Mentions, links and AI search

By Bill Hartzer · Hartzer Consulting · 25+ years in search, domains and expert witness work

The question behind the question

When people ask whether links still matter for AI search, they are usually asking something more specific: has the unit of evidence changed? For twenty-five years the unit was the followed hyperlink — a machine-readable pointer with an attribute that determined whether it counted. The question now is whether a plain mention of your brand, with no link at all, does some of the same work.

My answer, with the uncertainty stated up front: probably yes, in some contexts, to a degree nobody outside these companies can quantify. That is an unsatisfying answer and it is the only defensible one. This page is written to be clear about which parts are established, which are reasonable inference, and which are speculation dressed as insight elsewhere on the internet.

What is reasonably established:

• Retrieval-based answer systems select documents and generate answers grounded in them. This is publicly documented at a general level by the companies building them.
• Those documents are ordinary web pages, drawn from indexes built by crawling.
• Named entities — people, companies, products, places — are represented in knowledge structures that predate generative answers by many years.

What is not established: how any specific product weights any specific signal, whether mentions feed classical ranking, and how citation targets are chosen from a set of retrieved documents. I do not know these things, and neither does anyone writing confidently about them.

Brand mentions and co-occurrence

A mention is your name appearing in text without a link. The reason mentions have become interesting is mechanical rather than mystical: language models and retrieval systems work on text. A sentence that says your company builds a particular kind of product is legible as a statement about your company whether or not the words carry a hyperlink.

The concept worth understanding is co-occurrence: how often your name appears alongside a topic, a category, or a set of competitors, across independent sources. Being named repeatedly in the same context as a subject builds an association that a single link cannot. That association is what a retrieval system has to work with when a question about that subject arrives.

This is not a new idea. Co-citation — two sources being referenced together — has been discussed in search for a long time, and Module 3 covers reading it in the link graph. What has changed is that the text around the citation is now processed far more richly than it once was.

The practical consequence: coverage that names you counts for more than it used to, even when it does not link. A trade publication article that describes what you do, a roundup that lists you among alternatives, a forum thread where people discuss your product — these produce text associating your name with a subject. That was previously worth almost nothing to search; it is now plausibly worth something.

Where I would stop short: I would not claim that mentions are a ranking factor in classical search. That is a common assertion and it is not established. What I would claim is that the value of coverage is no longer contingent on whether it linked, which is a narrower and better-supported statement.

Entities, not strings

The most actionable part of this whole subject has nothing to do with AI and everything to do with being unambiguously identifiable.

A string is a sequence of characters. An entity is a thing in the world — a specific company, a specific person — that a system can distinguish from other things with similar names. The move from matching strings to resolving entities has been under way in search for well over a decade, and it is the mechanism by which a mention of your name gets attached to you rather than to a similarly named business three states away.

The work here is unglamorous and largely within your control:

• Name yourself consistently. One canonical form of your company name, one form of your product names, used the same way everywhere. Variation splits the evidence.
• Structured data. Organization and Person markup, with sameAs pointing to the profiles that corroborate identity.
• Corroborating profiles. Consistent details across the places that describe organizations — professional profiles, industry bodies, official registries, reference sources that maintain entity records.
• A clear about page. Who you are, what you do, who runs it, where. Written for a reader, parseable by a machine.
• Author identity. Named authors with consistent bios, connected to their work elsewhere.

None of this is speculative. Entity resolution is a documented, long-standing part of how search works, and getting it right helps whether or not a single claim about generative answers turns out to be true. That is why I put it first when advising anyone on this topic: it is the part with a floor under it.

Citation in generative answers

Generative answer systems commonly show sources. The natural question — how do I become one of them — deserves a careful answer, because this is where the industry's speculation is thickest.

What is publicly documented, at a general level: these systems retrieve candidate documents and generate an answer grounded in them, with citations pointing at documents used. The retrieval step draws on a web index. That much the companies have described.

What follows reasonably from that:

• You cannot be cited if you are not in the index. Crawlability and indexation remain prerequisites.
• A page that answers a specific question directly and unambiguously is easier to ground an answer in than one that buries the answer in narrative.
• Clear attribution — who wrote this, what the source of a claim is, when it was updated — makes a document more usable as a citation.
• Being one of several independent sources saying the same thing is a stronger position than being the only one.

What is speculation, and should be labeled as such whenever you read it: any claim about how a specific product ranks candidates, how much weight it gives to any signal, whether it prefers particular formats, or that some technique reliably produces citations. I have seen a great deal of confident writing on this. None of it is verifiable from the outside, the systems change frequently, and results vary between users and sessions in ways that defeat casual testing.

My honest position: the durable work is to be indexable, be clear, be corroborated, and be identifiable as an entity. Everything past that is guesswork, and I would rather say so than sell certainty I do not have.

Why the dofollow question matters less

For years the first question asked about any coverage was whether the link was followed. If it carried rel=nofollow, many people counted the placement as worthless. That reflex made sense in a world where the only value of coverage was the equity the link passed.

Several things have eroded it.

• Nofollow became a hint rather than a directive. Google announced years ago that it would treat the attribute as a signal it may choose to use, rather than an absolute instruction, and introduced sponsored and ugc alongside it. What that means in any specific case is not knowable from outside — but the binary certainty is gone.
• Major publishers apply nofollow by policy. Some of the most valuable coverage available is nofollowed as a matter of house style. Discounting it to zero means discounting the best placements you will ever get.
• Syndication multiplies reach regardless of attribute. A national story republished across dozens of outlets produces name recognition, referral traffic and text that names you, whatever the markup says.
• Mentions do work now. If the text around the link carries value, the attribute on the link is no longer the only thing determining whether the coverage counted.

The reframing I would suggest: stop asking whether a placement was followed and start asking whether it was real — did a person choose to reference you, in front of an audience, in a relevant context. That question has a stable answer and it predicts value better than the attribute does. Module 9 covers this in depth, along with the measurement problem it creates.

What to actually do, and what to measure

The work that follows from all of this is less exotic than the discussion around it.

1. Fix entity consistency first. Naming, structured data, sameAs, author identity, corroborating profiles. Highest certainty, lowest cost.
2. Pursue coverage without the link filter. Judge a placement by audience, relevance and whether a human chose to publish it — not by the attribute.
3. Answer questions directly on pages. Not for a trick, but because a page that states an answer plainly is more usable as a source by every system that reads it, including humans.
4. Publish things worth citing. Original data remains the most reliable way to become the source rather than a summary of one.
5. Track mentions as well as links. Most link programs monitor only links, which now means monitoring a subset of your coverage.

On measurement, be modest. Tools that track brand presence in generative answers are immature. Answers vary by user, by session and by model version; sample sizes are small; and there is no equivalent of Search Console reporting what any of these systems did with your site. Building a rough monitoring practice is sensible. Reporting it as a precise metric is not, and clients who are given precise numbers here will reasonably expect them to be reproducible.

The one thing I would not do is redirect a working link building program into speculative work on the strength of a trend nobody can measure. Earned coverage serves both purposes. That is the fortunate part of this whole shift: the durable answer to the old question and the new one is the same — be worth referencing.