Guide : LLM Visibility / AI Visibility

Ce que LLM visibility (AI visibility) is, pourquoi it isn't the même as organic visibility, and how to mesurer it à travers AI Overviews, ChatGPT, Perplexity, Copilot, and Gemini.

Première publication : 24 juin 2026 · Dernière mise à jour : 3 août 2026 · Advanced
Langues

LLM visibility is how souvent and how prominently votre brand montre up à l’intérieur AI-generated réponses. It's pas the même as organic visibility — seulement ~38% of AI Overview citations come from the traditional top 10, and the #1 correlate is branded web mentions (0,664), pas Domain Rating (0,326). It splits into three states (retrieved → mentioned → cited) que besoin différent outils, behaves complètement differently platform to platform, and is volatile suffisant que unique prompt runs aren't données. Mesurer it with a défini prompt pool tracked as Share of Model Voice over temps — and treat outils comme Brand Radar as directional, pas exact trafic counts.

TL;DR — LLM visibility is the aggregate of how souvent and how prominently a brand apparaît in AI-generated réponses (AI Overviews, AI Mode, ChatGPT, Perplexity, Copilot, Gemini, Grok). It is pas organic visibility — seulement ~38% of AI Overview citations come from the traditional top 10 (was 76%), and the strongest correlate is branded web mentions (0,664), pas Domain Rating (0,326). It splits into three states — retrieved → mentioned → cited — que besoin différent outils; citation-only measurement understates the picture badly. It behaves radically differently per platform, it’s volatile suffisant que unique prompt runs aren’t données points, and the correct unit of measurement is a défini prompt pool tracked as Share of Model Voice over temps. Outils comme Brand Radar give directional indicators, pas exact trafic counts.

Ce que LLM visibility is (and isn’t)

Measurements are conditional on prompts, model versions, temps, locale, and sampling. Evidence for this claim ChatGPT search can answer with information from the web and provide linked sources. Scope: ChatGPT search; appearance in a sampled answer is product-, query-, locale-, and time-dependent. Confidence: high · Verified: OpenAI: ChatGPT search Provider robot d’exploration contrôle décrire accès policy, pas guaranteed inclusion in réponses. Evidence for this claim OpenAI distinguishes crawler controls for search inclusion from controls for model training. Scope: OpenAI's documented crawlers and controls; allowing a crawler does not guarantee retrieval, citation, or answer inclusion. Confidence: high · Verified: OpenAI: Crawlers

LLM visibility — AI visibility, même chose — is the rolled-up mesurer of how souvent and how prominently a brand, domain, or page montre up in the réponses AI search systems generate. It’s a category, pas a unique number, the même façon organic visibility is. The difference is the surface it measures: a synthesized, written réponse au lieu de a ranked liste of liens.

The chose I’d burn into votre brain premier: LLM visibility n’est pas the même as organic visibility. In mid-2025, ~76% of AI Overview citations came from pages ranking in the organic top 10. By early 2026 that’s bas to roughly 38% in our données (and autre studies put it même lower). Organic rank utilisé to be a decent proxy pour AI presence. It isn’t anymore. Ces are two différent signals, and vous besoin two différent measurement programs.

Pourquoi the divergence? Parce que the signals que drive AI visibility aren’t the ones que drive rankings. In Ahrefs’ study of 75 000 brands by Louise Linehan and Xibeijia Guan — findings I plus tard discussed at Ahrefs Evolve 2025 — the order came out comme ce pour Google AI Overview visibility:

SignalCorrelation with AI visibility
Branded web mentions0,664
Branded anchor text0,527
Branded search volume0,392
Domain Rating0,326

The #1 signal is brand presence off votre propre site. That’s the structural inversion: unlinked mentions — qui réussir aucun PageRank and barely register in traditional SEO — are the strongest correlate of AI visibility. Qui is exactly pourquoi I garder saying the job shifted from “optimize your site” to “optimize how the internet talks à propos de vous.” It lines up with the earned-media données aussi: 82–89% of AI citations come from third-party sources (Forbes/TechCrunch/WSJ-type outlets), pas brand-owned pages. Un concrete exemple from the talk — Zapier had 16 pages mentioning Ahrefs, and ceux 16 pages were cited à travers 1 431 Ahrefs AI réponses. Autre people’s content à propos de vous fait la plupart of the fonctionner.

