A Ghost Ranking happens when AI uses your content as a source but recommends your competitors instead. Your brand becomes a ghost: present backstage, never on stage. The problem has a second, equally sneaky form: an AI ranking that changes with every test, even though nothing changed on your site. According to Seer Interactive, across 541,213 analyzed responses, a brand that is cited but not mentioned sees its citation rate drop to 10.6%, versus 53.1% when it is named (source at the bottom of the page).
And your brand, does ChatGPT recommend it?
Measure your presence and identify the brands cited in your place. No credit card.
Key takeaways
- A Ghost Ranking covers two problems: your content gets cited without your brand being named, and your AI ranking changes on every query.
- AI first decides which brands to recommend from its parametric memory, then looks for sources to justify them: the citation comes after, never before.
- The ranking itself is not reliable: ChatGPT returns the same list less than once in a hundred. The only solid metric is Share of Voice aggregated across dozens of prompts.
- You fix a Ghost Ranking through brand entity, not more content: Schema.org, Wikidata, brand language inside the text, and third-party mentions.
What Is a Ghost Ranking?
A Ghost Ranking is a phantom AI visibility: artificial intelligence uses your content but does not name your brand, or shows a ranking that changes on every query. The term covers two distinct conditions that marketers often confuse.
The first form is the ghost citation. Your article clears the engine’s relevance threshold. It gets retrieved as a trusted source. Yet in the visible answer, AI recommends a competitor and places your URL as a mere footnote. You are funding someone else’s recommendation.
The second form is the phantom rank. You test a query in ChatGPT and appear third. You run it again the next day and you are gone. Nothing changed on your site. The ranking itself is unstable to the point of meaning nothing in isolation.
Both problems share a common root. AI separates the recommendation task from the citation task. Understanding this split changes your entire AI visibility strategy.
Ghost Ranking or Ghost Citation: What’s the Difference?
The difference lies in the observed symptom. A Ghost Ranking is the umbrella term; the ghost citation is one specific case. Confusing them makes you apply the wrong fix.
The “cited without mention” case
Your domain appears in brackets under an answer. The text covers a topic you master. But your brand name appears nowhere in the prose. AI extracts your knowledge and leaves your identity behind. Seer Interactive calls this stranded brand equity: the value was created, the brand did not capture it.
The “unstable rank” case
Here you are sometimes named, sometimes absent, in an order that varies on every test. The problem is not your brand entity. It is the probabilistic nature of the engine. A SparkToro study shows there is less than a one-in-a-hundred chance of getting the same brand list twice when querying ChatGPT or Google AI a hundred times.
The table below summarizes how to tell them apart and what to do.
| Criterion | Ghost Citation | Unstable rank |
|---|---|---|
| Symptom | Content cited, brand never named | Presence and order change on every test |
| Cause | Weak brand entity in the knowledge graph | Probabilistic generation by the model |
| Fix | Strengthen the entity: Schema.org, Wikidata, brand language | Measure Share of Voice across many prompts |
| Common mistake | Produce more content | Report a “position” to a client |
Why Does AI Cite Your Content Without Naming Your Brand?
Because the AI recommendation and the citation rely on two separate systems. AI first decides which brands to name, then looks for sources to back that choice. The citation is the bibliography, not the brainstorm.
Here is how a ghost citation unfolds, as documented by Seer Interactive through behavioral tests. A user asks for the best solutions in a category. The model reaches into its parametric memory, the knowledge encoded during training, and surfaces the brands it associates with that category: your competitors. Only then does the retrieval step look for relevant documents. Your article is found. Your URL becomes a supporting footnote. The answer goes out: competitors recommended, your content cited, your brand silent.
The data signature is clear. When a brand is mentioned in a response, its citation rate climbs to 53.1%. When it is not, that same rate drops to 10.6%. If retrieval drove recommendation, those two numbers would be close. They are not.
This model remains a behavioral hypothesis, not proven architecture: no one outside the labs observes the generation logs. But the consistency of the numbers makes it very solid.
The consequence is direct. Adding content does not fix a ghost citation. You get more retrievable pages, with the same non-mention pattern. The problem is a brand entity recognition gap, not a content shortage.
Why Does Your Rank in ChatGPT Change on Every Test?
Because these engines are probability machines, built to generate a different answer every time. The rank you see one day has no reason to hold the next.
SparkToro ran twelve prompts through 600 volunteers, nearly 3,000 times in total, across ChatGPT, Claude and Google’s AI Overview. The verdict comes down to three variations in every response: the list shown, the order of recommendations, and the number of items. For order it is even more extreme: it takes around a thousand runs before seeing the same sequence twice.
The ranking also suffers from deep statistical fragility. MIT researchers showed that platforms ranking LLMs can flip by dropping a handful of votes. On one of them, removing 2 votes out of more than 57,000, just 0.0035%, was enough to change the top-ranked model.
- Never report an isolated rank position to a client: it does not survive a second test.
- Multiply your queries: SparkToro estimates you need 60 to 100 runs of the same prompt to get a stable signal.
