Ghost citations: Why AI search cites your content, not your brand

Throughout roughly 16 million model appearances, new information from Writesonic discovered that round 40% of AI citations didn’t identify the supply model within the generated reply. On Perplexity, the speed reached 52%. (Disclosure: The article co-author, Samanyou Garg, is the founder and CEO of Writesonic.)

Getting cited by an AI engine is usually touted as a win. You probably did the work, and now your web page seems within the sources, your URL exhibits up in quotation reporting, and the engine has discovered your content material helpful sufficient to help its reply.

However one factor folks in our subject aren’t typically speaking about is the truth that even with a quotation, the reader could by no means see your model identify.

We name this a ghost quotation: an AI engine hyperlinks to your web page as a supply however doesn’t point out your model within the reply itself.

That distinction modifications what a quotation really tells you. It confirms that your content material contributed to the reply, however not that your model was seen to the reader.

Key findings

• Round 40% of citations didn’t identify the supply model within the reply.
• Ghost quotation charges ranged from 19% to 52% throughout engines.
• Perplexity had the best price at 52%, adopted by Google AI Mode at 49%.
• Google AI Overviews reached 41%, whereas ChatGPT reached 37%.
• Gemini, Grok, and Microsoft Copilot had decrease charges at 25%, 22%, and 19%, respectively.
• The engines broadly break up between those who hyperlink steadily and those who identify manufacturers extra constantly.
• Quotation price and model visibility ought to due to this fact be measured as separate outcomes.

A quotation doesn’t imply your model was seen

A quotation means an AI reply hyperlinks to your web page as a supply. It could seem as an inline reference, a footnote, or a hyperlink inside a sources panel.

A point out means your model identify seems within the generated reply textual content. Contemplate these two solutions:

With a quotation and a point out, the discovering is linked on to Writesonic, and the underlying analysis can be linked:

• “Writesonic’s evaluation of roughly 16 million model appearances discovered that roughly 40% of AI citations didn’t identify the supply model.”

However a ghost quotation appears totally different. The Writesonic web page nonetheless seems within the sources, however except the person notices and opens the quotation, there’s nothing within the reply connecting the discovering to the corporate that produced it.

• “One evaluation discovered that roughly 40% of AI citations didn’t identify the supply model.”

Each solutions may hyperlink to the identical Writesonic web page. However solely the primary one tells the reader who produced the analysis.

That’s the ghost quotation hole: your content material is seen to the engine, whereas your model could stay invisible to the reader.

This hole turns into measurable when citations and mentions are tracked individually. As an alternative of treating each supply hyperlink as the identical consequence, groups can see the place they had been cited, the place they had been named, and the place the 2 alerts separated.

Perplexity drops the model identify in 52% of citations

The ghost quotation price diverse sharply by engine.

Ghost quotation price by AI engine. Notice: Google AI Mode (49%) is omitted for readability. Its conduct intently mirrors AI Overviews.

Perplexity had the best price within the research. In 52% of its cited appearances, the engine linked to a model’s web page with out naming the model within the reply.

Google AI Mode adopted at 49%, whereas Google AI Overviews reached 41%. ChatGPT sat barely beneath the general common at 37%.

The speed fell to 25% for Gemini, 22% for Grok, and 19% for Microsoft Copilot.

This implies two manufacturers can earn the identical variety of citations and obtain very totally different ranges of recognition relying on the engine.

Out of 100 cited appearances on Perplexity, the model can be absent from the reply textual content in roughly 52. On Copilot, it could be absent in roughly 19.

So the quotation depend may look equivalent in a report, however the variety of occasions the reader really sees the model wouldn’t.

AI engines break up into namers and citers

The distinction turns into clearer when quotation price and point out price are in contrast.

• Namers, together with Gemini and Microsoft Copilot, usually tend to embrace manufacturers within the reply textual content however hyperlink to their pages much less steadily.
• Citers, together with Perplexity and Google’s reply engines, hyperlink to pages extra steadily however are much less constant about naming the manufacturers behind them.

ChatGPT and Grok fall between the 2 teams. Neither consequence is right.

A point out with out a quotation makes the model seen however could not present the supporting supply and encourage referral clicks to the web site. A quotation with out a point out makes the web page accessible however could depart the model out of the response, arguably lacking the extra essential consequence: model recognition.

The strongest consequence is receiving each. This break up additionally signifies that to enhance visibility throughout AI engines, your optimization technique ought to differ by engine, which other industry research backs up, too.

Chasing extra citations on Perplexity could improve your supply depend with out rising how typically your model seems within the reply. On Gemini or Copilot, nevertheless, you’re extra more likely to be named, however you have to be cited extra constantly to drive referral site visitors.

Why do ghost citations occur

Ghost citations are possible the results of how totally different AI engines course of and synthesize sources moderately than any deliberate option to omit model names.

However the trustworthy reply is: our subject doesn’t have a whole clarification but.

