57% of AI Citations Go to Third-Party Sources

Most of what an AI engine cites when it answers a buying question sits somewhere other than your website. Profound's citation research found 57% of AI citations go to third-party sources. That one number reorders the work.

If more than half the evidence sits on pages you cannot edit, the first question is which off-site sources carry your category, what they currently say about you, and which of them you can honestly influence.

Honest is the operative word, and it works as a constraint on everything below. An accurate profile is earnable, so is a review program that asks every customer, and so is a factual correction. Everything adjacent to that is manipulation.

The stakes come from the buyer. G2 found AI chatbots are now the single largest influence on B2B shortlists, 51% of software buyers begin research inside an AI chatbot, and one-third bought from a vendor they had never heard of before.

How much of an AI answer sits on sources you do not own?

The majority, and the rest is settled by which pages the engine could retrieve at all. Seer found 87% of SearchGPT citations matched Bing's top results in a 500-citation sample, so pages a conventional index cannot surface are unlikely to reach the retrieval layer at all. Ranking there gates retrieval; being cited is a separate decision the model makes on top of it. An AirOps analysis of 548,534 pages mapped which page traits correlate with being pulled into an answer.

Off-site sources win that competition because they were built for it. A review platform's category page, a best-tools roundup, and a long forum thread all answer the question the buyer typed. Your product page answers a different one. Ahrefs' analysis of 4 million AI Overview citations and Semrush's AI Visibility Index across 126 million AI search prompts both look at which domains recur across large citation samples.

So a content calendar aimed only at your own domain is working the minority of the evidence. Publishing on your own domain still earns its place, and knowing the split first is what makes the quarter's allocation defensible.

Which off-site sources are worth working on first?

Start from the sources your own buyer questions return. Four categories cover almost all of it: review platforms, roundups and listicles, community threads, and reference sources. They differ in how they are earned, how fast they move, and how long the result lasts.

Review platforms are slow to build and durable once built. Roundups shift in weeks and decay as writers refresh them. Community threads are the fastest surface and the least controllable. Reference sources move slowest and sit closest to permanent.

Effort spread evenly across four categories usually loses to the same effort concentrated where your engines already cite.

Rank them by observation. Run your buyer questions, read every source the answers cite, and count which domains recur. Research on brand dynamics in LLM recommendation systems shows prior associations are sticky and unevenly distributed across brands, so a source that already carries your category is worth more of your time than a larger one that does not.

What can you legitimately earn on a review platform like G2?

Four things, and none of them is a review. First, an accurate profile: current pricing model, current category placement, current feature list, correct description of what the company does. A stale profile is a factual error the engine repeats back to your buyer.

Second, a standing review program that asks every customer on a schedule. Third, replies to critical reviews, which are public text the engine reads alongside the complaint and the only place you answer in the source's own voice. Fourth, the structured facts the platform exposes. Google's guidance on high-quality review content sets out what makes a review page useful, and the same attributes make a profile extractable.

What you cannot do is choose which customers get asked. Gating requests by expected sentiment is the behavior the FTC's advertising guidance warns about, and it is self-defeating anyway: a profile with no critical reviews reads as unverified, both to a buyer and to a model weighing which source to trust.

How do you earn a place in a roundup or listicle?

By being comparable, then by being cheap to compare. Most roundup writers assemble from what they can verify in an afternoon: pricing, integrations, category, one differentiating claim, a screenshot. When those facts are ambiguous on your site, you cost more to include than the competitor whose page states them plainly.

Extractable evidence is the mechanism. The GEO study (Aggarwal et al., KDD 2024) tested whether adding statistics, quotations, and citations changes what gets cited, and its earlier preprint reported a 41% improvement over baseline for the best method it tested. The gains came from carrying evidence.

Be honest about what phrasing does not buy. C-SEO Bench (NeurIPS 2025) tested conversational-SEO rewrite methods across two tasks and six domains and found only 3 of 54 unilateral conditions produced statistically significant citation-rank gains. The lever is being worth listing on the merits, with facts a writer can check in minutes.

The second earnable action is correction. When a roundup includes you and describes you wrongly, a short note with a verifiable source usually gets fixed, because the writer wants the piece to be right.

How do reference sources differ from community threads?

Reference sources reward consistency and threads reward specificity, and each fails in its own direction. A reference entry earns its citation because your facts match everywhere else they appear. A thread earns one on the strength of the single answer inside it, and participation that reads as marketing gets removed by the community long before an engine sees it. Threads get their own treatment in a separate piece.

