Do Backlinks Still Matter for AI Search, or Just Mentions?

Links earned their twenty-year reign for a specific reason: PageRank-style ranking treats a link as a vote, and votes decide positions on a results page. An AI answer is built differently. The engine retrieves passages, synthesises a response, and decides which brands to name and which sources to cite. A language model's sense of which brands belong in an answer is shaped by how often and in what context those brands appear across the text it has read, linked or not. A brand discussed in a hundred articles, forum threads and videos exists to the model; a hyperlink with no surrounding discussion barely registers as a fact about anything.

That is the theory. What made 2025 and 2026 the point where the question got real is that the theory met large-scale measurement, the two studies in the next section. The timing also matters for budgets: link building is the single most expensive recurring line in most SEO programmes, and finance teams noticed that the channel it was built for is shedding share to answers. "Do backlinks still matter?" is a budget question wearing a technical costume, and it deserves a data answer rather than a reflex from either the link-building industry or its disruptors.

One definition pair before the data, because the whole debate turns on it. A mention is an AI response saying your brand name; a citation is an AI response linking a source. They diverge wildly in practice. Ahrefs' own tracking found AI Overviews attaching a link to only 10.7% of its brand mentions, against 26.9% on ChatGPT, and Semrush's companion research found 62% of AI citations are "ghost citations" that never produce a brand mention, a split we unpacked in which agency specializes in getting brands featured by AI. Mentions and citations are different outcomes, earned by different work, and backlinks sit upstream of only part of it.

Ahrefs, 75,000 brands

Correlating link and brand metrics against AI Overview brand visibility, the study found brand web mentions at 0.664, branded anchors at 0.527, brand search volume at 0.392, and backlinks at 0.218. The top three correlating factors were all off-site brand signals, not link signals, and brands in the top quartile for web mentions earned roughly ten times the AI Overview mentions of the quartile below.

A later Ahrefs analysis at the same scale added the most quotable single finding of the year: YouTube mentions were the strongest correlated signal of all for AI visibility across ChatGPT, Google AI Mode and AI Overviews, outperforming backlinks, domain authority and content volume. Ahrefs itself flags that every tested factor came back moderate to weak and that correlation is not causation, which are honest caveats worth carrying forward.

Semrush and Kevin Indig, 1,000 domains

Analysing about 35,000 datapoints across ChatGPT, ChatGPT with Search, Gemini, AI Overviews and Perplexity against core backlink metrics, this study found a strong positive relationship between Authority Score and AI mentions, with major gains appearing only in the top authority tiers. It also produced three findings that matter operationally. First, unique linking domains matter more than total backlink counts. Second, engines weight link quality differently, with ChatGPT Search and Perplexity valuing it least. Third, nofollow behaviour splits: Gemini and ChatGPT reward strong nofollow links (think Reddit, Quora, Wikipedia) somewhat more, while AI Overviews and Perplexity lean toward followed links.

The headline correlations side by side

Both studies are correlational and say so, and engine behaviour also shifts between model releases. Read them as a map of association, not a causal invoice.

At first glance they argue, since one says backlinks barely correlate and the other says authority strongly does. The reconciliation, worked through in Something Inc.'s side-by-side of the two datasets, is that Ahrefs measured link counts against how visible already-visible brands are (roughly 74% of its sample had at least one AI Overview mention). Semrush's coarser first question was whether a strongly linked domain shows up in AI answers at all. Once presence is set aside, the same analysis finds Semrush's authority-to-share-of-voice figure landing around 0.23, right on top of Ahrefs' backlink number.

That yields the cleanest mental model the data supports. A gate is the minimum authority a site needs before AI engines will include it in the answer pool at all, and a dial is how often a brand is named once it is inside, which tracks mentions. In short, links are a gate, mentions are a dial.

• The gate: a site with no authority struggles to enter the answer pool at all. Authority, accumulated through quality links from unique domains, still correlates strongly with whether you appear. Semrush's top-tier finding sharpens it: the gate effect concentrates at the high end, so moving from weak to moderate authority matters more than stacking links on an already strong profile.
• The dial: among brands past the gate, how often you are named tracks brand mentions about three times more closely than it tracks links. The growth lever, once present, is being talked about in the sources the engines read. That is the mechanism behind our earlier finding that third-party brand mentions matter for GEO, and why Reddit and forums shape what AI says about brands out of all proportion to their link equity.

The model also explains the YouTube result without mystery. YouTube is the single largest reservoir of brand discussion the engines ingest, and nearly none of it carries SEO-meaningful links. A signal that strong and that unlinked is the dial announcing itself.

Four jobs, and pretending otherwise overcorrects into a new mistake.

