AI Citations Are Topic-Local. Backlinks Are Domain-Global. That Is the Whole Relationship.
Across six measured topics, 78.8% of the domains AI answers cited appeared in exactly one topic. Domain authority buys breadth in search. Citation is won per question. Here is the measurement.
Backlinks and AI citations are not the same currency, and the difference is measurable: backlinks buy domain-level breadth, AI citation is won per question. Across six topics we measured on 6 and 10 September 2026, 78.8% of the domains that AI answers cited appeared in exactly one of the six. A domain that earns citation on one question mostly does not carry that standing to the next one.
That is the honest answer to the question people keep asking in two incompatible forms — "do backlinks still matter for AI search?" and "is domain authority dead?" Both questions assume citation behaves like ranking. It does not behave like ranking, and the measurement below shows where the two systems actually touch.
The mechanical reason is not mysterious. The link economy is built on a graph-level score — PageRank and its commercial descendants estimate the standing of a domain or page within a web graph. Generative answers are assembled by retrieval over passages, the architecture introduced as retrieval-augmented generation by Lewis et al. in 2020, where the unit competing for inclusion is a passage matched to a query. Different unit, different competition, different winner.
What we measured
We ran topic-level citation snapshots against a live AI-answer corpus for six topics: techcrunch company categorization, machine relations, author credentials AI citation, how to measure AI visibility, answer engine optimization, and ai citations vs backlinks seo. Five were measured on 10 September 2026; how to measure AI visibility was measured on 6 September 2026.
Each snapshot returns the set of domains cited across sampled AI answers for that topic, with a mention count per domain. Total cited domains per topic ranged from 31 to 3,012. For cross-topic comparison we use each topic's top-25 cited domains, because that is the slice the provider returns in full for every topic.
Combining those six top-sets gives 99 distinct domains. Here is how many topics each domain appeared in:
| Appeared in | Number of domains | Share |
|---|---|---|
| 1 topic | 78 | 78.8% |
| 2 topics | 8 | 8.1% |
| 3 topics | 6 | 6.1% |
| 4 topics | 3 | 3.0% |
| 5 topics | 3 | 3.0% |
| 6 topics | 1 | 1.0% |
One domain — youtube.com — appeared in all six. The domains appearing in five were en.wikipedia.org, reddit.com, and linkedin.com.
Note the direction of the bias here. Truncating to top-25 by mention count favours finding recurrence, because the domains with the most mentions are the ones most likely to be large and cross-topic. The 78.8% single-topic figure is therefore a conservative floor. On the full cited sets, topic-locality would be higher, not lower.
Finding one: the recurring set is platforms, not authorities
Twenty-one domains appeared in two or more topics. Sorted by how many topics they reached:
| Domain | Topics | Type |
|---|---|---|
| youtube.com | 6 | Platform |
| en.wikipedia.org | 5 | Reference |
| reddit.com | 5 | Platform |
| linkedin.com | 5 | Platform |
| quora.com | 4 | Platform |
| developers.google.com | 4 | First-party documentation |
| searchengineland.com | 4 | Category publication |
| indeed.com | 3 | Platform |
| medium.com | 3 | Platform |
| semrush.com | 3 | Category vendor |
| coursera.org | 3 | Platform |
| seo.com | 3 | Category publication |
| forbes.com | 3 | General publication |
The pattern is not "high domain authority sites recur." It is that user-generated platforms and first-party documentation recur, and almost nothing else does. YouTube, Reddit, LinkedIn, Quora, Medium, Coursera and Indeed are not authorities on these topics; they are containers that hold a passage about nearly every topic. Wikipedia and developers.google.com recur for the opposite reason — they are canonical first-party or reference sources with genuine coverage breadth.
Strip those two groups out and the recurring set is four names: Search Engine Land, Semrush, SEO.com, and Forbes. Three of the four are category publications that cover this specific subject matter constantly — the same reason Search Engine Journal turns up in two of the six. That is topic coverage, not portable authority.
Finding two: inside a topic, mention count does not track link authority
The clearest test is the topic that is itself about backlinks. Here is the full top-25 cited set for ai citations vs backlinks seo, measured 10 September 2026 across 377 total cited domains:
| Domain | Mentions |
|---|---|
| youtube.com | 62 |
| reddit.com | 36 |
| semrush.com | 21 |
| linkedin.com | 14 |
| adsmurai.com | 12 |
| developers.google.com | 12 |
| digitalmarketinginstitute.com | 12 |
| seo.com | 10 |
| coursera.org | 9 |
| signal-ai.com | 9 |
| searchengineland.com | 7 |
| onelittleweb.com | 7 |
| sitepoint.com | 7 |
| caylor-solutions.com | 7 |
| seranking.com | 6 |
| searchatlas.com | 6 |
| quora.com | 5 |
| moz.com | 5 |
| manhattanstrategies.com | 4 |
| evergreen.media | 4 |
| en.wikipedia.org | 4 |
| forbes.com | 4 |
| riverworksmarketing.com | 4 |
| seocrawl.ai | 4 |
| behindrankings.com | 4 |
On a question about backlinks and search authority, caylor-solutions.com was cited seven times — more than Moz, the company that invented Domain Authority as a metric, and more than Forbes. behindrankings.com and riverworksmarketing.com tied Forbes and Wikipedia at four each. searchatlas.com, with the smallest AI search volume in the set, was cited more often than Wikipedia.
