1. What LLM visibility actually means
LLM visibility is the measurable rate at which a large language model retrieves and cites your brand when composing an answer to a query in your category. It is an observable metric, not a proxy. Where SEO visibility asks "where do we rank?", LLM visibility asks "when the model composes an answer, are we one of the sources it uses — and how prominently are we named?"
The distinction matters because the underlying mechanics are different. A search engine returns a ranked list of ten pages, and the buyer decides which to click. An answer engine composes a single response, cites two to five sources inline, and the buyer treats the composed answer as the shortlist. There is no page two. There is often no page one. There is an answer, and there are the brands the answer mentions.
LLM visibility is a discipline in its own right — separate from SEO, adjacent to PR, dependent on both. This guide is the working manual: what to measure, how to measure it, and what to build.
“LLM visibility is not a proxy metric. It is an observable one — either your brand appears in the composed answer, or it doesn't. Everything else is either a leading indicator or a downstream effect.”
2. The eight retrieval surfaces that matter in 2026
A meaningful LLM visibility program tracks visibility across a defined set of surfaces. The eight that matter for most B2B tech brands in 2026, ranked by our estimate of commercial retrieval weight:
- ChatGPT (with browsing / Deep Research / Atlas). Weighted heavily on tier-one press, analyst reports, .edu/.gov, and Wikipedia; increasingly retrieves from Substack, GitHub, and expert-tagged corners of Reddit and Stack Exchange.
- Google AI Overviews and AI Mode. Draws from Google's ranked index but with a strong bias toward pages that expose structured data cleanly and answer conversational queries in extractable form.
- Perplexity. The most citation-transparent surface — always shows sources, attributes clearly, and rewards pages with clean structure and named authors.
- Claude (with web search). Weighted toward long-form, authoritative sources; less prone to citing thin content than ChatGPT.
- Gemini. Draws heavily from Google's index but composes differently. Brand authority on Google translates well but not perfectly.
- Microsoft Copilot. Bing's index plus enterprise context. High-authority business sources (WSJ, FT, Bloomberg, Gartner) are weighted particularly heavily.
- Meta AI. Growing for consumer research; less relevant for most B2B tech, but worth monitoring in categories with prosumer overlap.
- TikTok search AI and other emergent surfaces. Category-dependent. For B2B tech, usually a secondary priority.
You do not need to instrument all eight. Pick the two that matter most to your buyers and track those on a running basis; treat the others as beneficiaries of the same infrastructure.
3. The signals each retrieval system weights
No LLM vendor publishes its retrieval ranking. But the observable behaviour across thousands of tracked queries — ours and those of the small research community that studies this — converges on a consistent set of signals. Ranked by observed weight:
Source authority in the curated corpus. LLM retrieval indices lean on a smaller, curated set of sources than Google's general index — tier-one national press, major trade publications, established analyst firms, .edu/.gov, Wikipedia, and a long tail of category-specific outlets. A domain absent from that corpus is comparatively invisible.
Entity resolution clarity. The model needs to attribute the extraction to a specific entity. A brand referred to inconsistently across the retrieved sources will either be dropped from the citation or misattributed to a competitor.
Passage extractability. Structured pages with labelled summaries, FAQPage schema, clear H2/H3 hierarchy, and explicit attribution outperform equivalent content without those signals. LLMs retrieve passages, not URLs.
Recency and freshness. Fresh pages with clear dateModified beat stale pages on time-sensitive queries. Signal freshness honestly with lastmod in sitemaps and updated JSON-LD dates.
Corroboration across sources. A claim that appears in three or more independent credible sources is far more likely to be extracted than one that appears in a single source, even a high-authority one. This is why distributed proof — the referring-domain layer — is prerequisite to the tier-one layer above it.
Owned assets on canonical URLs. Named frameworks, original data, and pillar pages hosted at stable, canonical URLs accumulate citation weight over time. Anonymous listicles and syndicated content do not.
4. Measuring visibility: the composite scorecard
A serious LLM visibility program measures five signals on a running basis. None is honest alone; together they produce a scorecard that survives audit.
- Prompt-set citation rate. Define 30–100 tracked prompts that represent the queries your buyers actually ask. Run them against your priority LLM surfaces on a schedule (weekly for competitive categories, monthly otherwise). Record: were you cited, in what position, and against what competitors. This is the visibility metric proper.
- Share of voice against category competitors. Over the same prompt set, how often are you cited relative to the five competitors that matter. This is the metric that travels to the boardroom.
- Referring-domain growth from the curated corpus. Weighted by whether the referring domain is one LLMs retrieve from. A backlink from a Substack a model reads is worth more than a backlink from a directory it doesn't.
- Branded and direct search lift. The dark-funnel proxy. If AI visibility is working, branded search volume rises without a corresponding paid or campaign spike.
- AI-referral session count. Where attributable — Perplexity and ChatGPT increasingly pass referrer strings. Instrument for it; don't rely on it as the primary metric.
Report the composite monthly. Report the individual signals to whoever owns them internally. Never present a single number as "our AI visibility" — the composite tells a truer story and withstands the questions a CFO will ask.
5. Instrumentation: how to actually collect the data
The tooling market is nascent. Several dedicated LLM visibility platforms are usable in 2026 (Profound, Peec, Otterly, AthenaHQ, Goodie, and others), and the mature enterprise SEO platforms have shipped LLM tracking modules of varying quality. Most Crackle clients run a hybrid: a dedicated tool for prompt-set tracking, a spreadsheet-and-audit process for the qualitative layer, and a lightweight internal script for the AI-referral session count.
