TL;DR
The GEO-BENCH study (KDD 2024) found adding direct, verifiable expert quotes improves a page's extraction into AI-generated answers by 30-40%, alongside statistics. This mechanism is identical to how fabricated quotes can produce a short-term citation bump, since the engine cannot distinguish sourced from invented ones at ingestion. Google's own quality raters consider trust the most important E-E-A-T signal, and an unnamed "industry expert" quote fails that test.
Only sourcing methods where a reader can independently verify the individual—original interview, expert-network platform with linked bio, or a publicly recorded statement with a link—carry low fabrication risk. The article's verdict is that quoting a credible, checkable expert works, while any quote-shaped text without an accountable source is indistinguishable from fabrication and will not survive inevitable scrutiny.
What "expert quotes for AI search" means
Direct answer: That's the whole definition. Everything below is about how to do it correctly, why it moves AI citation behavior, and where the line into fabrication sits — because the difference between "expert quote" and "quote-shaped text a model generated to look credible" is not stylistic. It's whether the sourcing chain can be checked by someone who isn't you.
The research behind the claim
GEO: Generative Engine Optimization (Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, and Deshpande — Princeton, IIT Delhi, Georgia Tech, and the Allen Institute for AI) was posted to arXiv in November 2023 and later published at KDD 2024. The authors built GEO-BENCH, a set of roughly 10,000 queries paired with the actual source documents a generative engine would draw on, then tested nine content-modification strategies against a Bing Chat-style generative engine (later validated against Perplexity), scoring each on two metrics: Position-Adjusted Word Count (how much of the generated answer traces back to a source, weighted by where it appears) and Subjective Impression (human-rated relevance and influence).
Adding direct quotations from credible sources was one of the two highest-scoring interventions the paper tested, alongside adding statistics — a secondary breakdown of the paper's results by DerivateX puts Quotation Addition in the same 30-40% improvement band as citing sources outright, while pure fluency polishing and keyword stuffing produced little to no measurable lift. The mechanism the authors propose is not mysterious: a quote signals that a claim is grounded in an identifiable outside authority rather than resting on the page author's say-so, which is exactly the kind of corroboration a generative engine's summarization step is built to reward when selecting what to include in an answer.
Two caveats worth stating plainly, because they don't survive most summaries of this paper: the benchmark simulates generative engines rather than measuring live ChatGPT, Google AI Overviews, or Perplexity traffic at scale, and the lift is in extraction/visibility within a generated answer — not search rank, not click-through, not conversion. It's real evidence for a real mechanism, not a guarantee about any specific engine's behavior in production.
Why this also aligns with Google's own quality framework
Direct answer: Independent of the GEO paper, Google's guidance on content quality has been pointing the same direction for years. Google's Creating Helpful, Reliable, People-First Content documentation asks creators to self-assess with questions like whether a page "provides original information, reporting, research, or analysis" and whether it demonstrates the qualities that fall under E-E-A-T — Experience, Expertise, Authoritativeness, and Trustworthiness. A named, verifiable expert quote is one of the more concrete ways to demonstrate two of those four letters at once: it borrows Expertise from someone with a track record, and it builds Trustworthiness by making the claim checkable rather than asserted.
Google's Search Quality Rater Guidelines — the manual given to the human raters who benchmark Google's ranking systems — state that "Trust is the most important member of the E-E-A-T family because untrustworthy pages have low E-E-A-T no matter how Experienced, Expert, or Authoritative they may seem." A quote attributed to "industry experts" with no name, or to a name with no way to confirm the person exists, fails that trust test even if the sentence reads persuasively. Google's 2023 guidance on AI-generated content makes the same point from a different angle: the company has no blanket objection to AI involvement in content production, but expects the output to be "backed and evaluated by someone with that expertise" — which rules out quote-shaped filler with no accountable source behind it.
Quote sourcing methods, ranked by credibility
| Sourcing method | How it works | Verifiable by a reader? | Fabrication risk |
|---|---|---|---|
| Original interview or direct correspondence | You contact a named expert and get a fresh, on-record statement | High — expert can confirm, often has their own public profile | Very low if attributed accurately |
| Expert-network platforms | You request quotes through a source-matching service (successor platforms to the old HARO/Connectively model) | Medium-high — profiles usually link to LinkedIn or a professional bio | Low, but verify the person actually responded to your specific query |
| Re-published public statement | You quote something the expert already said publicly (a talk, an interview elsewhere, a published paper) with a link to the original | High if linked to the original; unverifiable if not | Low if sourced and dated correctly |
| Pulled from another article, uncredited to that article | You lift a quote from someone else's piece without linking back to where it first appeared | Low — reader can't trace provenance, and it can look plagiarized | Medium — misattribution risk even if the quote itself is real |
| Anonymous or unnamed "expert" | "Industry analysts say..." with no name attached | None | High — indistinguishable from invention |
| Model-generated quote attributed to a real or invented person | Text written to sound like something an expert would say, then wrapped in quotation marks and a name | None until someone checks and finds no record of it | Fabrication — not a gray area |
The GEO paper's lift applies to the top rows of this table. A generative engine extracting "grounded in outside authority" as a signal has no way to distinguish a well-sourced quote from a fabricated one at ingestion time — which means the credibility payoff and the ethical failure mode are the same mechanism working in opposite directions. Fabricated quotes may produce a short-term citation bump before anyone checks; they don't survive scrutiny, and the scrutiny is getting easier to apply, not harder.
