structured ones — Wikipedia, well-regarded industry publications, official documentation, professional bodies. We work to establish your entity in those exact source classes, and we tune ClaudeBot access so the public side of your site is fully readable while sensitive pages are appropriately gated. The reward for this work is a quieter but more reputational kind of citation — one that buyers tend to trust on first sight.

N° 03

Perplexity

Real-time citation

Perplexity is a search-citation hybrid: every answer comes with sources visible at the top, and the citation layer is the product. That makes it the most measurable AI surface in the stack — and the most rewarding for sites with strong, current, source-grade content. We optimise for PerplexityBot’s crawl, write the kind of clean, sourced, recent answer pages that Perplexity prefers, and structure your content so it can be quoted directly into Perplexity’s answer cards. We then track the share of citations you capture for the queries that matter to your business, query by query, week by week.

N° 04

Google AI Overviews (Gemini)

Search-grounded

AI Overviews — the Gemini-powered answer block at the top of Google search — pull from the same index that already ranks your pages. That makes traditional SEO a direct input to AI Overview visibility, but the citation rules differ: Overviews favour pages with answer-formatted content, structured data, declared expertise, and clear topical authority. We implement the schema, content patterns, and on-page structures that AI Overviews specifically reward, and we monitor your overview presence on the queries that matter — which terms trigger an Overview, who is cited, and what content shape gets pulled into the answer.

N° 05

Microsoft Copilot (Bing)

Bing-powered

Copilot — across Windows, Microsoft 365, Edge, and the consumer chat — is grounded in Bing’s index. Bing’s signals overlap with Google’s but diverge on weighting: Bing tends to value direct keyword relevance, IndexNow submissions, and Bing Webmaster Tools verification more than Google does. The audience also matters; Copilot is increasingly the default AI surface for enterprise users in Microsoft 365 environments. We treat Bing as a first-class indexation target — submit cleanly, verify in Webmaster Tools, monitor coverage — and we tune content so that Copilot’s enterprise queries surface your business when relevant.

N° 06

Emerging models & bilingual surfaces

Watching

Grok, Mistral, DeepSeek, Qwen, and the new wave of agentic browsers are not hypothetical — they are routing real buyer questions today. We add new models to your tracking set as they cross a meaningful traffic threshold, and we extend the monitoring to bilingual surfaces where it matters. For our EN · ZH clients, that means parallel coverage of how a brand is cited in English-language and Chinese-language prompts, including by the Chinese-trained models that an Australia-or-Canada-only setup typically misses entirely.

N° 07

The other three pillars, in concert

Linked sub-disciplines

Per-LLM tuning sits on top of the foundational pillars — visibility audit, citation build, and schema. We coordinate the four so that what you publish for one model accrues to the others, rather than being thrown away. Click through for the full AI Search Optimisation service.

02 — Our Approach

Track per model.
Coordinate the tactics.

The temptation is to optimise for the one model that happens to be in the news that month. The discipline is to monitor all of them, weigh them by your buyers’ actual usage, and choose the work that lifts more than one at a time. Our approach is built around that discipline.

i

One question, six answers

For every priority query, we record what each of the major models returns — the cited domains, the answer phrasing, the rank order in their respective citation surfaces. Five identical questions can produce five almost entirely different source lists. The diff between those lists is where the per-model work lives, and it is what we report on every month.

ii

Crawler access tuned per bot

GPTBot, ClaudeBot, PerplexityBot, Google-Extended, OAI-SearchBot, Bingbot, Applebot — each has its own user-agent, its own behaviour, and its own set of opt-out signals. We configure robots.txt and meta directives so the bots you want citing you have full access, the ones you’d rather opt out of are politely declined, and the public side of your site is consistently readable across the lot.

iii

Content shaped for each model’s preferences

Perplexity rewards source-grade clarity; Claude rewards considered authority; AI Overviews reward concise, well-structured answers; ChatGPT rewards entity presence in canonical sources. We do not write four versions of the same page — we write one that hits the overlap, and we add the specific structural patterns each model rewards. The result is content that earns citations from more than one surface for the same investment.

iv

Bilingual coverage where the buyer is

For clients selling into both English- and Chinese-language audiences, we extend the monitoring to both linguistic surfaces and the model sets that serve each. A brand that is cited well by ChatGPT in English but invisible to the Chinese-trained models is leaving half the conversation on the table. We close that gap, deliberately, and report on it as a single coordinated picture.

03 — Who It’s For

Businesses being
asked about by name.

Per-LLM optimisation is most valuable for businesses whose buyers are already past the awareness stage and are actively researching providers, products, or services in the conversational layer. If your competitors are showing up in answers and you are not, this is the work that rebalances that picture.

Businesses where the model’s answer carries weight with the buyer.

