Blog · Third-party sources

The third-party sources AI engines keep citing

Published 5 June 2026 · Updated 8 October 2026 · 8 min read

The third-party sources AI engines keep citing
Photograph: Chris Linnett / unsplash

The short answer: The only well measured case is Google. Pew Research Center, analysing the March 2025 browsing of 900 United States adults, found Wikipedia, YouTube and Reddit together supplied about 15% of the links inside Google AI summaries, with .gov pages at 6% against 2% in ordinary results. The same shapes recur in the citation sets we collect on other engines: encyclopedia entries, large forums, video, government and standards pages, review directories, category roundups. The head of that list is not winnable. The four or five category pages behind it usually are.

That last claim is where most offsite programmes go wrong. People read a list of big domains, conclude they need a Wikipedia article and a Forbes mention, and spend a quarter on the two things least responsive to effort.

Why these pages, and not yours

An engine answering "best X for Y" does not run one query. Google describes its AI features as using a query fan-out, breaking the question into several related searches and assembling links from all of them. Each sub-query needs candidate documents, and the retriever has a budget.

A third-party roundup is cheap under that budget. One page already holds eight vendors, their prices, who each suits, and a verdict, in a structure a passage splitter can cut cleanly. Your product page holds one vendor and argues for it. For a comparison question, the roundup satisfies four sub-queries at once and your page satisfies a fraction of one. The engine is not preferring a brand. It is preferring density of answer per fetched document, which is also why the listicles you are missing from matter more than their traffic ever suggested.

Forums win for a different reason. They carry the sentences nobody publishes about their own product: what broke, what the migration cost, which plan is a trap. There is rarely a competing source for that, so retrieval has little else to reach for. We go into the specific case in why Reddit gets cited so much.

Government, standards and encyclopedia pages win on stability. They change slowly, they are linked densely, and they rarely contradict themselves. That is the whole of their advantage, and it is not an advantage you can copy.

The distribution is short at the head and very long after it

Two facts sit next to each other awkwardly. The top handful of domains take a visible share of all citations in the Pew sample, and we have not seen anything a normal company does move that share in the audits we have run. Meanwhile the pages that decide whether a mid-market buyer hears your name are small, category-specific and often unknown to the marketing team: a trade publication's annual roundup, a niche directory, one much-upvoted thread, a comparison blog run by two people.

Vendor research published over the past two years points the same way for error correction, with inaccurate third-party claims spread across a long tail of domains rather than concentrated in a few. Treat that as a direction, not a figure, because the underlying data is not open. The practical version holds either way: your movable set is small, specific to you, and will not appear on anyone's list of authoritative sites.

Engines also weight these differently from one another, which is why a source audit done on one assistant transfers badly to the next. The automated evaluation in arXiv preprint 2410.22349 (October 2024) measured three answer engines and found both inaccurate citation and engine-to-engine variation, so this is not only our impression. The mechanics of that split are in how each AI engine picks the sources it cites, and the single clearest case is in how Perplexity picks and ranks its sources.

A counterfactual test, using documented controls

Most offsite work is argued from correlation: we got mentioned there, visibility rose, therefore. You can do better, because at least one engine exposes source control in its API. Perplexity's search API takes a search_domain_filter holding up to 20 domains, in allowlist mode or denylist mode, not both at once.

That gives you a cheap experiment. Take a prompt you care about. Run it three ways:

  1. Unfiltered, several times, and record every domain cited.
  2. Allowlist mode, restricted to the third-party domains from run one, with your own site

    3. excluded. If your brand still gets named, those pages are carrying you.

  3. Denylist mode, blocking those same domains. If your brand vanishes, you have no

    4. independent corroboration anywhere else in the index, and your owned pages are not enough on their own.

The third run is the useful one. It separates two situations that look identical on a dashboard: a brand that is genuinely well-corroborated, and a brand whose entire presence depends on two roundups that could be rewritten next quarter. Run each configuration more than once. Answers move between identical runs in every prompt set we have run, so a single result is an anecdote. This test tells you about Perplexity's index and nothing directly about any other engine, because no other engine we use exposes the same control.

