A new Growth Memo analysis published last week, by Kevin Indig, ran 115 prompts across ChatGPT, Google AI Overviews, Gemini, and Google AI Mode in 14 countries, recording 3,981 domain appearances. The headline finding is that 61.7% of AI citations are “ghost citations.” A domain gets a source link, but the brand name never appears in the answer text the user actually reads. Only 13.2% of brand appearances produce both a citation link and a visible mention. The data was pulled from the Semrush AI Toolkit and represents one of the largest cross-engine analyses of AI citation behavior published to date.
The deeper finding is that the four major AI engines treat citations and mentions in fundamentally opposite ways. Gemini names brands in 83.7% of its appearances but only generates a citation link 21.4% of the time. ChatGPT does the opposite: it cites sources 87% of the time but mentions brand names in only 20.7% of answers. AI Mode behaves like ChatGPT (76.3% citation, 37.6% mention). Only AI Overviews come close to balanced, with 84.9% citations and 61% mentions.
Our read: most B2B AEO dashboards being sold today report a single “AI visibility” number that averages across these engines. That number is structurally misleading. A brand can look dominant on Gemini and invisible on ChatGPT for the same prompt, or get cited 9 times by Google without ever being named. For B2B teams whose buyers are evaluating shortlists inside AI tools, the question isn’t whether you appear; it’s whether buyers ever read your name. Citation counts have been the wrong KPI.
A practical AI visibility diagnostic can separate real citation risk from vanity dashboard movement.
Key Takeaways
- Growth Memo’s April 2026 analysis of 3,981 domain appearances across ChatGPT, AI Overviews, Gemini, and AI Mode found 61.7% of AI citations are “ghost citations,” meaning a source link is generated but the brand name never appears in the answer text.
- Only 13.2% of brand appearances produce both a citation link and a visible mention, which means most “AI visibility” reporting overstates real brand exposure by a wide margin.
- The four engines diverge sharply: Gemini mentions brands 83.7% of the time but cites 21.4%; ChatGPT cites 87% but mentions 20.7%; AI Mode 76.3% / 37.6%; AI Overviews 84.9% / 61%.
- Aggregator and academic sources (Medium, Wikipedia, Wired) are repeatedly cited but almost never named, while consumer brands are repeatedly named without being cited; the gap matters because buyers only register what they read.
- For B2B teams, the implication is to track citation and mention as separate metrics per engine, prioritize comparison and evaluation content (which produces mentions, not just citations), and accept that LLM-specific strategies are not interchangeable.
What “Ghost Citation” Actually Means for B2B Teams
A citation is when an AI engine includes a clickable source link in its answer. A mention is when the brand name appears inside the answer text the user reads. Most B2B teams have been told these are the same thing because both show up in AI visibility tools as an “appearance.” Indig’s analysis demonstrates they are not. 74.9% of domains in the dataset were cited; only 38.3% were mentioned. The gap of 36.6 percentage points is the ghost-citation problem in one number.
The mechanic matters because of how buyers actually use AI search. When a B2B buyer asks ChatGPT “best CRM for mid-market manufacturing,” the answer text lists three or four named vendors. Below that, in smaller text or a sidebar, are source links. Eye-tracking and click data from Perplexity’s referral decline and parallel ChatGPT studies show that fewer than 1% of users click citation links. The named brands in the answer body are the entire game; the cited domains are footnotes nobody reads.
This is why “we ranked first in citations” reporting has been misleading B2B marketing teams. A team can look like the most-cited domain in their category and still be invisible to buyers because the answer text names three competitors and treats the team’s content as anonymous source material. Aggregators take the worst of this: Medium.com was cited 16 times across the dataset and named zero. Wikipedia and Wired follow the same pattern. Consumer brands like Facebook show the inverse: Gemini named Facebook in 3 of 3 appearances, while Google AI cited Facebook 9 of 9 times and named it once. Same brand, same query, different engines, opposite outcomes.
Why the Four Engines Behave Differently
The Growth Memo data shows each engine has a distinct posture toward attribution.
Gemini behaves like a conversationalist drawing on brand knowledge. It freely names brands in its answers (83.7%) but rarely sources them (21.4%). Google’s reasoning, based on integration patterns, appears to be that Gemini is optimized for conversational fluency over academic rigor. For B2B brands, this is the highest-stakes engine for entity authority: if Gemini doesn’t already know your brand from web-scale training data, you won’t be named, and citing your way in doesn’t help.
ChatGPT behaves like an academic paper with footnotes. It cites sources 87% of the time but names brands in only 20.7% of answers. ChatGPT lifts content but synthesizes it into a generic recommendation paragraph. For B2B brands, this means optimizing for ChatGPT citation is a different exercise from optimizing for Gemini mention. The two strategies don’t transfer, and the ChatGPT citation patterns we documented earlier this year reinforce the point: review-rich third-party domains lift far more often than vendor sites do. The volatility 5W just documented shows the lift itself is unstable: ChatGPT’s Reddit citation share fell from roughly 60% to 10% in six weeks in late 2025 after a single Google parameter change.
Google AI Mode resembles ChatGPT (76.3% citation, 37.6% mention) but with about 17% more brand mentions in outputs. AI Mode is leaning toward citation discipline while still naming brands more freely than ChatGPT. The behavior pattern aligns with Google’s AI Mode advertising rollout incentives: Google needs sources for E-E-A-T signal but mentions to make the ads-adjacent surface feel branded. The May 19 upgrade of AI Mode to Gemini 3.5 Flash as the default model raises the stakes on both metrics, since the same engine now serves both one-shot queries and the persistent re-checks information agents will run against B2B vendor surfaces.
