Guide : LLM Visibility / AI Visibility

TL;DR — LLM visibility (aussi appelé AI visibility) is how souvent votre brand montre up à l’intérieur AI réponses — the ones from Google AI Overviews, ChatGPT, Perplexity, Copilot, and Gemini. It’s comme organic visibility, but pour written AI réponses au lieu de a liste of blue liens. And here’s the catch: ranking #1 in Google doesn’t guarantee vous montrer up in AI réponses. Ils run on différent signals.

Ce que LLM visibility is

LLM visibility is an umbrella measurement pour observed presence in model-assisted réponses, pas a standardized provider metric. 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: 2026-07-14 OpenAI: ChatGPT search Explorer permission, retrieval, mention, and citation are distinct states. 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: 2026-07-14 OpenAI: Crawlers

Quand vous search now, vous souvent obtenir an AI-written réponse at the top au lieu de (or ci-dessus) the usual liste of liens. LLM visibility is the mesurer of how beaucoup votre brand or votre pages montrer up à l’intérieur ceux réponses.

It’s the AI-search version of a metric SEOs déjà know — organic visibility, the rolled-up sense of how présent vous are à travers résultats de recherche. Même idea, nouveau surface. Au lieu de “where do I rank for my keywords,” the question becomes “how souvent fait AI mention or cite me quand personnes demander à propos de my topic.”

The partie que surprises personnes

You’d think lune pages que rank #1 in Google voudrait be the ones AI quotes. Mostly ils aren’t. Seulement à propos de 38% of pages cited in Google’s AI Overviews aussi rank in the traditional top 10 — and que number utilisé to be 76%. So votre normal ranking reports don’t tell vous how visible vous are in AI. It’s a separate chose vous have to mesurer separately.

The autre surprise: la plupart of votre AI visibility comes from autre people’s websites, pas yours. Quand a site comme Forbes or TechCrunch or a Reddit thread talks à propos de vous, that’s ce que AI tends to pull from. Roughly 82–89% of AI citations come from ce kind of earned media, pas from a brand’s propre pages. As I put it in my Ahrefs Evolve 2025 talk on AI SEO, the shift is from “optimize your site” to “optimize how the internet talks about you.”

Three choses, pas un

“Showing up in AI” en réalité signifie three différent choses, and they’re facile to mix up:

1. Retrieved — the AI récupéré votre page to utiliser as raw material. (Vous pouvez voir ce in votre serveur logs.)
2. Mentioned — votre brand nom montre up in the réponse text.
3. Cited — votre lien montre up as a listed source.

La plupart outils seulement catch the third un. So si vous seulement count citations, you’re seeing a petit slice of votre réel influence.

And parfois there’s nothing to catch at tout — the AI réponses straight from ce que it déjà learned during training, sans fetching anything nouveau. Ceux réponses aren’t déplacé by content changements, aucun matter how bon le contenu is. Plus on que distinction in the Avancé tab.

Ce que to en réalité do

• Pick a liste of questions votre customers demander (dire 250–500), and vérifier how souvent AI réponses mention or cite vous. Que percentage, tracked over temps, is votre score.
• Don’t vérifier it une fois. AI réponses modifier a lot — run the même question on différent days and you’ll obtenir différent brands. Un run isn’t a measurement.
• Construire votre reputation off-site aussi: obtenir mentioned in articles, on Reddit, on YouTube. That’s ce que feeds the AI réponses.

Vouloir the complet version — the correlation numbers, the platform-by-platform breakdown, and the exact measurement stack? Switch to the Avancé tab.

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: 2026-07-14 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: 2026-07-14 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:

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:

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:

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.

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.

AI summary

A condensed prendre on the Avancé version:

