The intuitive model is that an AI assistant searches the web, finds the best pages, and summarises them. If that were true, different assistants asked the same question would converge on similar sources.
They do not. An analysis of citation data across platforms found that only about 11 percent of domains cited by ChatGPT are also cited by Perplexity. Ask two assistants about tariffs or immigration policy and you are, for the most part, reading summaries of two different bodies of evidence.
Ahrefs examined 15,000 queries and found that roughly 12 percent of URLs cited by AI tools also appeared in Google's top ten results. The other 88 percent came from pages that rank nowhere near the first page.
This gap has widened quickly. In mid-2025, around three quarters of Google AI Overview citations came from top-ten organic results. By early 2026, Ahrefs put that figure closer to 38 percent, and BrightEdge's measurement was lower still. Whatever the exact number, the direction is consistent: what ranks and what gets cited are drifting apart.
For political topics that matters more than it does for product reviews. The pages a search engine surfaces have at least been filtered by decades of ranking work. The pages an assistant retrieves have been filtered by something newer and less examined.
The differences are architectural rather than editorial. ChatGPT leans on Bing's index, which gives its citations a strong correlation with Bing's own rankings and a documented preference for high-authority reference sites. Perplexity runs a live web search on every query, maintains its own crawl, and has the strongest recency bias of the major tools.
That last point has a consequence worth sitting with. Perplexity's most-cited source category in one analysis was Reddit, at nearly half of top sources. On a political question, an answer weighted toward forum discussion is a different kind of answer from one weighted toward reference works, even when both are presented in the same confident tone.
Retrieval works best when there is a settled answer to retrieve. Ask an assistant for the boiling point of water and any reasonable source gives the same figure. Ask whether a tariff helped or hurt domestic manufacturing and there is no equivalent ground truth, only a contested literature and a great deal of advocacy written to look like analysis.
An assistant has no reliable way to tell an economist's working paper from a trade association's briefing when both are well formatted and confidently argued. Structure and clarity are exactly the qualities retrieval systems reward, and they are qualities that lobbying material has in abundance.
This is not a claim that assistants are biased in a partisan direction. The research does not show that, and the platforms disagree with each other too much for a single slant to be a coherent description. The problem is narrower and more mundane: on questions where sources conflict, the selection of sources is doing most of the work, and that selection is invisible to the reader.
The presentation of these answers does not vary with their reliability. Assistants tend to deliver conclusions in a uniform register regardless of how thin the underlying evidence is, and independent testing has repeatedly found that a meaningful share of generated citations do not fully support the claim they are attached to.
A reader has no way to tell, from the answer alone, whether they are looking at a well-sourced synthesis or a confident guess assembled from three forum threads.
None of this makes AI assistants useless for political questions. It makes them a starting point rather than an authority, which is roughly how a careful reader already treats any single source.
Three habits help. Ask the same question of more than one assistant, because the source overlap is small enough that differences are informative. Open the citations rather than trusting the summary, since the link is the only part you can verify. And treat a confident tone as a stylistic default rather than a signal of confidence in the underlying evidence.
The deeper point is familiar. A summary is a compression, and something is always lost in the compression. That is true of a fact-check rating, of a headline, and of an AI answer about a contested political question. The useful move is the same in every case: find out what was actually being summarised.