AI Overviews are the closest to balanced at 84.9% citations and 61% mentions. AI Overviews are the most public-facing surface (used by 1.5 billion people monthly) and Google has had the most pressure from publishers to maintain citation discipline. For B2B teams, AI Overviews remain the highest-payoff surface to optimize because mentions and citations move together. That payoff just grew: a labeled Preferred Source slot now sits inside AI Overviews, so an owned audience can convert citation discipline into a visible, named appearance instead of another ghosted link.
Country-level differences add another layer. India and Sweden show the highest mention rates at 50%, suggesting more brand-forward query patterns. Italy, Brazil, and the Netherlands sit at 18-22% mentions with very high citation rates of 82-94%. Aggregate “global AI visibility” metrics mask all of this. They also mask the underlying retrieval-crawl footprint AI engines bring to B2B sites, which is what determines whether your domain is even eligible for the citation slot in the first place.
What B2B Teams Should Change This Quarter
Five operational adjustments based on the data.
Track citations and mentions as separate KPIs, per engine. Stop reporting a single “AI share of voice” number. Build a 4×2 grid: ChatGPT, Gemini, AI Mode, AI Overviews on one axis; cited and mentioned on the other. The 13.2% of appearances that produce both is your real brand visibility. The other 86.8% is either ghost citation or unsourced mention. Both have value, but they’re different problems with different fixes. For B2B agencies specifically, the measurement layer just got cheaper: Clutch’s free baseline dashboard for AI agency visibility handles the four-engine citation tally for any agency already listed on the platform, removing the $499-$699-per-month tool tax that previously kept this discipline out of reach for mid-size shops. The B2B brand-description drift audit adds a third KPI to that grid: whether the answer is accurate when the brand is named.
Prioritize comparison and evaluation content for brand mentions. Indig’s data shows that comparative content (X vs Y, best-of, evaluation) produces mentions, while informational content (definitions, explainers) produces ghost citations. For B2B teams, this means doubling investment in evaluation-stage content like vendor comparison pages, alternatives lists, and use-case-specific shortlists. These are the content types AI engines lift with the brand name attached.
Run an LLM-specific strategy, not a unified one. A Gemini-first content strategy (entity authority, brand recall, third-party citations from DA-60+ sites) is different from a ChatGPT-first strategy (citation density, structured FAQs, source-link-friendly formatting). Indig’s data shows minimal overlap in which sources each engine cites for the same prompts. Pick a primary engine based on where your buyers actually evaluate, then build the secondary strategy on top.
Accept that aggregators are working against you. If your AEO strategy depends on getting picked up by Medium, Wikipedia-style aggregators, or generic listicles, you’re producing ghost citations for someone else’s domain. The lift goes to the aggregator’s source link; the mention, when it happens, goes to the brand the aggregator named. Build your owned domain authority first, then use third-party placements to reinforce.
Audit prompt phrasing in your AI visibility tracking. Short conversational queries and long structured queries produce mentions and ghost citations at different rates. The same brand can appear “visible” on one phrasing and absent on another. Most B2B teams test 5-10 prompts; the Growth Memo dataset suggests 100+ prompts across multiple phrasing styles is the floor for usable measurement. The next question is who owns the work, because fewer than 15% of B2B teams have a dedicated GEO leader despite widespread experimentation.
The broader pattern: 2025’s “get cited by AI” advice was the wrong KPI for most B2B teams. Forrester’s GTM singularity argument is built on the premise that buyer preference is forming inside AI surfaces. If that premise is right, then mentions, not citations, are what shape preference. Ghost citations are work the brand did that the buyer never saw. Closing that gap is the AEO project for the rest of 2026. The supply of citable content is shifting under it too: CNN’s suit against Perplexity over 17,000 copied works signals the sources feeding AI citations may soon carry licensing terms, reshaping which domains engines are willing to cite at all.
Frequently Asked Questions
A ghost citation is when an AI engine like ChatGPT, Gemini, AI Overviews, or AI Mode generates a source link to a domain in its answer but never names that brand in the answer text the user reads. The term was coined in Kevin Indig’s April 2026 Growth Memo analysis. The Growth Memo dataset of 3,981 domain appearances found that 61.7% of AI citations are ghost citations, meaning the brand gets a footnote but no recognition in the body of the response.
Buyers register what they read in the answer text, not source links they rarely click. Mentions translate into brand awareness and shortlist consideration; citations contribute SEO-style authority signal but limited direct brand recall. For B2B teams whose buyers are evaluating vendors inside AI tools, mentions are the conversion-adjacent metric. Tracking only citations overstates real visibility by roughly 4-6x according to the Growth Memo data.
Gemini names brands in 83.7% of its appearances, the highest mention rate of the four major engines analyzed. AI Overviews come second at 61%. AI Mode mentions brands 37.6% of the time, and ChatGPT names brands in only 20.7% of answers despite citing sources 87% of the time. The data comes from Kevin Indig’s April 2026 Growth Memo analysis of 115 prompts across 14 countries.
Track citations and mentions as separate KPIs across each engine, build a 4-by-2 measurement grid covering ChatGPT, Gemini, AI Mode, and AI Overviews, and treat the small overlap between cited-and-mentioned (13.2% of appearances) as your real brand visibility number. Run at least 100 prompts across multiple phrasing styles, since short conversational queries and long structured queries produce different mention rates.
