How do AI search engines decide which companies to cite?

Retrieval first: they can only cite what they can fetch

Every AI search engine — ChatGPT with browsing, Perplexity, Google's AI Overviews, Claude with web search — answers a question in two steps. It decides what to look up, often breaking your question into several narrower searches, then reads what comes back and writes an answer with citations drawn from those pages. A company can fail at the first step without ever reaching the second. If the page that would answer the question is rendered by JavaScript after load, blocked by a robots.txt that never mentions GPTBot or PerplexityBot, or buried under a hero image and three paragraphs of positioning, the engine either never sees it or sees nothing quotable.

This is the unglamorous part of the answer, and it accounts for a large share of "why does it never name us." The page has to exist as server-rendered HTML, at a stable URL, with the answer near the top, and the crawlers have to be allowed in by name. None of that is optimization. It is admission.

Ranking on Google matters less than it did a year ago

Marketers assumed for a while that AI citations were just the top of the search results wearing a different hat. Ahrefs tested that twice, and the two results a year apart tell the story.

In July 2025 three quarters of cited pages ranked in the top ten for the same query, with a median position of 2 for the primary citation. By March 2026 that had halved: 38% of citations came from the top ten, and nearly a third came from pages that did not rank in Google's top 100 at all. Ahrefs attributes the shift to the engine's fan-out — it issues several related searches and cites from all of them, so a page can be cited for a question it never ranked for. YouTube alone accounted for 5.6% of all AI Overview citations in the 2026 dataset.

The practical reading is that a top-ten ranking is still helpful and no longer sufficient. Pages that answer a narrow sub-question well can be cited for the broad question a buyer actually asked, even if they never rank for it. That is the opening for a smaller company: you do not have to outrank the incumbents on the head term to be named in the answer.

Mentions beat links

The second thing engines weigh is whether they have encountered the brand before, and here the data is unusually consistent. Ahrefs' May 2025 study of 75,000 brands measured which signals correlated with being mentioned in AI Overviews. Branded web mentions led at a Spearman correlation of 0.664. Branded search volume was 0.392, Domain Rating 0.326, and the raw number of backlinks only 0.218. 5W's 2026 State of AI Citations report, built on Profound's dataset of 680 million citations across ChatGPT, Google's AI Overviews and Perplexity, found the same ordering for ChatGPT: brand search volume is the strongest known predictor of citation likelihood at 0.334, materially stronger than backlinks. Both reports say plainly that correlation is not causation, and both draw the same practical conclusion — the web talking about you predicts assistants naming you far better than the link graph does.

Where the web talks about you matters too. Ahrefs' September 2026 tracking of ChatGPT's most-cited domains puts Reddit first at 16.8% and Wikipedia second at 7.0%, followed by Consumer Reports, Forbes and YouTube. The engines treat community discussion, reference material and comparison sites as more trustworthy than a vendor's own page. A company that is named in three forum threads, one comparison article and a LinkedIn discussion is easier to cite than one with a beautiful website and no third-party footprint, and Ahrefs' brand study found about 26% of the 75,000 brands it examined had no AI Overview mentions at all.

Freshness and specificity

Two more filters sit on top. Ahrefs' July 2025 analysis of 16.975 million cited URLs found AI assistants cite newer content than Google ranks: an average of 1,064 days since publication across assistants against 1,432 for organic results, with ChatGPT the most recency-biased at 958 days and Perplexity at 1,166. Both ChatGPT and Perplexity also tend to list newer sources first. A page that has not been touched in three years is at a disadvantage against one revised last quarter, which is a strong argument for maintaining a small set of answer pages rather than publishing once and moving on.

The last filter is the one you control most directly. When an engine has several candidate pages, it quotes the one that answers the question in plain, specific language it can lift into a sentence: a price range, a timeframe, a named category, a direct comparison. A page that opens with "it depends" has given the model nothing to cite. A page that opens with a number has.

Put together, the decision looks like this: can I fetch it, does it answer the sub-question I just asked, is the brand one the web already talks about, is it recent, and can I quote it? Ranking helps with the second. Everything else is a publishing and reputation problem, which is why the companies that get cited are usually the ones that decided to answer buyers' questions in public, consistently, and kept doing it.

Where SalesARC fits

ArcClaw™ AEO is SalesARC's answer-engine service, and it is built around the decision above rather than around rankings. It identifies the questions your buyers ask assistants, samples each one repeatedly and reports your citation rate per question, then publishes the answer pages, forum responses and third-party articles that change the result — reported next to your Google visibility so you can see both. From $499/mo for 50 tracked questions. The Answers hub on this site is our own team's version of the same work, and this page is one of the tracked questions.

See ArcClaw™ AEO or build a plan.