How to add real expert quotes, step by step
- Identify the specific claim that needs outside validation. Not every sentence needs a quote — target claims that are contestable, numeric, or where "why should I believe this" is a fair reader reaction.
- Find a real, identifiable person whose expertise actually matches the claim. A quote from a nutritionist about server infrastructure doesn't help credibility even if the person is real — mismatched expertise is its own form of misdirection.
- Get an actual statement, via direct interview, email exchange, a source-matching platform, or a properly linked public statement. Do not write the sentence first and find a name to attach to it after.
- Attribute completely: full name, current title, organization, and a link to something that lets a reader confirm the person — a staff bio, LinkedIn profile, prior byline, or conference speaker page.
- Preserve the wording. Quotation marks mean verbatim. If you trim for length, use an ellipsis and don't let the cut change the meaning; never smooth an expert's phrasing into something they didn't say.
- Mark it up so both readers and machines can parse it as a quotation — use an actual quote block (a
blockquoteelement with aciteattribute, in HTML terms) rather than italicized body text, and consider structured data using schema.org's Quotation type, which lets you formally associate the quoted text with the author and source. - Date the quote, particularly for anything time-sensitive — a statement given in 2023 about a fast-moving topic shouldn't read as current in 2026 without a date attached.
- Get sign-off from the expert on how they're being quoted, especially if you're pulling from a longer conversation or editing for space. This is a courtesy and a fact-check step in one.
- Re-verify before publishing, and periodically afterward — confirm the person's title and organization are still accurate, and that any linked source page still exists. A quote that was verifiable at publication and silently rots into a dead link stops doing its job.
What this doesn't guarantee
Direct answer: It doesn't guarantee inclusion in any particular AI engine's answer. The GEO paper's effect sizes come from a benchmark built to simulate generative engines, not a live measurement of ChatGPT, Google AI Overviews, or Perplexity, and the paper itself notes effectiveness varies by domain. Treat the finding as evidence the mechanism is real, not a promised percentage lift for your specific page.
It doesn't guarantee search ranking improvement — Position-Adjusted Word Count and Subjective Impression measure presence and prominence inside a generated answer, which is a different thing from where a page ranks in a traditional results list.
It doesn't substitute for the underlying claim being accurate. A real, verifiable quote attached to a false or misleading claim is still misleading — attribution isn't a laundering step for weak content.
And it does not, under any circumstance, extend to inventing a quote, inventing an expert, or inventing credentials for a real person. This isn't a gray area with degrees of acceptability. The 2023 Sports Illustrated case is the clearest cautionary example: the publisher's parent company attached fabricated author biographies and AI-generated headshots to bylines, the fallout cost multiple executives their jobs, and the publication lost its license to the Sports Illustrated name entirely within months. Fabricated attribution is a trust failure that tends to surface, and the cost of getting caught is categorically worse than the cost of skipping the tactic.
One more practical wrinkle: the platform ecosystem for finding real experts has been in flux. HARO, the long-running source-matching service journalists and content teams used to find quotable experts, shut down in December 2024 after its acquisition by Cision, as reported by Search Engine Roundtable; newer source-matching platforms have since filled that gap, with varying degrees of verification built in. Whichever platform you use, the verification step described above — confirming the person, title, and organization independently — stays your responsibility, not the platform's.
FAQ
Direct answer: No. It fails the verifiability test that makes the tactic work in the first place, and it reads as filler to both human readers and Google's rater guidelines, which explicitly reward trustworthiness over the mere appearance of authority.
Can I quote something an expert said publicly somewhere else, without interviewing them myself?
Yes, provided you link to where the statement originally appeared and attribute it accurately — that's the "re-published public statement" row in the table above, and it's verifiable as long as the link is real and specific, not a link to a homepage.
How many expert quotes does a page need?
There's no fixed count in the research. Target claims that genuinely need outside validation; a single well-sourced quote on the page's central claim outperforms several thin, loosely-relevant ones.
Does this work the same way across ChatGPT, Google AI Overviews, and Perplexity?
The GEO paper's benchmark was built around a Bing Chat-style engine and validated against Perplexity; it doesn't claim uniform behavior across every generative engine, and the paper itself found effect sizes vary by topic domain. Treat it as evidence the mechanism generalizes, not a promise it behaves identically everywhere.
What if the expert I quoted changes jobs after I publish?
Update the attribution line rather than leaving a stale title, and note the quote's original date if the context has changed. This is part of why step 9 (periodic re-verification) matters — attribution that was accurate at publication can become misleading without any edit to the quote itself.
Is quoting an anonymous survey respondent acceptable?
It's weaker than a named expert but not the same as fabrication, provided you're transparent about the anonymity and the survey's methodology is real and disclosed. What's not acceptable is presenting an anonymized or composite quote as if it came from a specific, named individual.
