LLM visibility — what it is, why it matters, and how to measure it

Every few months, a client asks us why their content has disappeared from search. Before 2024, the answer was usually an algorithm update, a penalty, or a competitor with more backlinks. Now, the answer is different: their content exists on Google, but it doesn't exist in ChatGPT or Perplexity — and that's where a growing portion of their category traffic went.

LLM visibility is the measurable frequency with which your content gets cited by AI-powered search tools. We built our own tracker in late 2025 to measure it systematically. This post defines the term, explains why it is only now measurable at scale, and walks through the four-layer methodology we use in practice.

What LLM visibility actually means

LLM visibility is the measurable presence of your content in AI-generated responses — specifically, how often ChatGPT, Perplexity, Google AI Overviews, and similar systems surface your content when a user asks a question in your category.

It is not the same as traditional search visibility. A page can rank first in Google organic results and never appear in a single AI Overview response. A page can rank 11th and get cited by ChatGPT repeatedly. The signals that drive traditional ranking and LLM citation overlap (content quality, domain authority, schema markup) but the weighting is different, the citation mechanism is different, and critically, the measurement approach is completely different.

What LLM visibility is not: a replacement metric for SEO. It is an additional signal that matters in proportion to how much of your audience's search behaviour has shifted to AI tools. For most SMB marketing teams, that shift is already well underway. A Reuters Institute study tracking AI usage across six countries found that weekly use of generative AI tools nearly doubled in a year — rising from 18% to 34% — and that 61% of Americans reported having seen an AI-generated answer in response to a search query in the past week (Reuters Institute, via Nieman Lab, October 2025). AI-mediated discovery is no longer a fringe channel — for a majority of your audience, it is already a default surface they touch each week.

The implication is concrete: if your target audience is asking questions in ChatGPT or Perplexity instead of (or before) Google, and your content is not appearing in those answers, you have zero visibility for that portion of the query funnel — regardless of your traditional SERP position.

Why LLM visibility is now measurable (it wasn't 12 months ago)

Eighteen months ago, measuring LLM visibility was not practical for most teams. You could manually ask ChatGPT whether it would cite your content on a given topic — some teams were doing exactly this, spending hours on manual spot-checks with no systematic approach. But there was no infrastructure for repeatable measurement, no API access to AI Overview citation data, and no way to tie LLM-referred traffic back to specific content pieces at scale.

Three things changed between late 2024 and Q1 2026.

DataForSEO added AIO component data to their SERP API in August 2024 — the first time programmatic querying of AI Overview citations was possible at scale. It is not perfect — AI Overviews are personalised and non-deterministic — but across tracked keyword sets it gives a reliable citation frequency signal. Before that release, there was no programmatic way to detect AI Overview citations at all.

UTM attribution for AI-referred traffic also matured over the same period. Perplexity and similar tools now pass referrer data in a consistent format that Google Analytics 4 and most attribution platforms can classify correctly. Traffic from perplexity.ai that arrives via a cited article is trackable end-to-end. The measurement lag has shrunk from months to days.

The third shift was volume. OpenAI's own usage research, published with NBER in 2025, found that "Seeking Information" — described as a close substitute for web search — accounts for roughly 21% of the 18 billion weekly messages sent through ChatGPT (Chatterji et al., NBER Working Paper, 2025). That is billions of search-equivalent queries per week happening entirely outside Google's index. Beyond raw volume, among Americans who report seeing AI-generated search answers, only about a third say they "always or often" click through to source links, while 28% say they "rarely or never" do (Reuters Institute, 2025). That click-through compression is precisely why citation frequency has become the variable worth tracking: the citation often is the impression. By mid-2025, the volume in most marketing-adjacent verticals was dense enough to start detecting patterns.

Together, these three shifts made systematic LLM visibility tracking possible. We started building our internal tracker in October 2025.

How RP measures LLM visibility in practice

The tracker runs four measurement activities on a fixed cadence. This is the methodology we use on our own content and the framework we recommend to anyone building LLM visibility tracking from scratch.

1. Weekly DataForSEO AIO cron (automated)

For each tracked domain, maintain a keyword list — typically 50–150 target keywords depending on domain size and topic breadth. Every Monday, a Claude Code script queries DataForSEO's SERP API for each keyword, extracts the AIO component data, and logs whether the domain appears in the AI Overview for that keyword. This produces a weekly citation rate: "this week, this domain appeared in AI Overviews for 23 of its 80 tracked keywords."

2. Monthly manual ChatGPT / Perplexity / Claude runs

The DataForSEO data covers Google AI Overviews but not direct ChatGPT or Perplexity citations. Once a month, run a set of 10–15 representative prompts through ChatGPT-4o, Perplexity Pro, and Claude, note which responses cite the domain's content, and log the results. This is slower than the automated layer but necessary — the major LLM providers do not expose citation APIs.

3. UTM traffic monitoring in GA4

Monitor traffic from known AI referrer domains — perplexity.ai, chatgpt.com, claude.ai, and AI Overview landing traffic from Google — as a dedicated segment in GA4. When a piece of content sees a sudden uptick in this segment without a corresponding organic ranking change, it usually signals a new AI citation rather than a ranking movement. This tracking operates at aggregate level — you are measuring which content receives AI-referred sessions, not identifying individual users — so standard GDPR and CCPA compliance for GA4 applies without additional complexity.