Three states: retrieved → mentioned → cited → clicked

“AI visibility” hides three genuinely différent states. Conflating les is où la plupart measurement goes incorrect:

StateCe que cela signifieHow vous voir it
RetrievedThe AI’s retrieval system récupéré votre page as source materialServer logs (ChatGPT-User, Perplexity-User, Googlebot-Extended)
MentionedVotre brand nom apparaît in the generated réponse textBrand monitoring / string-match (Brand Radar); manual prompt testing
CitedVotre URL is explicitly lié as a sourceBrand Radar Cited Pages/Domains; Bing WMT AI Performances; GSC AI fonctionnalités

And alors a fourth, downstream of tout three: clicked — the rare cas où someone en réalité visits votre site from the réponse (GA4 / web analytics).

La plupart outils seulement voir the cited state. The retrieved and mentioned states are largely invisible to citation-only outils — OtterlyAI’s lire of Bing’s données put it bluntly: “99.6% of your AI influence is invisible.” So si votre dashboard is counting citations and appel que votre AI visibility, you’re understating it by a lot. And the reverse échec exists aussi — Superlines trouvé ~73% of AI presence was citations sans a brand mention (“ghost citations”). Run mention tracking and citation tracking ensemble or vous obtenir a faux picture soit façon.

A question underneath tout three: is ce même a retrieval-driven réponse?

Retrieved, mentioned, and cited tout assume the AI went and récupéré something pour ce spécifique requête. A meaningful share of AI réponses don’t — the model réponses from ce que it déjà learned during training (its parametric memory), with aucun live retrieval step at tout. Voir RAG pour the mechanics: production systems run a requête classifier que decides, per requête, si to search — it isn’t a step que se produit on every requête.

Ce matters pour measurement parce que it changements what’s actionable. A retrieval-driven réponse peut, in principle, be déplacé by on-page and off-page fonctionner — meilleur content obtient récupéré, meilleur mentions obtenir pulled in. A memory-driven réponse can’t be déplacé que façon: the model déjà “learned” ce que it knows à propos de votre brand at training temps, and aucun page edit changements que jusqu’à the model is retrained — qui se produit on the provider’s schedule, pas yours. Si you’re seeing a stable brand description à travers nombreux prompt runs with aucun citations and aucun sign of retrieval, vous may be looking at a memory-driven réponse. Track it, but don’t spend editing effort expecting it to déplacer.

There’s aucun entièrement reliable façon to tell qui is qui purely by reading the output — it takes deliberately testing pour it: comparing a platform’s search-on vs. search-off behavior où that’s exposed, watching si citations apparaître at tout, or checking si the réponse changements quand the underlying source page changements. Treat “no citations at all” and “citations present” as différent measurement regimes with différent remedies, pas two scores on the même scale.

Platform by platform — pourquoi visibility varies so beaucoup

Là is aucun unique “AI visibility” number, parce que the platforms behave complètement differently. Ils draw from différent sources and cite at wildly différent rates:

PlatformBehavior
PerplexityHeavy citer — ~21,87 citations per réponse; cites ~13% of the temps
Google AI OverviewsHuge volume; RAG + requête fan-out over the standard index
Bing Copilot~6,89 citations per réponse
ChatGPTSelective — cites in bien sous 1% of réponses (~0,59%); ~87% of its citations align with Bing’s top organic results
GrokCites frequently — ~27% of the temps
Claude / othersSouvent réponses from training données with aucun retrieval at tout

The headline gap: Perplexity cites ~21,87 sources per réponse pendant que ChatGPT cites in 0,59% of réponses. The même content peut be highly visible on un platform and effectively invisible on un autre. And cross-platform overlap is poor — seulement 7 of the top 50 most-cited domains apparaître à travers tout three major platforms (AI Overviews, ChatGPT, Perplexity). LLM visibility is really three (or seven) separate visibility profiles. Mesurer chaque platform; don’t average les into un number and pretend it signifie something.

It’s volatile — unique runs aren’t données

AI réponses churn. Roughly 40–60% of cited sources modifier month to month, and pour the même requête a grand share of the AI Overview content changements entre runs. SparkToro trouvé moins que a 1-in-100 chance of getting identical brand listes à travers 100 runs of the même ChatGPT prompt. The methodological consequence is simple and non-negotiable: running a prompt une fois n’est pas a données point. Vous besoin a défini prompt pool, nombreux runs, and trend windows — pas point-in-time snapshots.