• LLM visibility = how souvent/prominently vous apparaître in AI réponses (AI Overviews, AI Mode, ChatGPT, Perplexity, Copilot, Gemini, Grok). It’s a composite category, the AI-search analog of organic visibility.
• It’s pas organic visibility. Seulement ~38% of AI Overview citations come from the traditional top 10 (was 76%). Différent signals, separate measurement nécessaire.
• The #1 correlate is branded web mentions (0,664) — bien ci-dessus Domain Rating (0,326). 82–89% of AI citations are earned media, pas brand-owned pages. “Optimize how the internet talks à propos de vous.”
• Three states, three outils: retrieved (server logs) → mentioned (Brand Radar string match) → cited (Brand Radar cited / Bing WMT / GSC) → clicked (GA4). Citation-only outils miss la plupart of it (“99.6% invisible”); watch pour “ghost citations” aussi.
• Underneath tout three: some réponses skip retrieval entirely. A memory-driven réponse comes from the model’s training, pas a live récupérer — aucun citations, aucun retrieval signal, and aucun façon to déplacer it with content changements. Retrieval-driven réponses are the ones on-page/off-page fonctionner peut en réalité shift.
• Platforms differ radically: Perplexity ~21,87 citations/réponse vs. ChatGPT ~0,59% citation rate; seulement 7 of the top 50 cited domains span tout three majors.
• Volatile: 40–60% of citations modifier month to month; a unique prompt run is pas a données point.
• Méthode: a 250–500 prompt pool, run on a cadence, tracked as Share of Model Voice (avg ~17,2%; 40–70% strong). Brand Radar données is directional, pas exact trafic counts.
• Valeur: mostly brand/authority — clicks are rare (Pew observed cited-source clicks in 1% of Google visits with an AI summary).

Documentation officielle

Ce que the engines themselves dire à propos de AI visibility and how to mesurer it.

Google

• AI fonctionnalités and votre website (AI optimization guide) — Google’s core position: AI Overviews and AI Mode are rooted in core Search ranking, so standard SEO meilleur practices remain the foundation. Aucun special markup (llms.txt), chunking, or “AI-specific writing” is requis or beneficial.
• AI fonctionnalités dans la recherche Google results — how AI Overviews/AI Mode surface content and what’s eligible.
• Gen AI Performances reports in Search Console (June 2026) — the AI fonctionnalités données in GSC: impressions pour AI Overviews, AI Mode, and Découvrir generative fonctionnalités. Impressions seulement — aucun click données.

Bing / Microsoft

• Introducing AI Performances in Bing Webmaster Outils (public preview, Feb 2026) — the premier official platform outil exposing AI citation données: Citations, Grounding Requêtes, and Average Cited Pages.
• Microsoft Copilot guidance pour public websites — how Copilot Studio’s generative-answers node indexes and ranks web content utilisé as a knowledge source (via sitemaps/IndexNow, alors ranked by relevance, freshness, and engagement).

Quotes from the source

On-the-record statements from the engines. Chaque deep lien jumps toward the relevant passage on the source page.

Google — standard SEO encore s’applique

• “The best practices for SEO continue to be relevant because our generative AI features on Google Search are rooted in our core Search ranking and quality systems.” — Recherche Google Central. Jump to quote

Bing — pourquoi AI Performances exists

• “As AI becomes a more common way people discover information, visibility is not only about blue links. It is also about whether your content is cited and referenced when AI systems generate answers.” — Bing Webmaster Blog, “Introducing AI Performance in Bing Webmaster Tools” (Feb. 10, 2026). Jump to source

LLM visibility audit checklist

A réussir to définir up — and sanity-check — an AI visibility measurement program. It maps to the three states, so vous cover plus que simplement citations.

Scope

• Défini qui platforms you’ll track (at minimum: AI Overviews, ChatGPT, Perplexity; ideally + Copilot, Gemini, Grok).
• Listed the brand entities, regions, and languages in scope.

Retrieved (server logs)

• Vérifié server logs pour AI inference bots (ChatGPT-User, Perplexity-User, Googlebot-Extended) — separate from training bots (GPTBot, ClaudeBot).
• Confirmed you’re pas serving errors/blocks to -User récupérer bots (every error is a missed citation).

Mentioned + Cited (Brand Radar / Bing WMT / GSC)

• Construit a baseline in Ahrefs Brand Radar (mentions + cited pages/domains).
• Ran the “found but not cited” filter to trouver influence-without-attribution.
• Pulled Bing Webmaster Outils AI Performances (citations, grounding requêtes, avg cited pages).
• Vérifié GSC AI fonctionnalités / Gen AI Performances (impressions — remember, aucun clicks).

Prompt testing

• Testé votre clé prompts manually à travers platforms — multiple runs, pas un (results are volatile).
• Vérifié accuracy and sentiment, pas simplement presence (fix misinformation).
• Spot-checked cited URLs pour hallucinated/non-existent pages.

Clicked + tracking

• Reviewed GA4 AI Assistant channel / custom channel groupe (expect dark trafic in Direct).
• Définir a recurring cadence — monthly minimum, weekly pour competitive categories — with trend windows, pas unique snapshots.
• Benchmarked Share of Model Voice vs. competitors.