One potential clarification is that some engines deal with the linked quotation as enough attribution after which current the knowledge in their very own “voice.” 

Take this instance beneath. Google’s AI-generated reply to “how one can make GEO worthwhile” gives suggestions and takeaways just like these within the highlighted citations, however its model, tone, and phrasing are distinctive.

The engine has absorbed the supply supplies and rewritten them, leaving the quotation within the footnotes however the model identify out of the reply.

Content material construction may play a job. When a model identify isn’t intently paired with the declare it produced on the supply web page, synthesis has extra room to sever them.

A statistic buried three paragraphs beneath the corporate identify that generated it’s most likely simpler to anonymize than one the place attribution is constructed into the sentence itself. This isn’t a confirmed rule, however one thing I’d counsel testing.

What we don’t but know is the relative weight of every issue, or whether or not there are others. What the engine-level variance does verify is that the charges aren’t uniform, which provides us patterns price investigating even when we are able to’t but clarify what drives them. 

Tips on how to flip citations into model recognition

There’s no confirmed technique that ensures an AI engine will identify a model each time it cites its URL. However analysis and information we’ve seen level to those sensible steps:

• Begin with measurement, as a result of you’ll be able to’t repair what you’re not monitoring.
• Construct proprietary information as your model moat.
• Make attribution specific in your pages.
• Reinforce your model identify throughout third-party sources.

1. Measure mentions and citations individually

Classify tracked appearances into 4 teams:

• Cited and talked about.
• Cited however not talked about.
• Talked about however not cited.
• Neither cited nor talked about.

Then section these outcomes by engine. The research exhibits that an aggregated quotation rating can cover giant variations between platforms.

2. Construct named analysis your model can personal

Proprietary analysis or information can develop into your brand moat as a result of it could be more durable to ghost than generic recommendation. When a research, index, framework, or methodology carries your model’s identify, the information and the model develop into more durable to separate as a result of the discovering references you by definition.

Named analysis additionally provides journalists a stronger hook to cowl, since reporters usually tend to choose up research with a transparent methodology identify.

That earned protection issues doubly now: third-party content material will get cited extra closely by AI engines, so the offsite PR your named analysis generates compounds into extra AI citations over time.

3. Pair your model identify instantly with the declare, not simply close by 

One possible issue price testing is the proximity of your model identify to a key stat or discovering on the web page. The speculation is {that a} statistic that seems a couple of paragraphs away from the corporate identify that produced it’s simpler to anonymize than one the place attribution is constructed into the sentence itself.

So as a substitute of:

• “Our evaluation discovered that round 40% of AI citations didn’t identify the supply model.”

Use:

• “Writesonic’s Ghost Quotation Examine discovered that round 40% of AI citations didn’t identify the supply model.”

Named research, datasets, indexes, and frameworks could make the connection between an perception and its creator clearer.

That affiliation shouldn’t stay solely by yourself website. Impartial protection, knowledgeable commentary, buyer content material, and business discussions throughout the sources that the majority affect AI solutions can repeatedly join your model with the concepts it needs to personal.

4. Examine the solutions behind the metric

The ghost quotation price identifies areas of low brand mentions, and the subsequent step is to evaluate the person solutions to assist diagnose what to check. For essential prompts the place your web page is cited however your model is lacking, overview:

• Which web page was cited?
• What data was extracted?
• The place did the model identify seem on the supply web page?
• Have been different manufacturers named within the reply?
• Does the sample repeat throughout associated prompts?
• Does it differ by engine?

Then check whether or not clearer attribution, constantly named analysis, or modifications to web page construction have an effect on the consequence over time.

The following GEO metric is recognition

A supply hyperlink confirms that an AI engine used your content material. It doesn’t verify that the reader noticed your model. That’s why quotation price can’t be the one measure of AI visibility.

GEO measurement wants to trace not solely whether or not an organization appeared someplace within the sources, however whether or not it was acknowledged within the reply. Quotation price tells you that an engine discovered your content material. Point out price tells you whether or not your model really traveled with it.

We observe each individually for each shopper as a result of a quotation your reader by no means sees is visibility that solely exists in a dashboard.

Methodology

Writesonic analyzed a current 30-day window of AI reply information throughout 1000’s of manufacturers.

The ghost quotation evaluation lined roughly 16 million model appearances. The engine-level charges offered on this article cowl Perplexity, Google AI Mode, Google AI Overviews, ChatGPT, Gemini, Grok, and Microsoft Copilot.

• A quotation was recorded when an AI reply linked to a model’s web page as a supply. A point out was recorded when the model’s identify appeared within the generated reply textual content.
• A ghost quotation was recorded when a web page was cited, however the model wasn’t named within the reply.

Percentages had been rounded. The outcomes signify a snapshot of engine conduct in the course of the measurement window and don’t set up the causes of ghost citations or their impact on clicks, model recall, belief, or conversions.