Expert quotes are the strongest content lever, and almost nobody uses them
Updated:
Quick answer: adding attributed expert quotes to a page raises AI citation rates by 28% to 41%, which makes it the single strongest content change measured in the GEO-bench study. It also happens to be the change most businesses skip, because it requires finding a real person willing to say something specific on the record.
Why does attribution change how a model treats a passage?
A language model assembling an answer faces a problem it cannot solve directly: it has no way to check whether any given claim is true. What it can do is estimate how defensible a claim is.
An unattributed assertion sits alone. If it turns out to be wrong, the model has nothing to point at. An attributed one comes with a chain: the claim, the person who made it, their role, their organisation. When the model repeats it, the attribution travels with it, and the risk shifts from the model to the source.
This is why quotes reduce hallucination pressure. The model is not inventing a position, it is relaying one. Relaying is a safer operation, and systems built to avoid confident errors prefer safer operations.
There is a structural reason too. A quote block is naturally self-contained. It has a speaker, a claim and a boundary. Retrieval systems chunk pages into fragments, and a <blockquote> with a name attached survives that process better than almost any other element on the page.
- Attribution
- A statement of who said the line: name, role and place of work next to the quote itself.
- Expert quote
- A first-person passage attributed to a named person, not to a company or to "experts".
What does the measurement cover?
The finding comes from GEO-bench, the Princeton benchmark that tested individual content modifications against AI visibility.
Quotation addition: 28% to 41% increase in citations over baseline. In the same study, adding statistics measured at 33% to 41%, and citing external sources at 28% to 40%. Quotes sit at the top of that range.
What the study tested was specific: direct speech from a named expert, marked up so the attribution is machine-readable, with a role and an affiliation attached.
What it did not test, and what will not work: anonymous authority. "Experts agree that" and "industry leaders recommend" are not quotes. They are claims wearing a costume, and they carry no attribution chain at all.
Which technical shape gets detected?
Our auditing engine looks for a <blockquote> element that contains either a cite attribute or a name within two hundred characters of the quote. That is a deliberately narrow definition and it excludes a lot of what people call quotes.
The full shape worth aiming for:
<blockquote cite="https://example.com/author-page">
<p>Specific claim with a number or a mechanism in it.</p>
<footer><cite>Name Surname</cite>, role, organisation</footer>
</blockquote>Three things make this work beyond the markup. The speaker is named. The role is stated, which tells a reader why this person's opinion counts. And the claim itself is specific rather than decorative.
The last part is where most quotes fail. "Quality is very important in our industry" attributed to a named CTO is still worthless, because the sentence carries no information. A quote earns its place when removing it would remove a fact.
Where do you find a quotable expert as a small business?
The obvious objection: we have no industry-famous names to quote.
You do not need one. The research measured attribution, not celebrity. What matters is that a real person with a stated role said something specific.
Your own people are experts. The person who has done the work for eleven years is an expert on that work. Their observation about what clients get wrong most often is exactly the kind of specific, experience-backed claim the mechanism rewards.
The quote should carry the thing only they know. Not the company position, which you can state in your own voice. The detail that comes from having done it: what breaks most often, what the number usually is, what surprises clients.
Length is not the point. Two sentences with a number in them beat a paragraph of reflection.
Get it on the record properly. The person's name and role need to be real and checkable. An invented expert is worse than no quote, and not only ethically: it creates a claim that falls apart the first time anyone looks.
Why does almost nobody do this?
The measurement has been public for a while. The technique is not secret. It costs nothing beyond a conversation. And still, most business pages contain zero attributed quotes.
The reason is friction, not ignorance. Adding a real quote means finding a person, asking them to commit to a specific statement, and then not editing that statement into corporate mush. Every step of that is slower than writing another paragraph in your own voice.
That friction is exactly why the signal works. A page with three named experts saying specific things is expensive to fake at scale. Content farms do not do it, because doing it requires having people.
