AI search citations: inside the farm Perplexity cites

Two audits landed on the same day this week. Both point at the same soft spot, which is the citation layer of AI search. The first names three sites behind 215,128 machine-made “best software” pages. Perplexity’s grounded answers cite those sites across software categories (the report). The second found that a third of Perplexity’s citations lack the number they are cited for (the companion audit). I read both reports, then fetched the named sites myself to see whether the claims still hold. Most of them do. If your team quotes AI search answers in front of a client, this is your supply chain too.

What the audit actually measured

The setup is concrete, which is why it earned 448 points on Hacker News within a day. On September 2, Trellner Research put 380 buyer-intent categories to two Perplexity models through OpenRouter. The categories ran from “CRM software” down to “museum collection management software”. One prompt per category per model made 760 calls in total. Every call asked for a ranked top five, with the official homepage domain for each product. That produced 3,800 recommendation slots naming 1,807 distinct products. It also produced 7,534 citations across 2,055 domains, each checked against the Tranco top-1M list for September 1.

The topline is bleak. Some 59.8% of the 7,534 citations point at domains ranked worse than 100,000th. Another 23.4% sit outside the top million entirely, which means Tranco has never ranked them at all. The median citation that does rank lands at position 71,611, deep in the long tail.

The top of the citation table looks respectable: G2, Reddit, Gartner. The surprise sits third. Guideflow sells interactive product demos, and its marketing blog was cited 194 times across 96 of the 380 categories. That puts a vendor’s content hub ahead of Gartner at 158 and behind only G2 and Reddit. Nothing about the blog is deceptive, because listicles about markets you do not operate in are standard content marketing. The retrieval layer is what turned it into a top-three evidence base for buying advice. For scale: Wikipedia drew three citations out of 7,534.

The three sites are built to be retrieved

The finding that matters most is the audience the pages address, more than the scale, though the scale is absurd. Grounding is the retrieval step where a model fetches documents before it answers. Two of the three sites greet crawlers with an HTML title of “Facts & Grounding Page”. That detail is the one the rest hangs on. I fetched both homepages on September 3, and the titles are unchanged. The meta description offers verified facts in one machine-readable record, addressed to whatever software is reading. A homepage that is titled for a retrieval step is not talking to buyers.

The structure behind the titles is methodical. All three brands were registered between December 2023 and May 2024, on the same Cloudflare nameserver pair. Each runs the same page template and keeps a six-post blog about the other brands in the set. Their sitemaps list 103,578, 107,083, and 105,541 URLs. Some 215,128 of those are /best/-software/ buying guides. Worldmetrics even advertises custom research from EUR 5,000 on the same site that serves the generated lists.

The report pulled the same category page from all three sites: project estimation software. Each page states its ranking in JSON-LD, so the disagreement is on the record. Nine staff names are credited across three sites for one category.

Gitnux’s winner does not appear in Worldmetrics’ top five at all. The bylines carry an artifact that my own fetches confirmed: unrendered template variables, still reading “Within the next 41 days”. That sits oddly beside the editorial-process banner the pages announce. One page labels its verdict “AI-verified, expert reviewed”, which is a bold claim for a page with a broken byline.

One caveat matters: low rank is popularity, not guilt. The report is explicit that nobody has shown these sources change the answers, and it declines to claim that itself.

How much to trust this report

Treat it as a useful measurement from an unverified publisher, and weight the checkable parts over the prose. The Hacker News thread hit 448 points within a day, and the objections were quick. Commenters flagged the prose as AI-written and noted the founder returns zero search results. The same account has posted four similar reports from research sites the thread calls AI-generated. That track record would sink most sources. It does not sink this one, because the central claims are checkable, and I checked the supply side myself. The titles are live, the variables are unrendered, and the dryad.co redirect still lands on a gambling portal. Shared infrastructure is strong circumstantial evidence of common control, and the report words it exactly that carefully.

The limitations section is the report’s best feature. It flags the one-day snapshot, the single-engine scope, and the fact that the two Perplexity tiers share one retrieval layer. The tiers returned byte-identical citation lists in 289 of 380 categories, so treat this as one search stack sampled twice. It also declines to claim answers get worse when the sources are removed, because nobody has tested that.

Why AI search retrieval is easy to farm

Grounding is a retrieval step, and retrieval has no taste. Anything crawlable, structured, and fresh can become the evidence base for an answer. A page that states its verdict in JSON-LD spares the model interpretation, which is exactly what these sites do. The farm worked that out before most marketers did.

The failure is bigger than source choice, though. Haus Research checked whether cited pages contain the numbers they are cited for. About a third do not, per its report, which means claim support fails even when the source is legitimate.

There is money behind this, and plenty of it. Commenters on the same thread put one GEO vendor, Profound, at $155M raised and a $1B valuation. That figure is forum commentary I could not verify, so read it as a signal rather than a fact. Either way, the optimization industry has noticed that models can be steered. One comment described the method: measure the distance between your page and the model’s answer, then rewrite until it shrinks. That is programmatic SEO with a new customer, and the model is the reader. My programmatic SEO post covers the disciplined version of that job. This is the undisciplined version, pointed at a reader that never bounces. It is not a one-week panic either; my last AI search read covered the drift from the demand side.

What a marketing team should do about it

You cannot fix their retrieval layer, so work both directions: what you consume and what the engines say about you. On the intake side, treat AI search citations like a junior analyst’s links. The number enters a deck only after you open the page and find it there. It is the same gate I described for AI Overviews content, applied one layer down.

On the exposure side, ask the engines your category’s buying questions and log what they cite. If an unfamiliar vendor blog outranks Gartner in your citations, that is your retrieval environment. Knowing that beats guessing when a client asks why the AI recommended a competitor.

One thing not to do: publish your own facts page for the models. That is not a strategy; it is the farm, and you do not want your brand in that sentence. The pre-trust check I run before quoting an AI answer now looks like this:

• Open the citation before the number reaches a client, and confirm the page states it.
• Spot-check one cited domain per answer, because a human should still call it an authority.
• Ask the engines your top buying questions monthly, and log the cited domains.
• Watch for vendor blogs outranking review sites in your category.
• Discount any source whose homepage is addressed to models rather than buyers.