LLM SEO: How to Get Cited by ChatGPT and Perplexity

LLM SEO, answer engine optimization and generative engine optimization are three names for substantially the same practice. The market hasn't settled on one, and anyone insisting they're rigorously different is usually selling something.

What they all describe is making your content the source a model quotes when someone asks a question. This covers the two routes into an AI answer, the crawler settings that determine which route is even open to you, and what to change on the page.

At a Glance

Two routes into an AI answer: training data, which is slow and largely outside your control, and live retrieval, which is fast and actionable. Optimise for the second.
Those routes map onto different crawlers. OpenAI runs three, controlled independently in robots.txt, and blocking the wrong one closes the route you wanted.
Crawlers generally don't execute JavaScript. If your content only exists after hydration, it may as well not exist.
Lead with the answer. Models lift short self-contained passages, not narrative buildup.
Citation and clicks move separately. Expect mentions to rise while sessions fall, and track both.
Refresh quarterly and update dateModified. Retrieval systems favour content that looks current.

The two routes in

Everything divides into these, and confusing them is why a lot of LLM SEO advice is useless.

Training data. The model learned about you during pre-training. Durable, but the feedback loop is measured in model releases, you can't influence it directly, and by the time it works the information may be stale.

Live retrieval. The system fetches your page during the conversation and cites it. Fast, measurable, and responsive to what you publish this week.

Nearly all practical work targets retrieval. If someone promises to get you into training data, ask how they'd verify it.

The crawler layer, which most guides get wrong

Here's the part that catches people out. It isn't one setting.

OpenAI documents three separate crawlers, each controlled independently in robots.txt:

GPTBot collects content for training foundation models.

OAI-SearchBot handles ChatGPT search results.

ChatGPT-User fetches pages when a user's conversation triggers a live lookup.

The independence matters more than the list. OpenAI's own documentation gives the example directly: you can allow OAI-SearchBot so you appear in search results while disallowing GPTBot to keep your content out of model training. Those are separate decisions with opposite effects on visibility.

Which means a blanket block on "AI crawlers" — a common reaction — closes the retrieval route you actually wanted while achieving a training opt-out you may not have needed. And a robots.txt change takes roughly 24 hours to propagate to search results, per OpenAI's documentation.

Anthropic runs ClaudeBot and Perplexity runs PerplexityBot, both with published documentation and IP ranges. Check all of them rather than assuming.

What to do this week: open your robots.txt and read it properly. Not "are AI bots blocked" but which ones, and whether the block matches what you actually intended.

Technical foundations

Three things, in order of how often they're the problem.

Rendering. Most of these crawlers don't execute JavaScript. Content that appears only after client-side hydration is invisible to them, however well it renders for a human. Server-side rendering or static HTML for anything you want cited — documentation, guides, product pages.

Structured data. Helps machines parse facts unambiguously, though be realistic about it: no schema type is required for AI Overviews and Google has published no AI-specific markup. Organization schema for brand identity, Article schema with author and dateModified, Person schema linked to a real bio. We cover the current picture in schema markup for AI search, including which types were retired in 2026.

Freshness signals. Update dateModified when content genuinely changes. Retrieval systems weight recency, and a page that hasn't been touched in three years reads as stale whether or not it's still accurate.

Writing content that gets lifted

Models extract passages. The formatting implications are specific.

Answer first. State the claim in the opening 40 to 60 words, then support it. A page that builds to its point over four paragraphs gets skipped for one that states it immediately.

Short paragraphs. Two to four sentences. Long blocks extract badly or not at all.

Facts in tables and lists. Numbers, comparisons and specifications are far more liftable as structured elements than as prose.

Consistent heading hierarchy. H1 through H3, phrased as the questions people actually ask rather than as internal filing labels.

Self-contained sections. Any section should make sense lifted out on its own. No "as we saw above."

The underused source of this material is conversations you already have. Support tickets, sales call recordings and objection handling contain the exact phrasing your buyers use, and turning those into public Q&A is the cheapest liftable content available.

Amplification, which is the slow part

Retrieval systems weight corroboration. Information several independent sources agree on gets used more readily than the same claim on your site alone.

That makes third-party presence a genuine ranking input rather than a nice-to-have. Industry roundups, review platforms, community discussions, creator coverage. It behaves more like PR than SEO, with the timelines to match — expect a quarter before it registers, not a fortnight.

Measuring it

Citation share against a fixed set of buyer prompts. The set has to stay fixed or month-to-month comparison is meaningless.

Sampling matters. Answers vary between runs for the same prompt, so a single check is a reading rather than a measurement. Repeated sampling is the only way to get a real number.

Track clicks and citations separately. They move independently. Mentions can rise while GA4 sessions fall, because a cited answer often satisfies the question without a visit. Judging LLM SEO on sessions alone will tell you it isn't working when it is.

For the tooling, AI visibility tools and LLM visibility tools cover the platforms and what each actually reports.

What this changes commercially

Three practical consequences worth planning for.

Pages models cite behave like landing pages, so treat them that way. Test whether a short summary at the top increases the click rate on the ones that matter commercially.

Retarget AI-sourced visitors. If organic sessions dip while brand awareness rises, some budget belongs on capturing that intent rather than defending the traffic number.

Prioritise commercially valuable pages first. The pages closest to a buying decision are where citation is worth most, and they're usually not the ones with the most traffic.

Where Strivelabs fits

Strivelabs samples ChatGPT, Gemini, Claude, Perplexity and Google AI Overviews against prompt sets you define, three samples per prompt, working standalone. It also connects to Search Console, GA4, Google Ads, HubSpot and LinkedIn Ads, so a citation finding becomes a routed task with a named owner rather than a dashboard entry.

The honest limitation: three samples per prompt shows whether you appear consistently, occasionally or not at all, which is enough to act on but lighter than dedicated trackers running higher counts. It covers five engines rather than nine.

Track your citation share across ChatGPT, Gemini, Claude and Perplexity

Strivelabs samples five AI engines against prompt sets you define — and routes findings to a named owner.

Book a Demo →

Frequently Asked Questions

What is LLM SEO?

Optimising content so language models cite it when answering questions. Substantially the same practice as answer engine optimization and generative engine optimization, measured on citation share rather than clicks.

Does ranking well on Google guarantee AI citations?

No. Rankings help, since retrieval often draws on search indexes, but models select on extractability and factual clarity as well as authority. Pages outside the top ten get cited, and pages at position one get skipped.

Will blocking AI crawlers hurt my Google rankings?

No, these are separate systems. But it will remove you from the AI answers those crawlers feed. Worth knowing that OpenAI runs three crawlers with independent controls, so you can allow search inclusion while blocking training rather than choosing all or nothing.

How often should I refresh content for AI citation?

Roughly quarterly for pages you want cited, updating dateModified when the content genuinely changes. Retrieval systems favour recency, but a date change without a content change is just noise.

Does schema markup get me into AI Overviews?

No. Google has published no AI-specific schema requirement. Structured data clarifies facts for machines, which is useful and narrower than the claims usually made for it.

Why are my clicks falling while mentions rise?

Because a cited answer often resolves the question without a visit. That's the expected pattern rather than a failure, which is why citation and session metrics need tracking separately.

Where do I start?

Read your robots.txt properly and confirm which crawlers are actually blocked. Then pick one commercially important page, rewrite the opening as a direct answer, update the schema, and run a before-and-after prompt test in four weeks.