The stale citation problem nobody budgets for

A third-party page that mentions you is an asset until the day it is wrong. Prices change, plans get renamed, a feature ships, and the roundup written eighteen months ago keeps sitting in the index being retrieved. When the outdated page matches the question better than your current pricing page, an engine can repeat the outdated figure and attach a source to it, which makes it read as verified. Inaccurate citation was one of the common limitations measured across three answer engines in arXiv preprint 2410.22349, so a source attached to a claim is not evidence the claim was checked.

Owned content is not exempt. A stale comparison page of your own, or a help article still quoting last year's limits, contradicts your pricing page and leaves the engine to pick. In the contradiction cases we have examined, the passage that fits the question wins and not the newest one. We can see the output, not the ranking, so treat that as a pattern we have observed rather than a documented ranking rule.

So the offsite job has two halves, and the second one is the part that gets skipped: earn placements, then keep a list of every third-party page that states a fact about you and re-read it twice a year. Correction requests are boring, and they are an easier ask than a pitch, because you are offering an editor a free accuracy fix rather than asking for a favour. We have not counted acceptance rates, so take that as reasoning, not as a measured result.

What to do this month

Build the target list from citations, not from reputation. Collect the domains actually cited for twenty to thirty of your real buying questions, across at least three engines, several runs each. Sort by how often each domain appears, not by how impressive it is.

Then cut the list in two. Everything you cannot influence on a one-year horizon goes in a column you ignore. In the audits we have run, what remains is usually four to eight pages. For each, write down which state you are in: absent, present but ranked low, present and stale, or present and correct. Those four states have different fixes and only one of them is outreach.

A sample audit report shows how we lay this out, with the evidence for each cited source attached so a finding can be argued with rather than believed.

Honest limits

The one solid number here, the 15% and the 6% against 2%, describes Google AI summaries shown to 900 United States adults in March 2025, and nothing else. Assistant behaviour inside a chat interface is measured far less well, because nobody outside those companies publishes the logs, so where we describe other engines we are describing our own audit runs: samples of roughly 45 prompts across five engines, several runs each, not a census of anything. The counterfactual test is specific to Perplexity, and an absence of your brand under a denylist tells you what that index returned on that day, not what exists. The four-to-eight figure and the stale-citation pattern are things we keep seeing in client work rather than measurements we can hand you, and we have not published the underlying counts. We also cannot tell you how long a placement keeps paying: we have seen sources drop out after a model update with no change to the source page, and others persist for months, and anyone quoting a half-life for a citation is guessing.

FAQ

Do I need a Wikipedia article? It helps with entity resolution when it exists and is accurate. It is also governed by notability rules you do not control, and attempting to create one about yourself tends to go badly. Treat it as an outcome of coverage, not an input.

Are directory listings still worth claiming? Yes, if the directory shows up in your own citation data. Claim it, fill every field, keep the pricing current. The value is a machine-readable, frequently refetched statement of what you sell, which is also why a neglected listing does real damage.

How many third-party sources should I be targeting? In our audits, four to eight that you can realistically affect. Below three the dependency is fragile. Above ten the quality of attention drops and nothing moves.

Will a paid placement get cited? Sometimes, and it carries a risk: sponsored roundups are often thin, duplicated across sites, and quick to go stale. A page that nobody updates is a liability once your pricing changes.

Can I just block the pages that get me wrong? You cannot block someone else's site. You can ask for a correction, publish an unambiguous, well-marked version of the fact on your own domain, and make sure your own pages agree with each other.

Sources

Related reading

Why Reddit gets cited so much, and what to do about itThe listicles you are missing fromHow each AI engine picks the sources it citesHow Perplexity picks and ranks its sourcesWikipedia, and the brand facts engines inheritYouTube, transcripts and video citations