What did the MIT citation trust experiment find?
The MIT experiment by Li and Aral (2025) found that reference links and citations in AI search results significantly increased trust in those results, and that the effect held when the links were wrong or hallucinated. Aral, Li and Zuo summarise it in their 2026 paper on the global effects of AI search, where they warn that AI search "affects our trust and behavior in potentially dangerous ways" and that AI designs "can increase trust in inaccurate and hallucinated information."
The effect was not spread evenly. In the authors' words, "references increased trust in AI search results significantly more for people with lower education levels and for those who did not work in technology related industries." Put plainly, the buyers least likely to check a citation are the ones a citation persuades most.
Two commercial consequences follow. A brand named inside a citation-formatted answer receives a trust transfer, whether or not the description is correct. And that transfer is strongest among SMB owners, procurement staff outside tech and general consumers, which is most of the market for many businesses.
How accurate are AI search citations?
AI search citations are wrong often enough that citation-driven trust is a real risk. Aral, Li and Zuo pull together several independent measurements, and a separate study by Wallat et al. adds a further angle: a citation can point to a source that says the right thing without the model actually having used it.
| Study | What was measured | Finding |
|---|---|---|
| Columbia Journalism Review, Tow Center (2024) | ChatGPT Search identifying the source of quotes | "Confidently wrong" in 146 of 200 attempts (73%) |
| Columbia Journalism Review follow-up (2025) | News source citations across AI search engines | Incorrect sources 60% of the time; Grok 94% |
| 2025 audit of LLMs with web access | Whether cited sources support response statements | 50% to 90% of statements not fully supported; about 30% of claims unsupported even for a top system |
| Wallat et al. | Citation faithfulness in retrieval-augmented answers | Up to 57% of citations unfaithful |
| New York Times (2025) | Hallucination in newer reasoning models | Problem reported as getting worse, not better |
Combine these error rates with the trust effect and the risk becomes concrete. An AI system that confidently places a specialist firm in a generalist category, or attributes services it does not offer, is generating trust in that misrepresentation among the buyers most likely to act on it without checking. How this plays out for brands is covered in our post on AI hallucination.
Why does citation presence affect trust more than citation accuracy?
Citation presence affects trust more than accuracy because AI search has replaced a list of sources with a single synthesised answer. A list invites comparison. A synthesis signals that the comparison has already happened.
Aral, Li and Zuo describe classic web search as training users in "the discipline of triangulation": opening several tabs, comparing claims, checking credentials and dates. That habit was built into the mechanics of a results page. An AI answer removes the mechanics. The citations underneath look like proof that the sources were weighed, not like alternatives the reader should open.
The authors call this the moment when "classic position bias turns into presentation bias". The polished answer at the top of the page "inherits an aura of authority" that the sources below it can no longer contest. Checking an individual citation means stepping outside that framing, and most people, most of the time, do not.
This shift is also why fewer people click at all. Our analysis of zero-click search looks at the behavioural data behind that change.
What does the citation trust effect mean for brand visibility?
The citation trust effect makes every accurate AI citation of your brand worth more, and every inaccurate one more damaging.
On the opportunity side, a buyer who sees your brand cited in a synthesised answer reads it as an implicit endorsement: the system looked at sources and chose yours. Iyappan (2026) documents a 14.2% conversion rate for AI-referred traffic, substantially higher than traditional organic search. Part of that premium plausibly reflects the trust-in-citation effect.
On the risk side, imagine an assistant that calls your firm a "full-service marketing agency" when you are a specialist in AI search visibility. Buyers who trust that answer, as citation formatting encourages them to, may never shortlist you for the work you actually do, or arrive with the wrong expectations.
Entity clarity decides which side you land on. Kargaev (2026) identifies brand entity signals as the dominant GEO factor, with an NIS of 0.918. When those signals are clear and consistent across sources, AI systems have reliable material for describing your business. When they are thin or contradictory, the system fills gaps with inferences from adjacent content, and those inferences then receive the full trust boost of a cited answer. Weak entity signals plus strong citation trust is the highest-risk combination. Our guide to brand entity SEO covers how to build those signals.
How does the effect differ by buyer and platform?
The citation trust effect is strongest for less technical buyers and early in the buying journey, and weaker where buyers verify sources.
| Buyer situation | Strength of citation trust effect | What matters most |
|---|---|---|
| SMB owners, non-technical procurement, consumers | Stronger | Accurate category and service description in AI answers |
| Enterprise IT, technical practitioners, researchers | Weaker | Cited page must survive verification |
| Awareness stage (first discovering vendors) | Stronger: the AI answer is the first impression | Correct positioning from the first mention |
| Evaluation stage (comparing vendors) | Weaker: more click-through and checking | Specific, verifiable claims on the cited page |
Platform matters too. Iyappan (2026) describes Perplexity as the platform most used by professional researchers, with explicit citation display that shows which sources shaped the answer. On Perplexity, a cited source is visible and gets scrutinised, so being cited from an accurate, specific page is worth more than appearing via a weaker source that may not hold up. Our comparison of how AI search platforms rank you differently goes into each platform's citation design.
