LLM Optimization: What You Can and Can't Influence

LLM optimization is the effort to influence how large language models talk about you. That covers whether they mention your brand, whether they cite your pages, and whether the facts they state are right.

The term overlaps heavily with GEO, AEO and AI SEO. The labels differ mostly in who coined them, and our guide to AI search acronyms sorts them out. For a short definition, see what LLMO means.

The useful question isn't what to call it. It's which parts of the process you can actually change. Most LLMO advice skips that question, and the answer decides where your time should go.

An AI answer is not one system. It's a chain, and each link has a different owner.

The stages are sequential. A page that fails access never reaches retrieval. A page that never reaches retrieval can't be cited, however well it's written.

The parametric stage deserves a plain statement. OpenAI's model docs list a February 16, 2026 knowledge cutoff for its GPT-5.6 models and August 31, 2025 for the Instant path. When an answer is generated without live search, nothing you publish today reaches it. Any change to training data lands a model generation later, if at all.

Retrieval is also not guaranteed. As of October 2026, ChatGPT searches at the model's discretion, Claude.ai search is a user toggle, and Google AI Overviews often don't trigger. So any single answer may or may not have used the live web.

Most LLMO tactics on the market are content rewrites: add statistics, add quotations, change the tone, restructure into short blocks. The research behind them shares one design choice. The pages were already placed inside the model's context window, with no live search involved.

That means the studies measured stage four (generation) and then got sold as fixes for stages two and three. A critical survey of 45 GEO studies (arXiv:2607.14035, July 2026) concluded that already-retrieved content can causally alter its citation, but no reviewed technique shows a stable, cross-platform effect on discoverability.

The best-designed test points the wrong way for rewrite sellers. C-SEO Bench (NeurIPS 2025) found that most conversational SEO methods are "not only largely ineffective but also frequently have a negative impact on document ranking." The same benchmark found gains shrink as more sites adopt a tactic, which makes it a zero-sum race.

The widely quoted "up to 40% more visibility" figure comes from the original GEO paper (Aggarwal et al.). It was measured in an engine the authors built, with sources supplied in context. It says nothing about getting found in the first place.

So the honest reading is narrow. Wording can change how a page is used once it's retrieved. Nobody has shown that rewriting gets a page retrieved. Our comparison of on-page rewrites vs access fixes lays out the evidence side by side.

The levers with evidence sit almost entirely at the access stage. They work by removing blocks, not by adding signals.

Crawler permission, per bot

robots.txt rules apply bot by bot. OpenAI's crawler docs say GPTBot governs training, OAI-SearchBot governs ChatGPT search, and "each setting is independent of the others." This combination is valid and common:

User-agent: OAI-SearchBot
Allow: /

User-agent: GPTBot
Disallow: /

The frequent error is a blanket AI block that quietly removes ChatGPT search eligibility along with training. Perplexity (PerplexityBot) and Claude (Claude-SearchBot) use their own search crawlers too.

Network reachability

A firewall or CDN can block a crawler that robots.txt allows. OpenAI's help center says eligibility needs both: allow OAI-SearchBot, and confirm your host or CDN accepts traffic from OpenAI's published IP ranges. How often this breaks is not measured, but it's cheap to test.

Server-side rendering

Vercel's crawler study found ChatGPT and Claude crawlers fetch JavaScript files but don't execute them, so they can't read client-side rendered content. If your pricing or product facts only appear after JavaScript runs, those crawlers may never see them. Googlebot does render JavaScript, so this mostly affects non-Google engines.

Index and snippet eligibility

Google says a page must be indexed and eligible for a snippet to appear as a supporting link in AI Overviews or AI Mode, with "no additional technical requirements." Check that noindex, nosnippet, max-snippet:0 and data-nosnippet are only where you intended.

Specific facts on a fetchable page

Stating a price, a number or a one-line definition on a page crawlers can fetch gives an engine a sentence it can quote. It's the only positive experimental signal for on-page content, and it's modest. Our page on structured data vs plain-text facts covers why the plain sentence matters more than markup. For a platform-specific view, see ChatGPT visibility factors graded by evidence.

Be clear-eyed about these, because many LLMO services charge for them.

• Training data. You can allow or block training crawlers. Nobody, including the vendors, has measured whether allowing them improves how a model recalls your brand.
• The retrieval query. OpenAI says ChatGPT search rewrites your prompt into one or more targeted queries. Google may fan a query out across subtopics. The prompt you test is not the query the engine runs.
• Position in the context. C-SEO Bench found position in the model's context list matters a lot. No publisher can set it.
• Brand prominence. Large, well-known sources dominate citations across studies. You can't buy that this quarter.
• Whether an answer used the web at all. You often can't tell, unless citations appear.

