Ghost Citations: Cited by AI, Never Recommended

A ghost citation is when an AI engine links to your page as a source and recommends someone else in the same answer. Your content earned the footnote; your brand did not earn the sentence. Seer Interactive analyzed 541,213 LLM responses across 20 brands and six AI platforms in February 2026 and found the gap is structural: a brand’s citation rate was 53.1% when the answer mentioned it and 10.6% when it did not. Being named and being cited are different games, and most startups are only playing one of them.

The two decisions inside one AI answer

Read any AI answer about a product category and you are actually looking at two separate outputs stitched together. The first is the recommendation: the three or four brands named in the prose. The second is the source list: the links under or beside the text. Founders read the answer as one thing, so a citation feels like a win. It often is not.

Seer’s hypothesis, which they support with six independent behavioral tests across 362,188 responses, is that citations are post-hoc. The model drafts the answer from what it already knows, choosing brand names out of its training, and only then retrieves pages that back up what it just said. In that order, your carefully structured comparison page is not competing to be recommended. It is competing to be the bibliography for a recommendation someone else won.

This explains a pattern that otherwise makes no sense. You do everything the generative engine optimization playbook asks. Your citation count climbs. Your name still never appears. The content layer is working and the brand layer is empty.

Why it stings more than a missing ranking

A missing ranking costs you a click. A ghost citation costs you the click and hands the purchase intent to a competitor, using your own research to do it. The reader gets your data, your framing, and your definitions, then reads a sentence recommending a product that is not yours.

It also flatters your dashboards. If you built the GA4 setup that isolates AI traffic, ghost citations show up as healthy: sessions from chatgpt.com and perplexity.ai, arriving on your best pages. The traffic is real. What the report cannot show you is the sentence above the link, where a competitor was named. Two brands can have identical AI referral charts and completely different outcomes.

Seer found that the problem concentrates where it hurts most: awareness-stage prompts, the ones where a buyer is still deciding which brands exist, carried the highest competitive ghost rate in their data. Category leaders showed near-zero ghost citations. Being the answer and being the footnote sort out early.

What actually correlates with getting named

If citations are post-hoc, then the lever is whatever put your brand in the model’s memory in the first place, and that is almost entirely off your own website.

Ahrefs studied 75,000 brands and ran Spearman correlations between branded signals and AI visibility across ChatGPT, Google AI Mode, and AI Overviews. The ranking is uncomfortable for anyone who spent the last decade on on-site work.

The top three are all things other people do about you. The bottom two are the things most teams control and optimize obsessively. Ahrefs is explicit that correlation is not causation, and you should hold that caveat firmly: nobody has proven that buying mentions moves citations. But the shape of the data matches the post-hoc model of how answers get built, and two independent research teams arriving at the same shape is worth acting on.

Note what “number of site pages” being near zero implies. Publishing more articles does not, by itself, make a model name you. It gives retrieval more to grab, which is exactly how you end up with a great citation count and a ghost problem.

Diagnose it in an afternoon

You cannot fix a ratio you have never measured. The audit is manual, boring, and takes about two hours.

1. Write 20 to 30 prompts a real buyer would type, weighted toward the awareness end: “best tools for X”, “alternatives to Y”, “how do startups handle Z”. Include a few where you would expect to win.
2. Run every prompt through ChatGPT, Perplexity, Gemini, and Google AI Mode. Use a logged-out or temporary session so personalization does not flatter you.
3. For each answer, record two columns: named (your brand appears in the prose) and cited (your domain appears in the sources).
4. Count the rows where cited is yes and named is no. That is your ghost citation set.
5. In those same rows, write down who was named. That list is your real competitive set in AI search, and it is frequently not the one in your pitch deck.

Two numbers come out of this: your mention rate and your citation rate. Track them monthly. A rising citation rate with a flat mention rate is the signature of the problem, and it is invisible in every rankings tool you already pay for.

The fixes, in order of leverage

Get mentioned in the places that get retrieved. Listicles, category roundups, comparison posts, and community threads are where models learn which brands exist in a category. That means real distribution work, the kind covered in content distribution for founders: showing up in other people’s content, not only your own. An unlinked mention of your name in a roundup can be worth more here than a dofollow link nobody reads.

Fix your review profiles. This one has the cleanest before-and-after evidence. Seer, in partnership with Trustpilot, analyzed 804,491 AI responses across 1,926 brands, 15,783 prompts, and four AI platforms in March 2026. Brands with no active review profile had a median citation rate of 1%. Brands with a minimal profile, as few as 1 to 13 reviews, hit 53.5%. That is a step change bought with a week of asking customers.

Two honest caveats. Trustpilot co-authored the study, so read the exact figures as vendor-adjacent. And the authors themselves note that 99.5% of Trustpilot citations arrive through organic search ranking rather than direct platform preference: the review site ranks, the model reads the ranking page, your name is on it. The mechanism is ordinary SEO working on someone else’s domain. It still works, and it explains why the effect appears at all volumes.

Make your entity unambiguous. Models and Knowledge Graphs both need to resolve “which company is this” before they can recall you. Google’s Organization structured data exists partly for disambiguation, and sameAs is how you connect your site to the profiles that represent the same entity elsewhere. Keep the name, the description, and the social profiles identical everywhere. If your brand name collides with a common word, a drug, a city, or a bigger company, this stops being housekeeping and becomes the whole job.

Do not expect the markup itself to buy citations. As covered in schema markup that gets you into AI answers, the measured uplift from adding JSON-LD to AI citations is roughly nothing. Schema is how you get identified, not how you get recommended. Those are different problems and they need different budgets.

Keep ranking. Retrieval still starts from search results. Everything above sits on top of a crawlable site that AI engines can actually read, which, as AI crawlers explained covers, is less common than teams assume once JavaScript is involved.

What not to do

Do not buy mentions. Paid placement in listicles and mass-produced guest posts create the surface pattern without the substance, and the correlation data explicitly separated paid branded signals, which sat near the bottom, from earned ones.

Do not write another comparison page and call it a strategy. If your ghost rate is high, more on-site content is the one lever the data says barely moves.

And do not let the citation count become the KPI. It is the easiest number to grow and the least connected to revenue. Mention rate is harder, slower, and closer to the sale, the same argument that makes answer engine optimization worth doing even when the clicks disappear.

Where this fits with the rest of your marketing

Ghost citations reframe what AI search work actually is. It is not a content problem with a technical fix. It is a brand awareness problem being surfaced by a new interface, and the old answer applies: be present where your buyers and, now, their models are already reading.

That is the loop Gantra runs for startups. Its daily SEO and GEO audits, free on the entry plan, watch how your pages perform in both classic search and AI answers. The article agent writes for extraction so retrieval finds you, and publishing runs on autopilot for LinkedIn and X with one-click drafts for Reddit, Hacker News, Medium, and Dev.to, which is the off-site surface that feeds the mention side. One agent writes; the rest put it where models read. See the full AI marketing team for how the twelve agents split the work.

The takeaway

Being cited is not being recommended. The evidence from two independent 2026 studies points the same way: models decide which brands to name from what they already know, then go find pages to support the answer, and a brand’s citation rate more than quintuples once it is named in the text. Measure both numbers separately, starting with a two-hour manual prompt audit. Then spend on the side that is losing, which for most startups is presence off their own domain: review profiles, third-party roundups, communities, and video. Publishing more pages will grow your citations and leave the ghost exactly where it is. If you want agents running the audit and the distribution while you keep score, Gantra starts free.