Position within the answer adds a further layer. Luther and Touboul-Cohen (2026) show that average position in AI responses is a metric independent of mention frequency: a brand mentioned fourth is less prominent than one mentioned first, even with identical citation formatting. Citation presence produces the trust transfer; a prominent position amplifies it.
How do you build content that AI systems cite accurately?
Content that AI systems cite accurately makes specific, verifiable statements about what a business does, and declares the same facts in machine-readable form.
Start with specific claims. "AI search visibility audits for Dutch SMEs, covering schema, crawler access and source citations" gives a model a bounded description to reuse. "We deliver excellent results" gives it nothing, so it fills in the specifics itself, sometimes wrongly.
Then declare your identity in structured data. Organization schema with complete knowsAbout and serviceType properties tells AI systems which category you belong to and what expertise you hold. Without it, the category is left to inference, and inference is where misdescriptions start. Our research into the AI visibility of Dutch SMEs shows how common that gap is: 52% of the sites studied had no Organization or LocalBusiness schema, and only 3 of 354 pages containing figures or claims linked to an independent source.
FAQ content with FAQPage markup gives AI systems ready-made question-and-answer pairs to extract. If the answers are specific and operationally clear, the extracted citation inherits that precision. If they are vague, so is the citation.
The research points the same way. Iyappan (2026) reports AI citation rates of 85% for content with statistics and citations, and 92% for long-form contextual content. Haddad (2026) finds that structured content completeness drives +8.7% AI-assisted inclusion, and the same completeness gives the model more precise material to describe you with. Our post on content quality and dwell time explains how operational specificity helps both human conversion and AI accuracy.
How do you monitor AI search credibility, not just frequency?
Monitoring AI search credibility means checking what AI systems say about your brand, not only how often they mention it.
- Run 20 to 30 category-relevant queries in ChatGPT, Perplexity and Google AI Overviews, and record every answer that includes your brand.
- For each answer, check four things: is the description accurate, is the category right, are any specific claims correct, and is the brand cited for its actual expertise or only generically?
- Compare those descriptions with your service pages, Organization schema and case material. Every discrepancy points to a gap: missing specificity, inconsistent entity signals or schema that misstates your positioning.
- Trace sources where you can. Perplexity shows its citations, so you can see whether the answer drew on your own site, press coverage or a directory listing. If weaker third-party descriptions keep winning, improve your own primary content first.
- Repeat the same query set at fixed intervals to see whether content and schema changes move the descriptions.
Frequency and position tracking are covered in our AI search monitoring framework; accuracy belongs next to them as a third dimension. The GEO checklist lists the structured content and entity work that drives it.
Frequently asked questions
Why do AI systems describe brands incorrectly?
AI systems describe brands incorrectly when brand signals are incomplete or inconsistent, so they fill gaps with inferences from related content. A business without clear Organization schema can be pushed into the wrong service category by keyword associations. One with inconsistent naming across the web may be described in the framing of whichever source the model drew on. The error rates in news citations, 60% to 73% in the Columbia Journalism Review studies, show how often that inference goes wrong.
Does an inaccurate AI citation harm my brand?
Yes, usually, though the damage depends on the buyer. Less technical buyers tend to trust the cited description and form a wrong impression, which can lower lead quality while citation counts look healthy. Technical buyers are more likely to spot the error, and a confident misdescription they discover can hurt credibility more than no mention at all.
Can I correct inaccurate AI citations?
Not directly, because AI assistants have no correction form comparable to search engine webmaster tools. The fix is content: clearer Organization schema, more specific service pages and, where needed, editorial mentions that describe your business in the right category. Those give the system better source material the next time it retrieves information about you.
Does structured data help AI systems cite brands accurately?
Yes, structured data is the most direct lever for citation accuracy. Organization schema declares your category and expertise in machine-readable form, and FAQPage markup supplies precise question-and-answer pairs. Haddad (2026) links structured content completeness to +8.7% AI-assisted inclusion.
How does citation trust relate to zero-click search?
Citation trust matters more because so many AI-assisted searches end without a visit. Aral, Li and Zuo cite Similarweb data showing a median zero-click rate of 80% for searches with an AI Overview, against 60% for searches without one. For buyers who never click, the AI description is the only contact with your brand before they decide, and the citation makes them less likely to question it.
How is AI search credibility different from reputation management?
AI search credibility work focuses on the machine-readable signals that shape what AI systems say about you, while traditional reputation management focuses on what people say in media, reviews and social channels. The two overlap, since press coverage feeds both. But only the AI side requires Organization schema with knowsAbout, FAQPage markup and consistent entity data across every source a model might read.
Does a brand benefit from citations if it is not mentioned first?
Yes, but less than a brand mentioned first. Any citation carries some trust transfer, while Luther and Touboul-Cohen (2026) show that average position is a separate metric from mention rate. Tracking both, as described in our AI brand visibility framework, gives a truer picture of how buyers meet your brand in AI answers.
Want to know how AI systems describe your business today? Get a free analysis of your website.