There is one indirect route. When engines cite other sites instead of yours, being named on those sites is associated with showing up in the answer. It's an association, not a promised effect. See your own site vs third-party mentions.

Work through the stages in order. Each step is checkable, and each removes a reason to be invisible.

1. Check crawler permission. Read your robots.txt for OAI-SearchBot, PerplexityBot, Claude-SearchBot, Googlebot and bingbot. Decide on training bots (GPTBot, ClaudeBot, Google-Extended) separately.
2. Check reachability. Request a page as each AI crawler and look for 403s or challenge pages. The AI Crawler Checker does this and shows which crawlers your firewall turns away.
3. Compare raw HTML with the rendered page. Turn off JavaScript and see whether your pricing, product names and key claims are still there. If not, server-render them.
4. Audit suppression controls. Search templates for noindex, nosnippet, max-snippet and data-nosnippet that were added by accident.
5. Confirm index status per engine. Google Search Console and Bing Webmaster Tools tell you different things. A healthy one says nothing about the other.
6. Fix contradictory facts on your own site. Make your price, founding year and product names match across pricing, about and footer pages.
7. Measure each platform separately. The free checker on this site shows whether AI answers name you for up to three buyer questions. Treat it as a snapshot, and expect answers to vary run to run.

These recommendations are common, and primary sources contradict them.

• "Add schema to get cited." Google says structured data isn't required for generative AI search. Ahrefs' matched test found AI Overviews citations fell 4.6% on pages that added schema. Existing markup does no harm, but it isn't a lever.
• "Publish an llms.txt." Google says Search ignores it. No other engine documents reading it.
• "Improve your E-E-A-T score." Google says E-E-A-T isn't a specific ranking factor. There is no score.
• "Allow GPTBot so ChatGPT can find you." GPTBot is for training. ChatGPT search uses OAI-SearchBot.
• "Chunk your content for LLMs" or "write AI-friendly." Google rejects both and says there's no ideal page length. Our digest of Google's generative AI optimization guide quotes the relevant lines.
• "Track your AI rank." Google's AI report exposes impressions only. No position data exists to check a rank against.
• Hidden text or prompt injection aimed at AI crawlers. It exists, and research shows it can sway some search agents. It's manipulation, it risks penalties and reputational damage, and it's not optimization.

The pattern is the same each time. A tactic that sounds additive has no evidence. The work that holds up is making sure nothing stops engines from reading what you already publish.

Is LLM optimization different from SEO? Partly. Google says there are no special requirements for AI Overviews or AI Mode beyond being indexed and snippet-eligible, so Google's AI surfaces inherit normal SEO hygiene. Other engines add their own crawlers and rarely execute JavaScript, which creates new access checks. The rewriting and markup tactics sold as LLM-specific have no credible evidence behind them. Treat LLMO as SEO hygiene plus per-engine access checks, not a separate discipline.

Can I get my brand into an LLM's training data? You can allow training crawlers such as GPTBot, ClaudeBot and Google-Extended in robots.txt. That signals permission, not inclusion. No vendor confirms what entered a training set, and nobody has measured whether allowing these bots improves brand recall. Any effect would arrive a full model generation later, after the next cutoff. Decide on training access as a policy choice, not as a visibility tactic.

Should I block GPTBot? That's a policy decision about whether OpenAI may use your content for training. It doesn't affect ChatGPT search. OpenAI's crawler documentation says each setting is independent, so you can block GPTBot and still allow OAI-SearchBot to appear in ChatGPT search answers. What you should avoid is a blanket AI block that disallows OAI-SearchBot by accident.

How long do access fixes take to show up in AI answers? Nobody publishes a timeline, and no fix guarantees inclusion. An access fix makes your page eligible; each engine still has to recrawl, index and choose to retrieve it for a given query. Recheck crawler responses right after the fix, then measure answers per platform over the following weeks. Expect variation between runs, because retrieval happens per query and at each engine's discretion.

How do I know if an AI answer used my website? Only when the answer shows citations or links. Without them, the answer may come from the model's training data with no live search at all. Even citations undercount use: OpenAI's API docs note that more sources are consulted than cited. Track per platform, note whether citations appeared, and treat mentions and citations as separate measures.