Whatever you use, three principles hold:
Own the prompt set. Do not let a vendor pick your queries. The prompt set is the measurement instrument itself — it must reflect the actual questions your buyers ask, phrased the way they phrase them, and it must include competitor-named prompts as well as unbranded category prompts.
Track competitors in the same run. A citation rate without a competitor baseline is not decision-useful. Include the five to ten competitors you actually lose deals to.
Sample at meaningful cadence. LLM answers vary run-to-run. Sample each prompt at least three times per measurement window and record the modal citation, not a single reading.
6. Engineering plays that move the visibility number
The plays below map to layers 1–3 of the AI Citation Pyramid. They are the engineering base without which no earned-media investment compounds.
Entity clarity
Publish a canonical Organization + Person JSON-LD graph on every page. Establish a single sameAs cluster that spans your LinkedIn, Crunchbase, Wikipedia entry (if you have one), GitHub org, industry association pages, and every press mention that will let you self-identify. Canonicalise ruthlessly: one brand-name spelling, one preferred domain, one URL per canonical page, redirects for every variant.
Structured facts
Publish a glossary of the terms your category actually uses. Every term gets a DefinedTerm schema block, a definition, an example, and a link to the pillar page it belongs to. Add FAQPage schema to every money page — not decoratively, but with the actual questions buyers ask and answers you can defend. Add HowTo schema where you're explaining a process. Validate every page through Google's Rich Results Test.
Owned depth
Build pillar pages for the two or three topics you intend to own in your category. Each pillar should be 5,000+ words, contain a labelled TL;DR (with Speakable schema), a table of contents, internal links to related glossary and framework pages, an FAQ block, and named authorship. Name and document your proprietary frameworks — an unnamed model is unciteable. Publish original data at least twice a year and syndicate it correctly.
Freshness signals
Bump lastmod in sitemaps and dateModified in JSON-LD whenever you edit a public page. Refresh your top 25 pages on a rolling schedule — every 90 days minimum. Freshness is a retrieval signal in its own right; stale flagship pages get de-weighted regardless of how good they are.
7. Content patterns LLMs extract cleanly
Structure is not decoration. It is the difference between being read and being extracted. The patterns below repeat across every high-cited page we operate.
- Labelled TL;DR at the top. An
asidewith a clear "TL;DR" label and three to five bullet-form claims. Wrap it in Speakable schema. LLMs cite this block disproportionately often. - H2 hierarchy that mirrors likely questions. Sections titled the way buyers phrase the query, not the way marketing teams write headlines.
- Attributed pull-quotes. A short, extractable sentence attributed to a named person with schema markup. LLMs are far more likely to quote an attributed claim than an unattributed one.
- Explicit lists over paragraphs. For any content that reduces to a list, use a list. LLM extractors love ordered and unordered lists.
- Named entities on first mention. Products, frameworks, and companies get their full canonical name the first time they appear, with an internal link to the entity page. This is how you build the entity graph the retrieval layer reads.
- Author bylines and Person schema. Every substantive page has a named author with a linked bio and JSON-LD Person markup. Anonymous pages under-perform on citation.
- An FAQ block that answers real questions. Not filler. Six to twelve questions, each answered in 60–120 words, with
FAQPageschema. This is often the block that actually gets extracted.
8. Distribution: earning the citations you can't buy
Engineering gets you extractable. Distribution gets you cited. The Earned Media Flywheel covers this end-to-end (see the framework page); the LLM-visibility-specific summary is:
Tier-one press remains the highest-weight input. A single well-placed WSJ, FT, NYT, Bloomberg, Reuters, Axios, or Information story that names your brand as the primary entity does more citation lift than any volume of self-published content. This is why the pyramid terminates in tier-one — nothing else replaces it.
Analyst coverage is under-priced. Gartner, Forrester, IDC, and category-specific boutique analysts publish into corpora LLMs weight extremely heavily. A single named appearance in a Gartner Hype Cycle is worth a quarter's worth of blog posts on citation lift.
Podcast citations compound faster than they look. Every episode with a durable URL and clean show notes creates an extractable source. Being a repeat guest on the podcasts your category actually listens to builds distributed proof at low marginal cost — one reason we invest in Under Embargo as a first-party asset.
Wikipedia matters more than people admit. If your company is genuinely notable (independent tier-one coverage exists) it belongs on Wikipedia. Do not attempt to write the page yourself; do make sure the tier-one coverage that supports it exists.
Substack, GitHub, and expert-tagged Reddit are the new mid-tier. LLM retrievers increasingly weight these surfaces when the authorship signal is strong. Sustained presence in the right sub-communities is a durable moat.
9. What good looks like: benchmarks by stage
Approximate ranges from Crackle engagements and public research. Treat them as goalposts, not promises — categories vary widely.
Early-stage startup (Seed to Series A)
Realistic target after 6 months of consistent work: cited in 10–20% of tracked unbranded prompts for the specific niche you're competing in; 100–200 total referring domains, half from the curated corpus; recognised as an entity by at least ChatGPT and Perplexity.
Growth stage (Series B to D)
Realistic target: cited in 30–50% of tracked prompts across the primary category; 500–1,500 referring domains with meaningful curated-corpus representation; measurable share-of-voice against the top three competitors in at least two LLM surfaces; a Wikipedia entry.
Late stage / public
Realistic target: cited in 60%+ of tracked prompts; multiple tier-one mentions per quarter; original data referenced by analysts and press unprompted; the brand functions as a category proxy in a meaningful share of AI answers.
If your program is stuck below these bands after 90 days of sustained input, the issue is almost always in the base layers of the pyramid — entity clarity or structured facts — not in the volume of content or press. Fix the base; the visibility follows.
10. Frequently asked questions
Next steps
Score your program against the pyramid, then talk to us about what to build first.