  • i High-consideration service firmsLaw, accounting, advisory, healthcare, migration. Buyers ask conversational questions like “best immigration lawyer in Vancouver” or “which firm handles small-business tax in Toronto” — the cited business wins the click before any ad does.
  • ii Brands already cited by one model and absent from anotherThe most common audit finding is uneven coverage — strong in Perplexity, invisible in Claude, partial in AI Overviews. Per-model tuning is exactly the work that closes those visible gaps without disrupting the wins you already have.
  • iii SaaS and technology companies in evaluation cyclesTechnical buyers ask AI models for vendor comparisons, integration questions, and feature-fit checks. Being cited as a recommended option in the model’s answer is a higher-intent signal than any banner placement, because the buyer is already in the decision.
  • iv E-commerce and product brands during research moments“What’s the best standing desk under five hundred” or “which Canadian brand makes organic dog food” are questions models answer with a finite list. Being on the list — across multiple models, not just one — is the difference between consideration and silence.
  • v Bilingual businesses serving EN and ZH audiencesIf half your buyers ask in English and half in Mandarin, you are already operating across two model ecosystems. Per-LLM optimisation across both languages is one of the few places where bilingual operators have a structural advantage worth pressing.

If your business is being asked about by name in any of these conversational surfaces, the question is no longer whether to invest in per-model optimisation but how soon to begin. Citation patterns compound — early authority is sticky and difficult for late entrants to dislodge once it has been established.

04 — A complimentary report

Curious how Google sees your site?

Send us your URL. We’ll send back a Premium SEO Report, prepared by hand, within 48 hours — domain authority, keyword rankings, backlinks, competitor gap, and the quick wins worth chasing first. The same report doubles as the starting point for AI search visibility, because the model layer reads many of the same signals search does.

No sales call required.

Each model has its own citation sensibility — its own taste in what counts as a good source. The work is to be a good source on each one’s terms.
— The Aureole Practice —
05 — Frequently Asked

Questions about
per-model tuning.

Per-LLM optimisation is the youngest sub-discipline in our practice and the questions that come back are pointed and specific. If a question is missing here, the contact link at the foot of the page goes to the person who would answer it.

i Why optimise per model rather than once for all of them?
Because the same query produces different answers depending on which model you ask. ChatGPT might cite three sources, Perplexity five, AI Overviews two, Claude one — and the overlap between them is often surprisingly small. Each model has different training data, different retrieval mechanisms, and different rules for what gets cited. Optimising once and hoping for the best leaves you visible on whichever model happens to match your existing content profile, and silent on the rest. Per-model tuning is the work that gets you cited consistently across the conversational layer rather than only in one corner of it.
ii Which models do you actively track and optimise for?
Our standard tracking set covers ChatGPT (including web browsing and SearchGPT), Claude, Perplexity, Google AI Overviews (Gemini-powered), and Microsoft Copilot (Bing-powered). For bilingual EN · ZH clients we extend the set to the Chinese-trained models that route a meaningful share of buyer questions in that market. We add new models — Grok, Mistral, agentic browsers, and so on — when they cross a usage threshold that makes them worth the additional measurement overhead. Anything below that threshold gets watched but not actively optimised; we’d rather do four models well than ten poorly.
iii How do you actually measure citations across so many models?
Manually and methodically. We maintain a query set per client — typically thirty to eighty prompts that mirror the questions your buyers actually ask — and run them against every model in the tracking set on a defined cadence, usually monthly. We log who is cited, where you appear or do not, what answer phrasing the model uses, and which specific URL was pulled. Some platforms (Perplexity especially) make the citation layer transparent; others require careful reading of the response and follow-up probing. The work is partially tooled but mostly disciplined human review, because no automated dashboard yet captures all of this reliably. Anyone claiming otherwise is overselling their tooling.
iv Will work that lifts one model help with the others?
Often, yes — and that overlap is exactly what we design the work around. Strong structured data, clean entity declarations, well-written answer-formatted content, and authoritative citations all feed into multiple models simultaneously. Where the optimisations diverge — Perplexity’s preference for very current sources, Bing’s IndexNow signal, Claude’s reward for measured authority — we layer the model-specific tactics on top of the shared foundation rather than building four entirely separate strategies. The result is that most of the work compounds across the stack, and only a small remainder is genuinely model-exclusive.
v How long before per-model tuning shows up in citations?
It depends on the model and your starting position. Perplexity, which crawls actively and cites in near-real-time, can show measurable change within four to six weeks of substantive content and structure work. Google AI Overviews tend to track Google’s index latency, so meaningful change often follows the same recrawl-and-reprocess cycle as traditional SEO — typically six to twelve weeks. ChatGPT and Claude move slower because their training cycles are longer, and full citation patterns can take three to six months to settle. Across the stack, sustained work over a quarter or two consistently produces visible per-model wins; one-off engagements rarely do.
vi Do you also handle Chinese-language model coverage for bilingual brands?
Yes. For bilingual EN · ZH clients we extend the tracking set to the Chinese-trained models that handle Mandarin queries — including the ones a Western-only setup tends to miss entirely — and we run query sets in both languages, not just translated mirrors of the English ones. The work is rarely a straight translation; the right answer in Mandarin is often a different shape from the right answer in English, both in phrasing and in the source classes the model trusts. We treat the two language surfaces as parallel optimisation tracks that share a foundation, and we report them as a single coordinated picture rather than two disconnected channels.
The Invitation

Find out where each
model sees you today.

We’ll run an AI visibility audit across ChatGPT, Claude, Perplexity, AI Overviews, and Copilot — and show you, model by model, where you stand and where the gaps are. No obligation, no sales department, no funnel.

Mon–Fri · 9–6 PT support@aureoleintelligence.com Reply within 1 business day