Reference sources are a different discipline. They reward correctness and consistency: one canonical company name, one description, one set of facts repeated identically wherever they appear. Entity-oriented retrieval research across 443 configurations shows how much retrieval quality depends on resolving a name to a single entity, and a brand that appears under three variants fragments into an ambiguous string.

The legitimate action on a reference source is narrow and worth doing anyway: correct factual errors, cite a verifiable primary source for the correction, and leave the framing alone. Writing your own promotional entry is the fastest way to have every mention of you deleted.

Where does earning end and manipulation begin?

The line is whether the signal survives disclosure. A review program that asks all customers survives it. Incentivized reviews, seeded threads from accounts you control, and paid placement presented as editorial fail it. Google's spam policies cover the search-side version of the same behavior, and the retrieval layer above still runs on that index.

There is a second cost that is easy to miss. Manufactured social proof is an attack surface: the mechanism that lets you inject a favorable signal lets a rival or a bad actor inject an unfavorable one, and a brand whose visibility rests on manufactured proof has no defense when the manufacturing turns hostile. That case is made in full in social proof is an attack surface, not a GEO tactic. This piece is its constructive half.

Promotional register carries its own penalty. Across 10 fictitious products, scarcity and exclusivity framing measurably reduced how often an LLM recommended one, which is a reason to watch the promotional register in copy that third parties quote back.

How does Trovance help you work the sources you do not control?

Trovance watches the answers and records what they cite. You define the buyer questions your market actually asks, and it runs them repeatedly across AI engines, preserving each answer run with its full citation set. What accumulates is a record of which off-site domains carry your category, how often each appears, and whether the answer that named a competitor was built on that competitor's pages or on someone else's.

That record turns off-site work into a ranked list. If a review platform's category page carries nine of your last thirty answer runs, then a full quarter spent on profile accuracy and the review program is defensible.

If one roundup keeps supplying the comparison that omits you, the action depends on why it omits you. Close a factual gap and earn the place; where the comparison itself is the missing evidence, publish it. On its own, an absence is a signal to test. Your Brand Core holds the claims you are entitled to make and the proof behind each, so a correction you send to a third party arrives with its proof attached.

Comparison over time is where it pays. Every answer snapshot keeps its sources, so when a profile is corrected or a roundup is refreshed, the next analysis cycle reruns the same questions and shows whether the cited sources changed and whether the answers followed. Drafts produced from your approved claims go to a person for review before anything publishes.

What Trovance will not promise is influence over a source it does not own. Nothing can make a review platform rank you higher, make a writer include you, or make a model recommend you, and a tool claiming otherwise is describing manipulation or selling a number it cannot support. Trovance preserves the citation record, diagnoses which off-site source is carrying the answer, and names the earnable action the evidence supports. Whether the third party acts stays their decision.

What should you do this week?

Four steps, in order. Run ten buyer questions across at least two engines on more than one day and record every source cited. A single run proves nothing: AI engines are inconsistent brand recommenders, and a 2026 variance-components study found run-to-run noise large enough to swamp real differences in small samples.

Second, count the cited domains and pick the two that recur most. Third, audit yourself on those two: is the profile current, is the description accurate, is there a factual error you can correct with a source attached. Fourth, start the review program, because it is the slowest item on the list and every week of delay costs a week of evidence.

Set the timeline honestly. Off-site change is measured in months, buyers evaluate roughly five vendors and most of that list is set before contact, and the sources move on their own schedule. If you want the source-level record without the manual runs, start a free Trovance analysis and see which off-site domains are answering for your category.

Where the answer actually comes from

• Where AI citations come from across industries - how much of the answer sits off your own domain.

• Where ChatGPT gets information about your business - the source mix behind a single answer.

• What gets cited by AI - the page traits that correlate with being pulled into an answer.

• Google rankings vs AI citations - why ranking and being cited are separate outcomes.

Earn the evidence

• How to build a competitor comparison page - the asset that replaces a roundup you cannot get into.

• AI agents need evidence, not more content - what makes a claim usable to a model.

• Why ChatGPT recommends your competitor - source displacement as one of five failure modes.

• Social proof is an attack surface, not a GEO tactic - why manufactured proof is a liability.

• The most-quoted study in AI search, quoted correctly - what the GEO benchmark actually tested.

FAQs