Two claims are made about this and both are wrong. The first is that AI search makes link building obsolete. The second is that nothing has changed and links work exactly as before.
What is actually true is narrower and more useful: the mechanism by which links produce value has shifted, and that shift makes some placements more valuable than they were and others close to worthless.
What we can say with confidence
Worth stating the epistemic position plainly, because this topic attracts more assertion than evidence.
Nobody outside the labs knows exactly how answer engines weight sources. What is observable is which sources get cited in generated answers, and the pattern is consistent enough to plan around: models draw disproportionately on publications with editorial reputation, and they favour pages that state facts plainly with attribution.
That set overlaps heavily — not perfectly — with the set of publications that pass meaningful link equity. Which is why the two systems, despite working differently, reward similar behaviour.
The mechanism, as best we can describe it
In classical search, a link is a vote that flows equity to your page, and your page then competes for a position.
In a generated answer, there is no position to compete for. A model synthesises a response from sources, and your company either appears in that synthesis or does not. What determines appearance is whether your product is named, in context, in sources the model draws on.
The unit has changed from the link to the mention. A link is one way of being mentioned; it is no longer the only way that counts.
This is why unlinked brand mentions — which pass no PageRank and have historically been treated as a consolation prize — now carry real value. A paragraph in a trade publication naming your product as one of the three tools practitioners use is exactly the kind of source a model reaches for when asked to compare the category.
What became more valuable
Original research and data
The clearest winner. A statistic with a published methodology, cited across dozens of publications, makes you the origin of a fact. Models summarising the topic name the source of the number, and that source is you.
Expert commentary and quotes
Being the named specialist in an article is a mention with attribution attached — close to the ideal shape for this. And the fact that these placements are frequently no-follow now matters considerably less than it did.
Category round-ups and comparison pages
Already the highest-value SaaS placement for other reasons, and now more so. When someone asks a model to compare tools in your category, round-ups are among the most directly relevant sources available.
Definitional and glossary content
Unglamorous and newly useful. Models explaining a category term lean on definitional sources, and being one is a durable position that almost nobody is competing for.
What became less valuable
Links from sites with no readership
Already a bad trade; now bad in a second dimension. A placement on a site nobody reads produces no referral traffic, no credibility, and no chance of being a source a model draws on. Its entire value was a ranking effect that may be discounted at any time.
Anchor-text-driven acquisition
Exact-match anchors were always a small part of a healthy profile. In a mention-based system they contribute nothing at all — a model does not care what words the link used.
Volume as a strategy
Fifty placements on marginal sites was a defensible approach when equity accumulated. Against a system that reads sources rather than counting votes, fifty marginal mentions in publications no model consults is fifty nothings.
The practical reframing
Before buying a placement, ask: if this link passed zero equity, would I still want it?
If yes — because the right people read it, because it generates referrals, because being named there is worth something on its own — then buy it, and treat the equity as a bonus.
If the only argument is the authority score, you are buying an asset whose value depends entirely on a mechanism that is becoming less central.
What to measure now
Three additions to a link building report, none of which existed as standard three years ago.
| Metric | Why |
|---|---|
| Unlinked brand mentions | They count now. Track them alongside links, not instead. |
| Citation of your data by third parties | The clearest signal that you are a source rather than a subject |
| Presence in category round-ups | Directly relevant to generated comparisons |
And one to stop over-weighting: average domain rating. It was always a weak proxy. It is now a proxy for a mechanism that is one of two rather than the only one.
What has not changed
Classical search has not gone anywhere. A large share of category research still happens through ranked results, and the pages that rank still need links. Any strategy built entirely around answer engines is making the opposite mistake to the one it is trying to avoid.
Nor have the quality standards changed. Verified readership, topical relevance, editorial identity and contextual placement were the right filters before and are the right filters now — with the useful side effect that a programme built to those standards was already optimising for both mechanisms without needing to know it.
A note on honest uncertainty
Everything above is inference from observed citation behaviour, not from documentation. The systems change frequently, and anyone claiming precise knowledge of how a model weights a source is overstating what is knowable.
What makes us comfortable acting on it is that the recommended behaviour is unchanged: acquire coverage in publications with real readers and real editorial standards. That was the correct strategy when only PageRank existed. It remains correct now, and it will remain correct under whatever comes next — which is a reasonable definition of a robust strategy.