One thing remains usable, though. If a brand shows up in 85 out of 95 responses, that appearance rate means something. Aggregated presence is solid; the one-off position is not.
How to Detect a Ghost Ranking on Your Brand?
You diagnose a Ghost Ranking by following three steps, matching three symptom branches. Each branch calls for a different fix.
Step 1 — Run the same recommendation prompt several times. Ask ChatGPT, Claude and Perplexity the typical question of your category, for example “what are the best tools to do X”. Repeat at least twenty to thirty times to neutralize inconsistency.
Step 2 — Record two things separately. On one side, is your brand named in the text? On the other, is your domain cited as a source? These are two distinct columns, never one.
Step 3 — Cross-reference with the diagnostic tree below.
- Cited AND named: no Ghost Ranking. Keep reinforcing your presence.
- Cited BUT never named: classic ghost citation. The work is on the brand entity.
- Neither cited nor named: an upstream visibility problem, often technical or editorial.
- Named one day, gone the next: unstable rank. The work is on aggregated measurement, not a page fix.
This manual diagnosis works for one brand and one category. Across several clients or several AIs, you move to a monitoring tool that automates these reruns and computes your Share of Voice directly.
How to Fix a Ghost Ranking? The 3 Levers
You fix a Ghost Ranking by working your brand entity across three layers. None produces overnight results: models re-index on their own schedule, so expect several weeks.
Lever 1: Make your brand inseparable from your key claims
Your name must become the grammatical subject of the ideas AI extracts from your pages. Stop writing “there are five approaches to the topic.” Write “at [Brand], our approach to the topic starts with.” If your name is not in the sentence, AI absorbs the idea and leaves your brand behind. You make name-free extraction impossible.
Lever 2: Build the entity graph AI reads
The gap that produces ghost citation sits between content relevance and recall of your name. You close it with machine-readable signals: a Wikidata entry, Organization-type Schema.org markup with a sameAs property on every page, a single canonical brand name everywhere, and an Author schema linking your named experts to the organization. Smaller competitors with a clean entity graph regularly outperform bigger brands here. Size does not protect you.
Lever 3: Earn third-party mentions in recommendation contexts
The model learned your competitors’ names somewhere: press, analyst reports, review platforms, industry publications. Every mention of your canonical name on an authoritative third-party domain strengthens the association between your brand and your category in the training data of future models. PR becomes directly useful to GEO again, as in the early days of PageRank. A mention reading “[Brand] is a leading player in the sector” in an industry headline is no longer a mere backlink: it is a training signal.
Which Metric Should You Track to Stop Getting Fooled?
Track aggregated Share of Voice, never rank position. Rank is a trap: it changes on every test and does not survive a check. Presence over volume stays statistically valid.
In practice, two indicators matter. The first is your appearance rate: across a hundred runs of a prompt, how often is your brand named? The second is the competitive ghost citation rate proposed by Seer: the percentage of your citations where you are cited as a source while a competitor is recommended. If it falls, your entity work pays off. If it rises, your content investment outpaces your brand investment, and the gap widens.
This metric shift protects your credibility. Presenting an AI rank position to a client means promising a number that collapses the moment they rerun the query themselves. Presenting a Share of Voice measured across dozens of prompts means delivering data they can verify without contradicting you.
FAQ
Is a Ghost Ranking a penalty from AI against my brand?
No. No AI deliberately penalizes your brand. A Ghost Ranking reflects an entity recognition weakness: the model does not associate you strongly enough with your category to name you when recommending.
How many times should I test a prompt to get reliable data?
Count on 60 to 100 runs of the same prompt to smooth out inconsistency, per SparkToro’s work. Below twenty, the result is still noise and should not be reported as a measurement.
Why doesn’t my cited content lift my brand mentions?
Because citation and mention rely on two separate systems. Content clears the retrieval threshold; mention depends on recalling your name from parametric memory. Strengthening one does not mechanically improve the other.
Is Schema.org markup enough to fix a ghost citation?
No, it contributes without being enough. Schema.org and sameAs feed the entity graph, but third-party mentions in recommendation contexts and brand language in your text matter just as much. The three levers work together.
How long before seeing an effect after a fix?
Several weeks, often two months. Models do not re-index in real time. A change made today acts on how the next model version perceives your brand, not on tomorrow’s answer.
Are Claude and Gemini affected by ghost citations?
The ghost citation diagnosis assumes the platform displays source URLs. Some responses do not expose these citations: the problem is not measured the same way there, but rank instability stays observable everywhere.
Sources
- John Lovett, Seer Interactive, “LLM Ghost Citations: Why Your Content Is Working and Your Brand Isn’t,” March 24, 2026 — seerinteractive.com
- Rand Fishkin, SparkToro, “New Research: AIs are highly inconsistent when recommending brands or products,” January 27, 2026 — sparktoro.com
- Adam Zewe, MIT News, “Study: Platforms that rank the latest LLMs can be unreliable,” February 9, 2026 — news.mit.edu
- Kevin Indig, Growth Memo, “The ghost citation problem,” April 20, 2026 — growth-memo.com
