Your engineering team can see it in the server logs. GPTBot, ClaudeBot, PerplexityBot: they're hitting your site regularly. Crawl volumes are up. Akamai measured 1.6 billion daily AI bot requests across its global CDN in early 2026, a 78% increase in six months. Your pages are being consumed at scale.
That information travels up to marketing as a signal: we're in the system. The bots are reading our content. We're being seen.
Except "being read" and "being credited" are two very different outcomes. And the distance between them is wider than most teams have measured.
The first filter: retrieval doesn't mean citation
Ahrefs published a study on April 15, 2026, analyzing 1.4 million ChatGPT prompts. The core finding: ChatGPT retrieves many pages in the process of generating a response, but it only cites roughly half of them. The rest get consumed (the model reads them, processes them, uses them to inform its answer) and then discarded without attribution.
The composition of what gets consumed but not credited is revealing. 67.8% of the non-cited URLs came from Reddit. The model is pulling heavily from community discussion, forum threads, and user-generated content. It's using that material to calibrate its understanding. And then it cites something else.
This is the first filter. Your content can be technically accessible, properly structured, and genuinely useful to the model's reasoning process, and still never appear as a source in the response the buyer reads.
The second filter: citation doesn't mean recognition
Now consider the content that does pass the first filter. It earned a citation. The URL shows up as a source link. On a GEO monitoring dashboard, this registers as a win.
Kevin Indig's research, published in Growth Memo in April 2026, measured something most dashboards don't track. His team analyzed 3,981 domains across 115 prompts, 14 countries, and four AI engines. The finding: 62% of all citations are ghost citations. The source URL exists in the response, but the brand name never appears in the answer text itself.
Think about what that means for the buyer. They ask ChatGPT a question. They get a detailed response. Your company's URL is listed in the footnotes. But your name, your brand, your identity as a solution provider: none of that showed up in the paragraph the buyer actually read. The model used your content and cited your page; it just never said your name.
Only 13.2% of brand appearances achieved both a citation (source link) and a mention (brand name in the answer text). Indig calls these "citation-mentions," and they're the only type of AI visibility that registers with a buyer the way a traditional search result does.
Seer Interactive's own ghost citation analysis found the same structural pattern: brands that are mentioned in the response text see a 5x citation rate lift compared to brands that aren't. Your content is doing the work. Your competitors are getting the recommendation.
AirOps measured this in September 2025: brands that are both mentioned and cited are 40% more likely to resurface across repeated runs. I built on that finding in Cited, arguing that the mentioned-vs-cited distinction should reshape how teams measure AI visibility. Indig's study, across 14 countries and four engines, settled it.
The platform split makes both filters worse
The ghost citation problem would be manageable if it behaved consistently across platforms. It doesn't.
ChatGPT cites sources 87% of the time. It mentions brands by name only 20.7% of the time. Gemini does the opposite: it mentions brands 83.7% of the time but only provides source citations 21.4% of the time.
Read those numbers again. ChatGPT is a footnoting machine that rarely says your name. Gemini is a name-dropper that rarely links to your site. They've developed opposite citation architectures, and a strategy optimized for one platform produces the wrong type of visibility on the other. The divergence isn't subtle: Superlines found that citation volumes for the same brand can differ by 615x between platforms like Grok and Claude.
A team tracking "AI visibility" as a single metric across all platforms is averaging signals that point in opposite directions. High citation rates on ChatGPT might mean your content is being sourced heavily but your brand is invisible in the answer text. High mention rates on Gemini might mean buyers hear your name but have no link to click.
What separates a ghost citation from a real one
The pattern in the data is consistent enough to be actionable: the type of content determines the type of visibility it earns.
Wix's March 2026 analysis found that informational content (articles, guides, how-to pages) gets cited at high rates on informational queries, with articles accounting for 45.48% of citations. But informational content feeds the model's reasoning without triggering a brand mention. It's structurally prone to ghost citation because the model is extracting facts, not endorsing vendors.
When a buyer asks "how does revenue operations work," the model pulls from your guide, uses your explanation, and cites your URL. It has no reason to say your name. The content answered a category question, not a vendor question.
Commercial and comparative content works differently. Listicles account for 40.86% of citations on commercial queries. Comparison pages, "best of" lists, evaluation frameworks: these formats put brand names into the answer because the content itself is organized around naming and evaluating specific solutions.
Growth Memo's March 2026 data reinforces this. Pages above 20,000 characters that cover "what is it," "who uses it," "how to choose," and "pricing" in a single URL averaged 10.18 ChatGPT citations each, compared to 2.39 for pages under 500 characters. The comprehensive pages that earn the most citations are the ones structured around the buyer's evaluation process. That's the context where the model names brands: when the content is explicitly about choosing between them.
The practical implication is that content strategy for AI visibility can't treat all citations as equal. Informational content builds the model's understanding of your category. That matters, but it produces ghost citations. Comparison and validation content builds the model's association between your brand and the buyer's decision criteria. That produces the citation-mentions that actually register.
What most teams aren't measuring
Stack the two filters together and the funnel looks like this: AI reads your content broadly. Roughly half of it never earns a citation at all. Of the half that does, 62% are ghost citations where the buyer never sees your brand name. The drop-off from "the model consumed your page" to "a buyer heard your name because of AI" is steep, and most GEO monitoring is measuring the middle of that funnel, not the end.
The teams reporting strong AI visibility numbers might be right about the citation counts. They might also be invisible to every buyer who reads those AI-generated answers without ever clicking a footnote. That's a measurement problem that looks a lot like the dark traffic blind spot: the numbers look fine until you realize you're reading the wrong numbers.
That's the gap. Whether you're measuring the right side of it is a different question.