4. Before/after structural experiments

When you make a structural change to a piece of content — adding FAQPage schema, restructuring H2s, adding a definition block — log the change date and track whether the AIO citation rate changes over the following 4–6 weeks. This is the only way to build a causal picture of what is actually moving citation rates on your specific domain.

The tracker is not a commercial product — we built it internally using Claude Code and the DataForSEO API, and you can see the full architecture and code in PortableText [components.type] is missing "span". The methodology is fully reproducible with DataForSEO API access and 2–3 hours per month for the manual verification layer.

What the research says moves the needle

The most rigorous published work on this question comes from the GEO paper — "GEO: Generative Engine Optimization" by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, and Deshpande (arXiv:2311.09735, accepted KDD 2024) — which systematically tested content optimisation tactics against Perplexity and measured citation impact. It is the closest thing the field has to a controlled experiment, and the findings cut against most of the advice currently circulating.

The highest-impact tactics were adding citations and sourced statistics (Statistics Addition showed improvements of up to 37% on one metric; Cite Sources showed consistent gains), writing in a more authoritative register, and including direct quotations from named experts. Keyword density and generic structural changes had near-zero effect. Mike King of iPullRank, who synthesised the GEO findings in his SEO Week 2025 keynote, put it plainly: "citing sources, being more authoritative in how you speak, and also having statistics are the things that get you in there most" (PortableText [components.type] is missing "span").

Schema markup type is where most guides go wrong. Article schema tells a crawler what type of content a page is. FAQPage and DefinedTerm schema create explicit, machine-readable answer structures: the kind of attributable answer that LLMs can extract and cite without inferring structure from prose. The GEO findings on authoritative, structured content map directly to this. Schema is the mechanism that makes structured content legible at machine scale.

H2 phrasing is likely a citation signal. Pages where each H2 answers a distinct question — "How do you calculate ROAS attribution across channels?" rather than "ROAS attribution" — should get cited more than pages with vague section titles. LLMs extract content by matching section text against query intent. When the H2 is already a question, the match is direct and the extraction is clean. We are watching for this in our own tracker data.

Answer density in the first 60 words of each section matters separately. LLM citation appears to operate at section level, not full-page level. If the opening paragraph contains a complete, standalone answer to that section's question, the content is easier to extract and attribute. This fits the GEO finding on fluency: systems reward content that answers directly, not content that builds toward an answer.

The domain authority floor is lower than for traditional SEO. Newer domains do appear in AI Overviews for competitive queries when their content has strong schema and dense answer-friendly structure. Traditional ranking requires sustained authority accumulation; AI citation rewards relevance and structure more directly. This matters if you are building on a newer domain with strong content.

Content recency is a factor, particularly with dateModified schema. Pages with a current dateModified value appear to have an advantage on recency-sensitive queries — for example "best [tool] in 2026" or "latest [tactic] approaches". Keep high-value pages actively maintained and update the schema timestamp when you do.

Internal linking creates topical authority clusters. When multiple pages on a domain get cited for related queries, the per-page citation rate tends to increase, as if the AI's confidence in the domain goes up when it finds multiple relevant answers in the same place. Deliberate topical clustering appears to reinforce this effect.

Common mistakes (what looks like it should work but doesn't)

Obsessing over keyword density is the most common wasted effort we see. Keyword density has no measurable correlation with LLM citation rate in the published research, and the GEO paper confirmed it specifically. LLMs evaluate whether the content answers the question, not how many times the query phrase appears. Teams spending time on density optimisation are solving for a signal that does not exist in this context.

Relying on Article schema alone is technically correct but functionally insufficient. The majority of guides on "ranking in AI Overviews" point at Article schema. It tells the crawler what type of content a page is; on its own it does almost nothing to increase citation probability. The citation-driving schemas are FAQPage and DefinedTerm. Without them, you have put a "library" sign on a filing cabinet.

Writing generic "optimised for AI" content is the most self-defeating mistake. The content most likely to be cited by AI is content that demonstrably knows more than the AI does about the specific topic. Generic, well-structured content that exists everywhere on the web does not get cited because the LLM already has that content in its training data. What gets cited is the piece with a specific number, a named expert, the before/after data, the worked example the LLM could not generate on its own.

Chasing citations for every query rather than strategic ones spreads effort too thin. Structuring 500 pages for AI citation is a significant investment. Structuring the 20 pages that matter most — the queries your target customers actually ask before buying — produces a much better return. Start with the pages where you already have content that is close but not being cited. The gap is usually smaller than it looks.

A 3-step framework you can use this week

Takeaway

LLM visibility is measurable, trackable, and improvable — the same way traditional SEO has been for two decades. The measurement infrastructure is newer, the citation signals are different, and the optimisation levers have shifted away from keyword density toward answer density, schema specificity, and genuine content authority. But the underlying discipline is identical: understand what the system is optimising for, measure whether your content meets that bar, and close the gap.

If you do not know your current LLM citation rate across your important keywords, you are operating without a baseline. The 3-step framework above gives you a manual starting point you can run this week without any tooling. For teams who want to track at scale, the methodology we use is described in detail in PortableText [components.type] is missing "span".

Once you have measurement in place, the next question is what to do with it: the structural changes, schema updates, and content improvements that move citation rates in practice. That is the focus of the companion piece: PortableText [components.type] is missing "span".