The honest answer starts with an admission: nobody outside the companies building these systems can tell you how a link is weighted, or whether it is weighted at all. There is no equivalent of a crawl log you can open and read. What you get is the output, and the output is a paragraph of prose with a few sources hanging off the side of it.
So the question keeps getting asked in a form that cannot be answered. Are links still a ranking factor, do they still count, how much of the signal is links now.
You can still learn a lot, though, by watching rather than theorising. Run the same question through a few answer engines over a few weeks, keep a note of which pages get cited, and go and look at those pages. That exercise tells you more than any confident claim about mechanics ever will.
What the cited pages tend to have in common
Start collecting cited URLs and a shape emerges fairly quickly. Not a rule though, more a tendency with plenty of exceptions.
The pages that keep turning up as sources are usually pages that other people have already pointed at. Sometimes that is a link. Often it is something looser: the page got discussed in a forum, quoted in a newsletter, pasted into a Slack channel that eventually became a public thread. The page had circulation.
And pages with no circulation at all rarely appear, even when they are better written than the ones that do. You notice this most painfully with your own content. The piece you are proudest of, the one nobody linked to, sits there unread by anything. Meanwhile a throwaway post from three years ago, the one that got argued about in a thread, keeps surfacing as a source for questions it barely addresses.
Whether that is a link effect or a discovery effect is exactly the part you cannot separate from the outside. A page that nothing points at is also a page that is harder to find, harder to crawl, slower to enter any index, and less likely to be sitting in whatever collection a system reaches for first. Those two explanations predict the same observation, which is why the debate never resolves and why the debate does not much matter to the work in front of you. The question of whether backlinks matter for AI may be unanswerable in the strict sense. The practical correlation stays stubbornly visible either way.
The sentence around the link is doing work the link is not
The shift that reframes placement entirely is not about the link at all.
In the old model, a link was a pointer with some weight attached, and the page it came from supplied context. The words immediately around it were a minor consideration. In a world where text gets read, summarised and reused, those words are not minor at all. They are a description of your page, written by somebody else, sitting on a domain you do not control.
Compare two mentions of the same resource:
- “There is a useful breakdown of this here.”
- “A breakdown of how answer engines pick between two near-identical sources, written by someone who ran the tests, sets the sequence out clearly.”
Both carry the same link. Only one of them says what the destination is about, who it is for, and why it might be worth reaching for. If a system is going to summarise the referring page, the second version travels. The first version is a dead end dressed as a citation.
You can test this on your own inbound links. Read the host sentences without the link, as plain prose. How many of them describe what you do? A great deal of link-building output, read that way, describes nothing at all.
Anchor text stopped being a keyword slot
The old anchor-text game was about exact match. You wanted the phrase you were chasing, sitting inside the link, repeated across enough domains to be noticed, and not so uniform that it looked bought.
That framing sits awkwardly now. An anchor reads as a label for a destination, and a label made of a stuffed keyword phrase is a worse label than a natural description of the same thing. The useful question is no longer whether the anchor matches a target term. It is whether someone reading the anchor alone, with no surrounding page, could say what they would find on the other side.
Which means variation is not a risk-management tactic any more. Different anchors describe different facets of the same destination, and a page described five ways by five sources is more legible than a page described identically five times. Whatever the first dozen placements agree to say tends to become the pattern for everything after, so the wording is worth settling before the campaign rather than auditing once it has run.
Small aside, but it matters: this also kills the argument for anchors like “click here”. Not because anything penalises them. Because they hand over a slot that could have carried meaning and fill it with nothing.
Links as routes rather than votes
There is a second function of a link that never depended on weighting, and it has quietly become the more defensible one.
A link is a path. Something following paths across the web reaches your page because a path existed. No path, no arrival, and the quality of your writing is irrelevant to a system that never got there. That has always been true of crawlers and it remains true of whatever is fetching pages on behalf of an answer engine.
This is why the linking patterns that look unfashionable keep earning their place:
- A mention on a site that gets crawled constantly, even if the link itself looks low value
- A listing in a directory or index that other sources routinely consult
- An internal link from a page that already gets cited, pointing at the page that does not
- A link from a page that is itself linked from somewhere well established
That third one is free and almost universally neglected. You have pages that already surface in answers. They are the closest thing you have to a front door, and most sites do not use them to point anywhere in particular.
Worth sitting with for a second, because it inverts the usual order of operations. The instinct is to build links to the page you want to perform. The cheaper move is to route attention from the pages that already perform, using assets you own outright, on a timescale of an afternoon rather than a quarter.
Where the old habits break
Some of what worked before now actively wastes budget, and it is worth naming which parts.
Volume for its own sake reads differently when the referring sentence is part of the payload. Two hundred placements with generic host sentences give you two hundred pieces of text that say nothing about you. The same effort spread across twenty placements, each with a sentence that describes what your resource covers, leaves a much clearer trail.
Chasing authority scores in isolation has a similar problem, because a high-scoring domain whose page nobody reads and nothing quotes contributes less than a mid-sized industry blog whose posts get pulled into threads and newsletters. Circulation and score are different things, and the tooling reports whichever one is easier to measure.
And the unlinked mention deserves more respect than the link-building trade has ever given it. A brand named in a sentence with a reason attached, on a page that gets read, does something. Not a link. Still something you would rather have than a link on a page nobody visits.
You see the same logic playing out in the outreach conversations. Publishers who once haggled over whether a link would be followed now field requests about how the surrounding paragraph should be phrased, which is a stranger negotiation to have and a more useful one. The follow attribute is a binary somebody else controls. The sentence is editorial, and editorial is where the description of your work actually gets set.
How to budget for this without pretending to know the answer
The reasonable position is to stop asking whether links still count and start asking what a given link would buy you.
Sometimes the answer is a path into a corpus you are otherwise absent from. Sometimes it is a well written sentence describing your page on a domain that gets consulted. Sometimes it is neither, and you are buying a number that goes up in a report.
Those are three very different purchases and the invoice looks identical.
Which is where the wording of a placement stops being a detail and becomes the thing you are actually buying. Settling the anchor and the host sentence up front, against something like an anchor-text guide for AI search, turns a placement into a description of your work rather than a line in a report. Publishers will usually agree to it, because a sentence that says what the destination covers is better editorial than a vague gesture at a resource.
The argument worth abandoning is the one about weighting. You cannot see it, the people who can are not going to publish it, and the argument produces no decisions. What you can see is which of your pages get cited, which of your inbound links have host sentences worth reading, and which of your best work is sitting in a corner with nothing pointing at it. Go and fix the third one. That part has never required knowing the formula.
