Do backlinks get you cited by AI answers?
Only indirectly, and less than most people building links assume. Large language models have no link graph inside them, so there is no PageRank-style calculation happening when an answer is assembled. Links matter at the retrieval step, because the search index an assistant pulls live results from still ranks pages partly on links. But the strongest measured predictor of being cited is how often a brand is mentioned across authoritative sources — roughly three times stronger a correlation than backlinks.
What the studies actually measured
Three findings from 2026 are worth holding onto, because they point the same way and they are not what most link building is currently aimed at.
The first is that brand mention frequency across authoritative sources correlates with AI citation at around three times the strength of backlinks. Brand search volume turns out to be the single strongest known predictor, at a 0.334 correlation — materially ahead of links.
The second is where citations physically come from. Omniscient Digital looked at 23,387 citations across the five major engines and found that 57% of branded citations came from reviews and social proof — review sites, listicles, press — with 17% from directories. Only about 4.5% came from the brand's own About, FAQ or home pages.
The third is concentration. Wikipedia accounts for close to half of ChatGPT's top-ten source share. Reddit accounts for roughly 46.7% of Perplexity's. Two sites you cannot buy your way into are doing an enormous share of the work.
| Source type | Share of branded citations |
|---|---|
| Reviews and social proof — review sites, listicles, press | 57% |
| Directories | 17% |
| The brand's own pages — About, FAQ, home | 4.5% |
Why links still matter, just not the way you think
It would be wrong to read the above as "links are finished". They are not, but the mechanism has changed and the change matters.
When an assistant answers a question about something current, it usually runs a search and reads the results. That retrieval step uses an index that still ranks pages partly on links. So a link that improves a page's ranking makes that page more likely to be retrieved, and therefore more likely to be summarised into an answer.
What has gone is any direct effect. There is no link authority stored in the model. A page with a thousand backlinks and a page with ten are, to the model itself, just text. The link only helps by improving the odds that the text gets fetched in the first place.
The practical consequence is that links pointed at pages nobody would retrieve for any real question do very little for AI visibility. That describes a lot of link building.
There is no link graph inside the model. Links help by getting your page retrieved, not by making the model trust it.
The part that costs us money to say
We sell placements on other people's publications, so the honest thing to do here is state where that does not help.
A placement on a site that exists to sell placements is not a brand mention on an authoritative source. It is a page with your name on it, on a domain no assistant has reason to retrieve, in a corpus no model weighted heavily. It may still move a keyword ranking. It will do close to nothing for whether ChatGPT knows who you are.
That is not an argument against buying links. It is an argument against buying them from anyone who will sell you forty thousand sites, because at that volume nobody has checked whether the publications are publications.
It also means the honest version of this advice is not "buy more links". For most agencies the higher-leverage moves are getting clients into the review sites and listicles that dominate the citation data, and being genuinely present in the places — Reddit and Wikipedia among them — that you cannot purchase at all.
Where editorial placements do fit
With that said, there is a real overlap, and it is worth being precise about where it sits.
A guest post on a publication with actual readers, an actual editor and an actual publishing schedule is a brand mention on an authoritative source. That is the category the data says predicts citation most strongly. The link is almost incidental — what is doing the work is your name appearing, in context, somewhere a model has reason to have read.
So the same purchase can be a good link-building buy and a good AI-visibility buy, or it can be neither, and the thing that decides which is whether the publication is real. That distinction used to be a quality preference. It is now the whole difference between a placement that compounds and one that does not.
What to actually do
- Stop measuring link campaigns purely on domain rating. It was always a weak proxy and it has no relationship at all to whether a model has read the site.
- Check whether a publication is retrievable before buying: does it rank for anything, does it publish on a schedule, does a real person edit it.
- Put effort into review sites, listicles and press, which produce 57% of branded citations. This is outreach and PR work, not link buying.
- Track brand search volume as a leading indicator. It is the strongest single predictor anyone has measured.
- Treat your own site's About and FAQ pages as necessary but not sufficient — they account for under 5% of citations.
Sources
- AI citations vs traditional backlinks — Searchable
- What gets cited by ChatGPT, Claude, Gemini and Perplexity — QuickSEO
- The state of AI citations 2026 — 5WPR
- Are backlinks still a ranking factor for LLMs — Pathmonk
Figures are quoted as the studies above report them. Where they disagree, the text gives the range.
Questions people ask
Publications, not placements
Every site on our marketplace was assessed as a publication first: real readers, a real editor, a real publishing schedule. That is the distinction this page is about, and you can check it yourself before you buy.
Browse the marketplace