LLM Optimization: Optimize the Inputs, Not the Black Box

Search “LLM optimization” and the results cannot agree on what you asked. Half the page is machine-learning engineering — quantization, KV caches, inference cost. The other half is marketing — “get ChatGPT to talk about your brand” — and most of it hands you the same ritual: type prompts into ChatGPT, see if your name comes up, tweak something, type the prompts again. That ritual has a fatal flaw. You cannot optimize a black box by guessing prompts at it. You can only optimize the inputs it retrieves — structured, extractable, corroborated facts on pages that already carry authority — and then measure whether you actually get cited.

This guide is about the second meaning of the term — LLM optimization for visibility, not for inference cost — done as an engineering discipline instead of tea-leaf reading. It sits under our AI search pillar and leans on its siblings for depth: answer engine optimization for the citation tactics, and closed-loop SEO for the verification mechanics. What this post adds is the argument the ranking pages skip: the model is not the surface you work on. The retrieval layer is.

“LLM optimization” means two different things

The live Google SERP for this exact keyword makes the split visible. Mirantis and the OpenAI developer docs rank for optimizing the model — making inference faster and cheaper. Semrush and Conductor rank for optimizing for the model — making a brand show up in AI answers. An AI Overview sits above all of them, there is no featured snippet, and the People-Also-Ask box carries both intents at once. The engines themselves have not decided what this query means, which is exactly why most content written for it is muddled.

Worth knowing before you invest: the keyword itself draws around 720 searches a month at a keyword difficulty of 1 — real, rising demand with almost no established authority on the visibility meaning. The category is young enough that the standard advice has not hardened into anything good. That is an opportunity, and a warning about what you will read elsewhere.

Why prompt-guessing is not optimization

The default LLMO workflow — ask ChatGPT about your category, note whether you appear, adjust copy, ask again — feels empirical. It is not, for four structural reasons. The outputs are non-deterministic: the same prompt returns different answers on different runs, so a mention appearing or vanishing between Tuesday and Thursday tells you nothing about anything you changed. The outputs are personalized and context-dependent: what the model says in your session, with your history, is not what it says in a prospect’s. The models update silently: a shift in answers may be a new model version, new retrieval behavior, or a new system prompt — none of which you caused. And the sample is hopelessly small: ten hand-typed prompts against a surface that fields millions of phrasings is not a measurement, it is an anecdote.

A mention that appears between Tuesday and Thursday proves nothing about what you shipped. Non-deterministic output cannot be your feedback signal — the inputs and the citations can.

None of this means AI answers are unknowable. It means the model’s output is the wrong layer to work on. Every consumer-facing answer engine that matters — Google’s AI Overviews and AI Mode, Perplexity, ChatGPT when it browses, Claude when it is connected to the web — composes its answers from retrieved documents. Retrieval is not a black box. It is the same observable, rankable, measurable web you already work on. The model is the chef; you do not optimize the chef. You optimize what is in the pantry.

Optimize the inputs, not the model

Working the input layer means making your pages the documents a retrieval step selects and a synthesis step can confidently quote. Three properties decide that, in rough order of how often they are the bottleneck.

Extractable structure

A model composing an answer under a token budget reaches for passages it can lift with minimal rewriting. That rewards a specific shape: a heading phrased as the question, followed immediately by a self-contained 40-to-60-word answer, followed by the evidence. One idea per paragraph. Definitions stated flatly, not teased. Schema — FAQPage, Article with a named author, HowTo where genuine — so the engine can classify each passage without guessing. The full tactical treatment lives in the AEO guide; the principle here is that structure is not cosmetic. It is the difference between a page a model can quote and a page it skips for one it can.

Corroborated facts

Synthesis engines cross-check. A claim that appears once, on your site, in your words, is a candidate. The same claim stated consistently across your site, your documentation, third-party mentions, and structured data is a fact the model can repeat without risk — and models are built to minimize the risk of repeating something wrong. This is why entity consistency matters more in LLMO than it ever did in classic SEO: the engine is not ranking your page, it is deciding whether to assert your claim. Contradictions between your pricing page and your docs, or between your site and your directory listings, quietly disqualify you. Distinctive, attributable claims — a specific number, a named method — survive synthesis best, because the model has a concrete thing to credit.