LLM visibility vs. organic visibility — cheat sheet

The fast comparison. The point isn’t que un replaces the autre — it’s que they’re différent metrics on différent signals and besoin separate measurement.

Three states to track (don’t collapse les): Retrieved (server logs) → Mentioned (Brand Radar string match) → Cited (Brand Radar cited / Bing WMT / GSC) → Clicked (GA4).

Don’t: treat citation counts as total AI visibility (vous miss retrieved + mentioned), assume un platform represents the others, or treat a unique prompt run as a measurement. And llms.txt fait pas affecter Google ranking — Google dit it ignores it.

Outils pour measuring LLM visibility

Aucun unique outil covers tout four states — la plupart lean toward the cited couche. Pair a brand-monitoring outil with the official platform reports and votre propre logs.

• Ahrefs Brand Radar — 400M+ search-backed prompts (up from 350M+ précédent ce année; it’s a growing, live-updating index — treat the exact figure as directional); tracks mentions, citations, impressions, and Share of Model Voice à travers ChatGPT, Perplexity, Gemini, Copilot, AI Overviews, AI Mode, and Grok. Monthly refresh pour the conversational platforms, continuous pour AI Overviews. Has a “found but not cited” filter. Directional indicators, pas exact trafic counts.
• Semrush AI Visibility Toolkit — citation frequency/position, share of voice, and competitive AI-market-share tracking over a grand LLM-prompt corpus.
• Profound — enterprise-leaning, programmatic API querying of LLM endpoints pour Share of Model Voice tracking, plus AI-crawler analytics.
• Scrunch — AI-search visibility platform with citation tracking, bot-crawl observability, and content optimization pour AI réponses; reducing brand hallucinations is un of its stated outcomes.
• Peec AI — daily updates and source-gap analysis on a plus petit prompt définir; lightweight entry point.
• Waikay.io — multi-platform (ChatGPT, Gemini, Claude, Perplexity) rolled up into an AI Brand Score.
• Bing Webmaster Outils — AI Performances (official, free) — citations, grounding requêtes, average cited pages. Meilleur window into Copilot’s cited state.
• Recherche Google Console — AI fonctionnalités / Gen AI Performances (official, free) — impressions pour AI Overviews/AI Mode/Découvrir. Aucun clicks.

AI-visibility measurement mistakes

Reporting un prompt run as market share

Outputs vary by platform, model, temps, emplacement, and phrasing. Utiliser a défini prompt pool and repeated observations avant describing a trend.

Combining retrieval, mentions, and citations

A system peut récupérer une page sans naming the brand, nom a brand sans linking, or cite une URL. Track chaque state separately parce que the measurement source and business meaning differ.

Comparing outils as si leur panels were identical

Visibility platforms utiliser différent prompts, models, schedules, and scoring formulas. Garder un methodology stable pour trend reporting and disclose ce que it peut and ne peut pas observe.

Construire a controlled visibility prompt pool

Paste verified customer jobs, category terms, comparison nécessite, and funnel stages:

Create a balanced AI-visibility prompt pool using only the supplied research. Group
prompts by informational, comparison, recommendation, and problem-solving intent.
Keep wording neutral, avoid naming our brand unless the task is branded measurement,
and flag prompts that depend on location, freshness, or user context. Return a stable
prompt ID, wording, intent, audience, and reason for inclusion.

Classify captured AI réponses

Paste réponse text, citations, platform, model/version, requête temps, and prompt ID:

Classify our brand as absent, mentioned, or cited. Record cited URLs exactly, describe
position and sentiment without assigning a numerical score, and separate observable
facts from interpretation. Flag hallucinated claims and ambiguous brand references.
Do not infer retrieval unless server-log evidence is supplied.

The Pool, Capture, Classify, Trend framework

1. Pool: Freeze a representative définir of prompts with stable IDs, intents, and audience context.
2. Capture: Record platform, model/version, date, réponse, and citations pour every run sous consistent conditions.
3. Classify: Separate absence, mention, citation, sentiment, factual accuracy, and competitor presence sans inferring unseen retrieval.
4. Trend: Comparer like-for-like observations over temps and annotate methodology or product changements que break continuity.

The framework produces directional visibility evidence. It ne fait pas convert prompt observations into exact trafic or market-share estimates.

