How AI answer engines decide what to cite, and why you should care twice

You ask an AI assistant a question squarely inside your expertise, and the answer arrives citing three sources. One is a content farm, one is a five-year-old blog post, and one is your competitor. The page you spent weeks making rigorous is nowhere.

The engine did not judge your competitor more correct. It never judged correctness at all. It retrieved candidate pages, extracted the passages that were easiest to lift and looked most like evidence, and assembled an answer. Understanding that mechanism changes how you read AI answers, and it changes how you write anything you want AI to find.

The GEO study from Princeton and collaborators ran a clean experiment: take real pages, apply one writing change at a time, and measure how often generative engines surfaced them in answers. Adding citations to sources, adding quotations, and adding statistics each boosted visibility substantially, in the range of 30 to 40% on their benchmark. Keyword stuffing, the old SEO reflex, made things slightly worse.

Read the list again and notice what it is: the markers of evidence. Engines quote pages that look like they have already done the work of proving things. Looking like evidence and being evidence are correlated, but they are not the same thing, and the gap between them is where a reader's judgment still earns its keep.

This mechanism is why an AI answer's sources deserve the same check as its claims. The engine rewarded extractability: clear structure, quotable sentences, visible numbers. A content farm that formats well can outrank a rigorous paper that does not. Columbia's Tow Center measured the result from the output side: even the best engine cited sources incorrectly 37% of the time.

So when you open a citation, you are not just checking whether the passage supports the claim. You are also asking the question the engine never asked: is this source worth believing at all? Who wrote it, on what evidence, with what incentive. Two layers, one habit.

The same findings are a fair playbook for being found. State the answer plainly near the top, because extraction favors pages that commit. Attach real statistics with their sources. Quote authorities accurately. Structure claims so each one stands on its own line with its evidence beside it.

Notice that every item on that list also just makes your work more trustworthy to humans. That is the honest version of GEO: the machines reward the surface features of rigor, and the durable strategy is to have the rigor underneath, because the audits that expose hollow sources are getting better every year. Write pages where the extraction and the truth are the same thing.

You may have noticed this post does the things it describes: an answer-first opening, named studies with links, specific numbers with their caveats. Every post on this blog does, and every answer inside Tatsulok goes further: assembled only from your documents and curated collections, with each sentence linked to the exact passage behind it, and an honest gap statement when the sources are silent.

That is our position in this ecosystem: the engines choose what looks like evidence, so we build the place where looking like evidence and being evidence are forced to coincide.

This week's drill, no signup needed: ask any AI assistant a question in your field and open every source it cites. Rate each one: would you have cited this? Count how many pass. That number is your calibration for how much the phrase "according to sources" should mean to you.