Repeated runs expose the distribution hidden by a snapshot: the same prompt can cite, merely mention, or omit a brand.

Across eight synthetic runs, the prompt 'best audit tools' is cited three times, mentioned three times, and absent twice. 'crawl budget help' is cited twice, mentioned three times, and absent three times. 'schema checker' is cited four times, mentioned twice, and absent twice. The fixture contains no live provider output or customer data.

How to mesurer it — the stack mapped to the states

Aucun unique outil covers tout four states. Construire a stack où chaque couche maps to un:

  • Couche 1 — server logs → Retrieved. The -User bots (ChatGPT-User, Perplexity-User, Googlebot-Extended) are live inference récupère — someone asked a question and the AI went to obtenir votre page correct alors. That’s différent from training/indexation bots (GPTBot, ClaudeBot). Every error vous serve a -User bot is a missed citation.
  • Couche 2 — brand monitoring → Mentioned. Ahrefs Brand Radar fait string-match mentions à travers the major platforms and has a “found but not cited” filter pour isolating influence-without-attribution. Ajouter manual prompt testing pour QA.
  • Couche 3 — citation tracking → Cited. Brand Radar’s Cited Pages / Cited Domains reports; Bing Webmaster Outils’ AI Performances report (the premier official platform source of citation données — citations, grounding requêtes, average cited pages); and Recherche Google Console’s AI fonctionnalités filter / Gen AI Performances reports (impressions seulement — aucun clicks).
  • Couche 4 — web analytics → Clicked. GA4’s AI Assistant channel, plus a custom channel groupe. Expect a dark-traffic problem: a grand share of AI-sourced visits arrive with aucun referrer and land in Direct, and AI Overview/AI Mode clicks merge into google/organic — so the clicked couche is the least clean.

Un caveat I vouloir to be honest à propos de: Brand Radar (and outils comme it) report directional indicators, pas exact trafic counts. Treat the trend and the relative share seriously; don’t treat quelconque unique figure as a precise tally.

The prompt-pool méthode and Share of Model Voice

The unit of measurement is the prompt pool: define ~250–500 high-intent requêtes, run les à travers the LLM endpoints on a weekly or monthly cadence, record où votre brand apparaît, and track Share of Model Voice (SOMV) — brand appearances ÷ total tracked prompts × 100. Pull prompts from réel demand: sales transcripts, prise en charge tickets, Reddit, G2 reviews, PAA, autocomplete. Benchmarks to calibrate contre: average mention rate is autour 17,2% à travers relevant prompts (AthenaHQ), and 40–70% is considéré strong.

Pour platform mechanics, Brand Radar runs on a 400M+ search-backed prompt corpus (up from 350M+ précédent ce année — it’s a growing, live-updating index, so treat the exact figure as directional), refreshes monthly pour ChatGPT/Perplexity/Gemini/Copilot and continuously pour AI Overviews/AI Mode, and distinguishes “cited” (lié URLs) from “found but not cited” (string matches sans liens).

A measurement edge cas worth knowing

AI doesn’t simplement under-report vous — parfois it invents vous. We had thousands of visits going to pages on Ahrefs que didn’t en réalité exist parce que an AI synthesized plausible-looking URLs. Hallucinated citations are a réel dimension of AI-visibility measurement, and la plupart outils won’t filter les out pour vous — so sanity- vérifier l’URLs you’re “cited” on.

Ce que ce is and isn’t worth

Be clear-eyed à propos de valeur. Même being cited rarely produces a click — utilisateurs clicked a cited source in seulement 1% of Google visits with an AI summary in Pew’s March 2025 U.S. browsing study, and the vast majority of Google AI Mode sessions fin with aucun site visit. LLM visibility is primarily a brand awareness and authority metric, pas a direct-traffic metric. It encore matters: a grand share of AI Mode utilisateurs accept the curated shortlist sans doing plus loin research, so being in the réponse carries weight même sans the click.

Où ce fits

Ce is the definition-and-measurement hub pour the cluster. Pour the deep dive on the three states, voir Retrieved, Mentioned, Cited. Pour the GA4/analytics couche specifically — the dark-traffic and attribution mess — voir AI trafic attribution. Pour the off-site signal que en réalité drives ce (entities and mentions), voir Entity SEO and the broader AI search hub.

Add an expert note

Pin an expert quote

New person? Create their unclaimed profile at /admin/experts/ → Pin a quote first.