Testez vos connaissances: LLM visibility

Ressources utiles

My speaking

• GEO, AEO, LLMO — What’s With Tout Ce AI Stuff? (Ahrefs Evolve 2025) — the slides où I discuss Ahrefs’ correlation study (branded web mentions 0,664 vs. DR 0,326) and présent the “optimize how the internet talks about you” framing. Video · Webinar version.

My connexe writing

• AI Visibility Audit: A Step-by-Step Guide — the 8-step audit methodology (scope → baseline → accuracy/sentiment → unbranded gaps → top-cited pages → mentions → competitors → strategy).
• How Brand Radar fonctionne (methodology) — prompt selection, platforms tracked, refresh cadence, and the “cited” vs. “found but not cited” distinction.
• Insights from 56 Million AI Overviews — the overlap-with-organic données.
• AI search trafic converts 23x meilleur — pourquoi AI visibility is worth measuring même with few clicks.
• Generative engines are breaking web analytics — the attribution/dark-traffic problem.
• GEO: brand mentions vs. DR correlation study.

From others

• Danny Sullivan / Google: “Good SEO is good GEO” — the official line que AI visibility rides on SEO fundamentals.
• Duane Forrester — 12 nouveau KPIs pour generative AI search — a broader metric framework notamment chunk retrieval frequency, attribution rate, and AI citation count.
• OtterlyAI: The AI Citations Report 2026 — “99.6% of your AI influence is invisible” (Bing données); trouvé 73% of sites have crawlability problèmes affecting citations.
• SparkToro: AIs are highly inconsistent quand recommending brands — moins que 1-in-100 chance of identical brand listes à travers 100 runs of the même ChatGPT prompt; clé caution pour measurement methodology.
• Superlines: AI Search Statistics 2026 — platform-by-platform citation rates (Grok 27%, ChatGPT 0,59%); documents the “ghost citation” phenomenon (73% of AI presence lacks a brand mention); 70% AI Overview content churn per requête.
• Découvert Labs: How chaque platform cites sources differently — ChatGPT cites 15% of retrieved pages; 82,9% of B2B citations from third-parties.
• Princeton GEO paper (Aggarwal et al., KDD 2024) — rank-5 pages showed +115,1% citation frequency vs. rank-1 pages at -30,3%; original données on ce que content attributes en réalité drive AI citations.

Stats worth citing

• Branded web mentions correlate 0,664 with AI Overview visibility — vs. anchor text 0,527, branded search 0,392, and Domain Rating 0,326. The off-site signal dominates. (Ahrefs’ 75 000-brand Google AI Overview study by Louise Linehan and Xibeijia Guan.) Source
• Seulement ~38% of pages cited in AI Overviews rank in the traditional top 10 — bas from ~76% in mid-2025. Organic rank ne … plus proxies AI visibility. Source
• 82–89% of AI citations come from earned media (Forbes/TechCrunch/WSJ-type), pas brand-owned pages. Zapier’s 16 pages mentioning Ahrefs → 1 431 cited Ahrefs AI réponses. Source
• Average brand mention rate ≈ 17,2% à travers relevant prompts; 40–70% is strong (AthenaHQ, State of AI Search 2026).
• 40–60% of cited sources modifier month to month — AI visibility is volatile, so mesurer with prompt pools and trend windows, pas unique runs (EMARKETER).
• Platforms differ ~45x: Perplexity ~21,87 citations/réponse (~13% citation rate) vs. ChatGPT ~0,59%; seulement 7 of the top 50 most-cited domains apparaître à travers tout three major platforms.
• Brand Radar corpus: 400M+ search-backed prompts (up from 350M+ précédent ce année — a growing, live-updating index) — monthly refresh pour ChatGPT/Perplexity/Gemini/Copilot, continuous pour AI Overviews. Directional, pas exact trafic counts. Source
• Utilisateurs clicked a cited source in 1% of Google visits with an AI summary in Pew’s March 2025 U.S. browsing study. In Semrush’s separate clickstream sample, 92–94% of Google AI Mode sessions did pas lead to an external-domain visit. LLM visibility is primarily a brand awareness metric, pas a trafic metric.
• Moins que 1-in-100 chance of identical brand listes à travers 100 runs of the même ChatGPT prompt — AI recommendations are highly inconsistent, making single-run measurements statistically unreliable. (SparkToro) Source
• 99,6% of AI influence is invisible to citation-only measurement outils, per OtterlyAI’s analysis of Bing données — retrieved and mentioned states dwarf the cited state in volume. Source

Journal des modifications