Ask three marketers whether backlinks still carry weight and expect three confident, contradictory answers. One will tell you link building is dead, replaced entirely by brand mentions and Reddit threads. Another will insist nothing has changed at all. The honest answer sits somewhere between those two positions, and a recent domain-level study finally puts some numbers behind it.
The question has become more pressing now that so much research starts inside a chat window rather than a results page. When ChatGPT, Perplexity, Gemini and Google’s AI Overviews decide which sources to cite, do they still weigh the web of links pointing at a domain, or has that signal quietly stopped counting? For any business weighing backlinks and AI search visibility against other places to spend its SEO budget, that is not an academic question.
Why the Question Won’t Go Away
Traditional search ranked pages against each other using backlinks as one of its clearest trust signals. Generative engines work differently: instead of ranking ten blue links, they retrieve a handful of sources and weave them into a single answer, which is why understanding how AI models choose citations has become its own discipline, separate from classic SEO. That shift in mechanics is exactly why so many marketers assumed the old signals, backlinks included, no longer applied.
What a 1,000-Domain Study Actually Found
Semrush’s AI Visibility Toolkit team tested that assumption directly in a 1,000-domain backlink study, tracking a random sample of domains across ChatGPT, ChatGPT with Search, Gemini, Google AI Overviews and Perplexity, then running correlation analysis against each domain’s backlink profile. The headline finding was not that backlinks carry no weight. It was that authority and quality predict AI visibility far more reliably than raw link volume, with Authority Score against AI mentions showing by far the strongest relationship in the entire study.
The volume story looked different. Authority Score against AI share of voice produced a much weaker linear relationship, and the researchers noted that gains tended to appear only once a domain crossed into a higher authority tier. A handful of extra low-value links rarely shifted anything. A handful of links from genuinely respected, topically relevant sites could.
Two further findings stand out. Follow and nofollow links performed almost identically across every platform tested, which undercuts the old habit of discounting nofollow placements from reputable publications. Image-based backlinks also correlated more strongly with AI visibility than plain text links, particularly for Perplexity and ChatGPT Search, suggesting that shareable charts, diagrams and infographics are earning their keep in ways a paragraph of guest-post text no longer does.
Where the Old Playbook Breaks Down
None of this means link building for AI search looks like link building in 2018. Search Engine Land’s guidance on AI link building points out that AI platforms lean heavily on signals that never touch a hyperlink at all: forum discussions, podcast mentions, review sites and social chatter. Reddit alone appears in roughly 3% of AI citations, and more than four in five of those answer exactly the kind of decision-seeking question, such as comparing two providers, that used to live entirely in branded search results. A domain can be genuinely well-regarded without a single backlink reflecting that reputation.
That is the real shift. The old mental model treated link acquisition as the goal in itself. A more useful model treats links as one visible trace of a broader reputation that AI systems are trying to read, alongside reviews, mentions in other people’s content and consistency about who a brand is and what it does.
Practical Ways to Build Signals That Actually Count
None of the research above argues for abandoning link building. It argues for being more deliberate about where the effort goes.
- Favour fewer, more relevant placements on genuinely authoritative sites over a high volume of low-quality links
- Diversify referring domains instead of repeating the same handful of guest-post outlets
- Treat nofollow mentions from reputable publications as worth pursuing, not automatically discounted
- Produce original, data-rich content, such as surveys or benchmark studies, that earns citations because it says something new
- Create shareable visual assets, like charts and diagrams, that are more likely to be linked and embedded as images
- Track citation share by platform alongside backlink metrics, since the two do not always move together
Platform Differences Worth Knowing
The correlations above were not uniform across platforms, and that detail shapes where a team should focus its energy. Standard ChatGPT showed the strongest relationship between link quality and visibility of any platform tested, while ChatGPT Search and Perplexity weighted it comparatively lightly, leaning more on freshness and direct relevance to the query at hand.
Google AI Overviews and Perplexity showed a slight lean toward follow links, while Gemini and ChatGPT leaned marginally toward nofollow, though the gap in every case was small enough that chasing one link type over another would be a poor use of a limited budget. The clearer pattern sits with format. Perplexity and ChatGPT Search showed the strongest pickup from image-based links specifically, a reminder that content earning citations increasingly needs something visual worth linking to, not just well-argued prose.
Taken together, the picture is less that backlinks are dead and more that they have grown pickier. A domain with a handful of links from sites nobody trusts is unlikely to see any lift at all. A domain building a smaller number of links from genuinely authoritative, topically relevant sources, alongside the unlinked mentions and reviews that AI platforms also read, has a real reason to expect one.
What This Means for Your Next AI Search Budget
For Singapore businesses weighing where SEO and digital PR spend goes as next year’s budgets take shape, this research offers a useful filter. Resist the urge to buy volume, and resist the opposite urge to cut digital PR entirely on the assumption that citations have replaced it. The more productive question is whether a proposed placement, mention or piece of content would hold up as a genuine signal of authority to a human reader, because that is increasingly the bar AI systems are trying to approximate too.
Getting that balance right, across backlinks, unlinked mentions and the structured content these platforms actually retrieve from, is ongoing work, not a one-off fix. If you would like a second opinion on where your own link and citation profile stands, feel free to get in touch with our team.