Host-page authority

Retrieval does not sample the web uniformly. Live-search engines pull heavily from what already ranks; Perplexity’s source set is visibly shaped by conventional search results, and ChatGPT’s browsing mode behaves the same way. A perfectly structured, thoroughly corroborated page on a site with no authority is a well-labeled jar in a pantry the chef never opens. This is the least fashionable part of LLM optimization and the most load-bearing: the boring accumulation of topical depth, internal linking, and earned authority is what gets your pages into the retrieval set at all.

Ranking is still the admission ticket

The most common LLMO mistake is treating AI visibility as an escape hatch from SEO — as if a site that cannot rank can somehow still be cited. The dependency runs the other way. Because retrieval leans on search, your classic organic position is your admission ticket to the answer; structure and corroboration then decide whether you are quoted once you are in the room. A site with no rankings has no retrieval presence, and no prompt engineering changes that.

But holding a ranking no longer guarantees the reward, and one anonymized example shows both halves at once. On a property we audited, a commercial query sat at average position 6.5 — page one, comfortably inside any retrieval set — and collected 952 impressions over 90 days for exactly zero clicks. The page was being surfaced and never chosen: present in the room, quoted by no one, clicked by no one. That is the zero-click pattern our zero-click SEO guide dissects, and it is the precise gap LLM optimization exists to close. Ranking got the page retrieved. Nothing about its structure made it the source the answer used.

The measurement loop is the point

Here is the thesis in one line: input work without a measurement loop is just a different flavor of guessing. Restructuring a page to be extractable is a hypothesis — “this change will get this page cited for these questions” — and a hypothesis without a committed metric and a re-read date is indistinguishable from the prompt ritual it replaced. What separates LLM optimization from LLM superstition is not which tactics you pick. It is whether every change ships with a way to find out if it worked.

Honesty requires saying that citation measurement is immature. There is no Search Console for ChatGPT. But the leading indicators are real, readable, and specific enough to judge individual changes — provided you commit to them before the change ships, the discipline our SEO KPIs guide calls verified causal movement.

A working LLM optimization loop

Put the pieces together and the discipline runs as a loop — the same find → fix → verify shape our closed-loop SEO guide describes, pointed at citations instead of clicks. Find: mine Search Console for question-shaped queries where you are surfaced but not chosen — impressions without clicks, rankings without reward. Fix: make the target page extractable and its claims corroborated, and record exactly what changed and when. Commit: before shipping, write down the metric that will judge the change and the threshold that counts. Verify: roughly three weeks later, re-read that exact metric on that exact query, alongside the scheduled citation spot-check. Confirmed, it becomes evidence about what works on your site; failed, the diagnosis reopens — which is also a result.

Run honestly, the loop produces the one thing the prompt ritual never can: an accumulating, site-specific record of which input changes actually earn citations — answer-first rewrites confirmed on commercial questions, schema additions that moved nothing, corroboration fixes that did. That record is what makes the next cycle smarter than the last. The loop is also relentless clerical work — committed metrics, timestamps, three-week re-reads across dozens of open changes — which is why it suits an agent and defeats vigilance. It is precisely the kind of work the agentic SEO operating model exists to run.

Where LLMO sits next to AEO and GEO

The vocabulary is still settling, and the three terms overlap more than their advocates admit. AEO — answer engine optimization — optimizes for being the cited answer to a question; its unit of success is a citation in a synthesized response. GEO — generative engine optimization — is the wider frame: inclusion and accurate entity representation across generative outputs generally, not only direct answers. LLM optimization, as the market uses the term, is best read as the umbrella over both — with the trap that its name points your attention at the model, the one thing you cannot touch. Whichever label wins, the work underneath is identical: structure the inputs, corroborate the facts, earn the authority, and measure the citations. Pick the term your stakeholders recognize and run the same loop.

Where to start: read your own retrieval evidence

The first move is not a tool purchase and not a prompt session. It is reading the retrieval evidence you already own: the Search Console record of every query where engines surface your pages, and the gap between how often you are shown and how often you are chosen. The queries with heavy impressions and no clicks are your LLM optimization backlog, already prioritized by demand. A free SEO report runs that find pass against your real data — sign in with Google, connect Search Console, and see which pages are being retrieved and ignored. If you would rather have the whole loop run for you — inputs restructured, metrics committed, outcomes verified on a schedule — book a call and we will walk through what it looks like on your site.