Do Backlinks Help You Get Cited by AI? Perplexity Cited Ahrefs and Two Unknown Startups in the Same Answer

Short answer: mostly no, and the clearest evidence I have is a single answer that quoted a domain with fifteen years of link equity alongside two startups nobody has linked to.

The answer in question

We store the URLs that answer engines cite on every visibility check we run for the products tracked in CueScout. Going through the grounded ones, here is the full citation set Perplexity returned for "How do I monitor when AI assistants cite my competitors but not me?":

• ahrefs.com/blog/brand-mentions/
• radarkit.ai/blog/monitor-competitor-mentions-in-ai-search/
• turboaudit.ai/ai-monitoring
• reddit.com/r/ProductMarketing/comments/1k10tlt/top_5_tools_to_monitor_your_brands_presence_in_ai/
• semrush.com/blog/ai-visibility-toolkit/

Ahrefs and Semrush are two of the strongest domains in the entire SEO category. radarkit.ai and turboaudit.ai are small products whose link profiles would not put them within sight of page one for that query. The engine read all five and treated them as five sources.

If authority were the dominant input to what gets cited, those two would not be in that list. They are in it because they had a page that answered the question directly.

It was not a one-off

The same pattern showed up across the six questions we went through, 23 unique cited URLs in total, collected between 17 June and 9 July 2026. The full teardown is here; three findings bear on the backlink question.

On "what's the best tool for finding potential customers on Reddit who are actively looking to buy", three of the four citations were linkeddit.com, ogtool.com, and replyagent.ai. A homepage and two small product sites. No Ahrefs, no G2, nothing with authority. For a commercial query, in a competitive category.

catchintent.com was cited three times on a single question, on three different pages of its own domain. They had written about buyer-intent signals on Reddit, so when someone asked about buyer-intent signals on Reddit, the engine had three of their pages to read. Nothing about that is a link-acquisition outcome.

Fourteen of the 23 cited URLs sat on the domain of a vendor selling something in the category being asked about. Zero were G2, Capterra, or Wikipedia, which are exactly the destinations anyone would name if you asked which sites AI trusts.

Twenty-three URLs from one engine is a small sample, and I have said so every time I have used it. But you do not need a large sample to observe that a retrieval set can contain a DR 90 domain and a domain with almost no links, side by side, answering the same question.

Our own numbers, which are embarrassing and useful

I can run this experiment on myself. In late July I checked CueScout in Ahrefs: 372 referring domains, domain rating 0.1.

Those two numbers do not belong together, and the explanation is that almost none of those links are real. Two days earlier the same tool had shown roughly 2 referring domains. What happened in between was a wave of scraper and directory sites auto-generating pages that mention new domains. It is noise, and I mention it only because it made a clean natural experiment: for a few weeks we had a backlink profile that looked superficially non-trivial.

AI citations to cuescout.com over that period: zero.

Meanwhile radarkit.ai and turboaudit.ai, which I would guess are in a similar position link-wise, were being cited by Perplexity for a real buying question, because they had each written a page on the thing being asked.

The same confusion runs through our Search Console data, and it is worth flagging because it is how people talk themselves into believing links are working. About 12 percent of our impressions turned out to be scrapers firing -site: operator queries, logged at positions 1 through 8. Averaged in, they made a blog post that ranks nowhere look like it was sitting on page one. Two vanity metrics pointing the same wrong direction.

Why the mechanism predicts this

An answer engine does not consult a ranking of brands. It runs a search, pulls a handful of pages, reads them, and writes a summary of what those pages say. Nothing in that loop counts links. The model reads text.

Links do get an indirect route in. The search layer the engine queries is shaped by classic ranking signals, so a page with no visibility anywhere is less likely to be pulled into a retrieval set. Authority influences which pages become available to cite. What it does not do is make you citable through an accumulated score, and that is the confusion behind most GEO packages being sold with a link building line item attached, because link building is what the agency already knew how to do.

Size matters too. A retrieval set for a competitive question might be five to ten pages. There is no gradual climb through it. You are inside the set or outside it, and what puts you inside is whether the pages that got retrieved happen to describe you.

There is a second reason the two systems diverge, and it is bigger than the link question. When we pulled every page the engines cited across one customer's question set and checked each against Google's top 100 for the query that surfaced it, 58 of 96 cited pages did not appear in the top 100 at all. If most cited pages are not ranking pages, then whatever produces rankings is not what is producing citations, and links are the main thing that produces rankings.

What the retrieval layer does reward

If not links, then what puts a page into the candidate set? From watching which of our own and our customers' pages get pulled, roughly:

Topical match at the passage level, first and by a long way. Engines retrieve chunks of a few hundred words, so a page containing a section that answers the exact question beats a page that covers the general area at greater length and higher authority.

Being crawlable and rendering server-side. Nothing exotic, but a surprising number of small-vendor sites put their entire value proposition behind a JavaScript bundle and then wonder why they are invisible.

Freshness, on questions that read as current. "Best X in 2026" pulls recent pages hard.

Having several relevant pages rather than one. catchintent.com was cited three times on one question because it had three pages worth pulling. That is the closest thing to a compounding asset in this game, and it looks a lot more like content depth than like link equity.

What this means if link building is your skill

I want to be careful here, because "backlinks don't work anymore" is exactly the kind of claim that gets written by someone selling an alternative, which is me.

The skills transfer almost completely. Finding the sites that matter in a category, working out who runs them, writing a pitch that a busy editor answers, following up without being irritating: that is the job either way. What changes is the target and the definition of success.

The target moves from "any site with authority that will link to me" to "the specific pages that show up in my category's answers", which is a much shorter and more researchable list. The success condition moves from a link to a mention with a good twelve-word description next to it. A nofollow mention in a cited listicle is worth more for this purpose than a dofollow link from a site no engine ever retrieves, which inverts about fifteen years of instinct.

So it is a retargeting rather than a redundancy. The agencies that will do well at GEO are mostly the ones that already know how to get a mention placed.

Where I could be wrong

Three places, and I would rather name them than have someone name them for me.

One engine, one time window, six questions. Perplexity cites more openly than most, which is why its sources are readable at all. ChatGPT with search, Google AI Mode, and Claude may weight authority differently, and I do not have clean data on them at this size.

Absence of G2 across 23 URLs is not proof G2 is irrelevant. In categories with heavy review volume and higher purchase consideration, I would expect directories to appear more often. Our sample cannot see that.

"No direct effect" is not "no effect". The indirect route through the search layer is real. A brand-new domain with no links anywhere has a lower ceiling simply because fewer of its pages get surfaced to be retrieved in the first place.

What I am reasonably confident about is the allocation question, which is what anyone reading this actually wants settled. Given a month and a goal of AI citations, getting named inside the pages your category's engines retrieve will move more than a month of general link acquisition. That is not a claim that link building is dead. It is a claim about which lever sits closest to the outcome.

What to do instead

Find out which pages supply your category's answers. That is an afternoon by hand, or a few minutes with our citation source radar. Then work the list by page type: ranked lists are an outreach problem, competitor blog posts are a content problem, review directories are a slow product-marketing problem, and forum threads are not a target at all.

The mechanics are split across three posts: how to find which listicles get cited, how to get added to one, and why this is not the same job as link building even though it uses the same skills.

If anyone has a larger dataset that contradicts this, I would like to see it. There are a lot of confident claims in this space resting on nothing, and I do not want to add one. The larger one I have since collected is the 533-URL corpus study, and it says the same thing about Ahrefs.