Links feed the rankings that feed retrieval. AI answers lean heavily on search indexes for candidate passages, and links still move those indexes. The chain is indirect but real: backlinks drive the organic rankings that feed AI retrieval systems, and a page that cannot rank is harder for a retrieval-augmented engine to fetch and quote.

Links are how crawlers find and revisit you. Discovery, crawl frequency and the freshness signals we covered in content refresh operations for AI search all travel the link graph.

Authority gates citation-worthiness. The engines are conservative, most of all in regulated categories, and that conservatism leans on domain-level trust. Link authority remains a large input to that trust.

A linked mention beats an unlinked one where it counts. Ahrefs' citation-rate tracking found the mentions that do carry links concentrate on high-volume queries, so linked mentions end up seen by more people even though unlinked mentions are more numerous. The ideal unit is not a link or a mention. It is a substantive discussion of your brand that happens to link.

What links no longer do is the thing the old budget assumed: scale visibility linearly. The tactics built purely to manufacture link count, above all syndicated wire placements, are optimising a 0.218 signal with tools the engines have learned to ignore. One 17-million-citation analysis found wire content earning 0.04% of AI citations.

Keep, reweighted: digital PR and editorial outreach, because the work was always about coverage and the KPI has now caught up. Judge placements by the venue's standing in AI answers (is it in the domains your engines actually cite? Your citation tracking knows) and by the surrounding discussion, with the link as a bonus that still compounds. Nofollow placements on Reddit, Quora and Wikipedia-adjacent surfaces now carry measurable weight on Gemini and ChatGPT, which quietly rehabilitates community work the old link maths wrote off.

Grow: the mention reservoirs. That means YouTube presence (reviews, comparisons and tutorials featuring your brand, yours and others') on the strength of the Ahrefs finding; community and forum participation done as genuine contribution; comparison and listicle inclusion on credible third-party round-ups; and the citable-asset flywheel, because original research is what earns substantive mentions, the dynamic we documented in what actually gets you cited by AI answer engines.

Cut: volume link building against domain-rating targets, syndicated wire distribution, and any placement whose entire value proposition is the hyperlink. These were weak before the AI data, and now they are measured weak.

Re-instrument: track unlinked brand mentions as a first-class KPI beside backlinks. The tooling exists in Ahrefs' Brand Radar, Semrush's AI toolkit and peers. Split reporting by engine, because the nofollow splits and quality weighting above mean the same placement buys different visibility on different engines. If your brand is below the authority gate, sequence honestly: foundation authority first, mention volume second, because dial-spinning before the gate opens moves nothing.

The model is how Lifewood's GEO programmes are scoped in practice. An authority audit decides whether a client is gated, in which case foundation link equity from unique, credible domains comes first. For clients past the gate, the growth plan is mention-weighted: placements in the venues the client's own citation tracking shows the engines reading, community presence where the category's buyers actually discuss vendors, and original research built to be quoted. Reporting carries both ledgers side by side, with linking domains and authority on one and unlinked mentions and share of answer on the other, split per engine and, for multilingual clients, per market, since the venues that constitute "the sources engines read" differ by language.

The evidence base here is unusually healthy, and worth using directly whoever runs your programme. Ahrefs and the Semrush and Indig collaboration published their methods and caveats openly; the ghost-citations research separates mention and citation outcomes so you can target the one your business model needs; and independent reconciliations like Something Inc.'s are the antidote to vendor cherry-picking in both directions. To compare providers directly, see our comparison of AEO and GEO agencies.

The one-line test for any link proposal you receive this year, from anyone: does this placement make the brand more discussed in a venue AI engines read, or does it only make a number go up in a link index? The first survives the 2026 data. The second is paying 2016 prices for a 0.218 signal.

• Pull both baselines: backlinks and referring domains from your SEO stack; unlinked brand mentions and share of answer from an AI visibility tool. You are about to manage two ledgers.
• Find your gate position: if domain authority is weak and AI presence is near zero, foundation link equity from unique domains is still step one. If you are visibly past the gate, the growth lever has moved.
• Re-grade every recurring placement by venue: does it appear in the domains your engines cite? Does the brand get discussed, or just hyperlinked?
• Cancel the pure-volume lines, meaning wire syndication and DR-target link buys, and redeploy the spend into the mention reservoirs: YouTube, communities, comparison inclusion and citable research.
• Add unlinked mentions to the monthly report beside backlinks, split by engine.
• Brief outreach teams on the new unit of value: substantive discussion in a cited venue, link welcome but secondary, with nofollow community placements back on the menu.
• Re-run the correlation sanity check twice a year against your own data, because engine weightings shift with model releases and the studies above are snapshots, not laws.