When I ask a client for a line from their own specialist, I almost always hear the same thing: hold on, let me check with marketing. That is where the quote dies. A live sentence goes through three approvals and comes back as something anyone could have signed.
What is the limit worth knowing?
If quotes are this effective, why not add five?
Because the same research base contains a countervailing finding. Piling trust triggers onto a page drops a challenger brand's breakthrough rate from 55% to 21.2% in Claude. The model treats an overloaded page as a manipulation attempt.
Our engine encodes this directly. Trust signals are scored with saturation: the first three carry the weight, and anything beyond that adds nothing. We also flag pages that combine five or more persuasion markers with thin body text and no original images.
The practical reading: one or two genuine quotes on a substantial page is the shape that works. Five quotes on a nine-hundred-word page is the shape that gets penalised.
What did we find on our own pages?
Our engine flags cite.expert-quote as failed on the home page of our own site, and puts it in the top of the fix list at four points, tied with the answer-first check for the largest single gain available on that page.
That is a fair result. The page describes what we do in our own corporate voice throughout. There is not a single sentence where a named person says something you could disagree with.
Writing that sentence turns out to be the hard part, and not for technical reasons. A quote that carries information exposes a position. "Most sites we audit block the crawlers they most need" is a claim someone can push back on. The safe version, "crawler configuration is an important consideration", says nothing and can be attributed to anyone, which is exactly why it counts for nothing.
The friction is the feature. A page carrying two specific, attributed claims is expensive to produce and hard to fake in bulk, and that is why the signal survives as a signal.
How do you run the conversation that produces a quote?
The blocking step is almost never the markup. It is getting a colleague to say something specific enough to be worth quoting, and there is a way to run that conversation that works better than asking for a quote.
Do not ask for a quote. "Can you give me a quote for the website" produces the corporate sentence every time, because the person switches into the register they think a website requires. You get "we are committed to quality" and no amount of editing rescues it.
Ask what goes wrong. "What do clients get wrong most often?" and "What surprises people about this?" produce specific answers, because they are questions about experience rather than requests for a statement. The person answers as themselves.
Ask for the number. Whatever they say, follow up with how often, how long, how many. "Most of the time" becomes "maybe seven out of ten" and that is the sentence worth printing.
Write it down verbatim, then read it back. Not a polished version. What they actually said, lightly cleaned of filler. People approve their own words readily and reject rewrites of them, which is convenient because their own words are what you wanted.
Attach the role, not just the name. "Anna, eleven years doing technical audits" tells a reader why this claim carries weight. A bare name does not.
The whole conversation takes about ten minutes and typically produces three usable quotes, not one.
Frequently asked questions
Does a customer testimonial count as an expert quote? Not for this mechanism. A testimonial says the service was good, which is an opinion about you. An expert quote states a fact about the domain. Different signals, both useful, only one measured here.
Can I quote myself? Yes, if you state your role and say something specific. A founder explaining what fails most often in their field is a legitimate expert source.
What if the expert wants to approve the wording? Normal and fine. What matters is that the final text is specific and that they stand behind it. Approval rounds tend to remove specificity, so push back on edits that turn a number into an adjective.
Does the quote need a link? It helps. A cite attribute pointing at the person's profile or the source of the statement strengthens the attribution chain. It is not required for the markup to be detected.
Do quotes work on service pages, or only in articles? They work anywhere they are honest. A service page where the person delivering the service explains the most common mistake is more convincing than a feature list.
How many quotes across a whole site? No useful number exists. Per page, one or two that carry real information. Across a site, as many as you have people with something specific to say.
The short version
The strongest measured content change in AI visibility is putting a named human being on the record saying something specific. Not a slogan, not a company position, not "experts agree". A person with a role, making a claim you could argue with.
It is slow, it requires talking to people, and it cannot be automated. That is not a drawback of the technique. That is why it still works.
Sources: GEO-bench measurements (Princeton University); research on brand bias in LLM recommendation systems; project research digest compiled from 165 sources on AI search, August 2026. Details in sources/research-notes.md.