If citation were a function of link authority, the Domain Rating ordering and the mention ordering would correlate. In this set they visibly do not. Small agency sites and category tools sit inside the same answers as Forbes and Wikipedia, on the one topic where link authority is the subject.
So what does transfer from SEO?
Three things transfer, and they are all upstream of ranking rather than downstream of it.
Retrievability transfers. A page that is not crawled, rendered and indexed cannot be cited. Google's own AI features documentation makes the dependency explicit: the same crawling and indexing eligibility that governs search results governs AI surfacing, a continuity Google set out when it introduced generative answers into Search. The same holds for answer engines that publish their crawler behaviour, such as Perplexity's documentation. Every technical SEO investment that gets a page fetched and parsed keeps its full value — your robots directives, your clean status codes, your structured data, and your presence in the open crawl corpora such as Common Crawl that many retrieval pipelines draw on. This is the part of the discipline that did not change.
Coverage transfers. Sites that publish continuously on a subject show up across that subject's questions. Search Engine Land recurring across four of six measured topics is not authority leaking across domains; it is a publication that has written about all four. Coverage is why the category publications recur and the general publications mostly do not.
Link-earned distribution transfers indirectly. Backlinks still drive the traffic, discovery and reference behaviour that get a page crawled more, cited by humans, and quoted on the platforms that themselves get cited. The mechanism is real but it is second-order — links get you retrieved and referenced, not selected.
What does not transfer
Domain authority as a portable asset. In search, a strong domain lifts a new page on a new topic before that page has earned anything. Our data shows no equivalent effect: 78.8% of cited domains were topic-bound, and the four non-platform domains that recurred did so through subject coverage rather than through domain strength. You do not arrive at a new topic pre-credentialed.
Ranking position as a proxy for citation. These are separate selection systems, and we have measured them diverging on our own properties. For the exact query "ai citations vs backlinks," our pages sit at positions 29.7, 58 and 85 — while for the broader question this page answers, an existing page ranks at average position 2.02 across 1,501 impressions and takes zero clicks, because it mentions backlinks twice in 124KB of text. Ranking on a query and answering it are different achievements. We have written about that divergence at length in the citation gap between ChatGPT citations and Google rankings and in why SEO rankings can leave you invisible in AI search.
What to do instead
Stop treating authority as an account balance and start treating it as a per-question claim.
- Pick the question, not the keyword. Citation is awarded to the passage that answers a specific question. Passage retrieval is the selection unit, so the asset has to be built at that granularity.
- Answer it directly and early. How AI search engines decide what to cite covers the selection behaviour; the practical consequence is that a page which circles a question ranks and does not get cited.
- Measure per topic, not per domain. A single portfolio-wide citation number hides exactly the topic-locality this data shows. Use citation rate and share of citation with the topic declared as the denominator.
- Expect to earn each topic separately. Why domain authority scores mislead in a niche expert strategy is the strategic version of this finding; the measurement above is the evidence for it.
- Watch the platforms. If YouTube, Reddit, LinkedIn and Quora appear in nearly every topic's citation set, presence in those containers is a distinct distribution channel, not an afterthought.
Limits of this measurement
We are stating what we measured and no more.
- Six topics. Six is enough to show the effect and not enough to size it precisely across an industry. The per-topic uniqueness rate ranged from 24% to 88%, so topic-locality varies substantially by topic breadth.
- Top-25 truncation. Cross-topic comparison uses each topic's top-25 by mention count. As noted, this biases toward recurrence, making 78.8% a floor.
- One provider, near-single date. Five of six snapshots are from 10 September 2026 and one from 6 September 2026, from one AI-answer measurement corpus. Different providers sample different engines and prompt sets; we have not replicated this across providers.
- Correlation, not causation. We did not measure backlink profiles for these 99 domains. The claim is that the citation ordering does not match the well-known link-authority ordering of the named domains in the set, not that we ran a controlled link study.
- No temporal component. A single snapshot cannot tell you whether topic-locality decays or compounds over time. That requires repeated measurement, which we are doing.
For the source-level companion to this work — which domains hold the citation economy on a contested category term — see the AEO citation economy source map on MachineRelations.ai. For engine-specific selection behaviour, see how Perplexity selects sources.
The short version: keep the technical SEO, keep earning links, and stop expecting either one to pre-pay for a topic you have not answered yet.