Getting Cited Isn’t the Same as Getting Heard.

Across 21,143 AI citations, being listed as a source and shaping the answer turned out to be two different things. The engine that cited the most sources shaped each answer the least.

When I worked for The Dallas Morning News marketing department, we often heard from editors that people love to see people. And decades have passed, but if I’m honest, there’s a particular satisfaction in watching a machine say my name. Do you feel that?

You ask ChatGPT where to buy a used shuttle bus, or a coated cast-iron skillet, or a divorce lawyer in Orange County, CA, and there in the little stack of sources sits your website, blue and underlined and looking very much like proof that all the money you spent was money well spent.

Well. Enjoy that feeling. It might not survive your intended customer’s contact with the data.

Three researchers ran 602 controlled prompts through ChatGPT, Google’s AI Overview, and Perplexity and collected 21,143 citations along the way, all of it logged in a public study on arXiv. Then they did the thing nobody selling you an “AI visibility score” wants done. They checked whether the cited pages did anything at all.

Turns out being cited and being used aren’t the same event. And the researchers, wild bunch that they are, split them on purpose and gave them names.

1. Selection is when the model runs a search and picks your page off the shelf.
2. Absorption is when your page hands the answer something it keeps. A sentence. A number, or a fact the model repeats because it got it from you.

You can have the first without the second, and most pages do. The engine lists you as a source, then writes its reply out of somebody else’s material, and you’re left standing in the footnotes like someone who got invited to the wedding and seated behind a pillar.

The three engines don’t behave the same way. Perplexity cites around sixteen sources per answer, Google about twelve, ChatGPT closer to seven. So Perplexity looks generous. It’s like it’s lonely and trying to make friends.

But when the researchers scored how much each cited page shaped the answer, on a zero-to-one scale, ChatGPT’s average page came in at 0.27 and Perplexity’s at 0.06. Google’s was lower still. Perplexity hands out more citations and means less by each one. If it were your friend, you’d say it’s thirsty and insincere. Lovely qualities. Oh, those researchers.

Now, I’m no logical genius, but my daddy didn’t raise a fool. The dashboard that counts mentions is scoring the generous engine as the winner and the stingy one as the loser. That doesn’t make sense. It’s got the truth exactly backwards.

The study went one level deeper, into what kind of page gets absorbed once it’s through the door. This is the part I’m putting in my Apple Notes. Pages with real numbers in them scored about 60 percent higher than pages without. Definitions, comparisons, how-to steps, code, all up by half or better.

And the one format every consultant on earth tells you to build, the FAQ, that tidy question-and-answer accordion, was the only genre that scored slightly negative. The machine doesn’t want your questions! It wants an answer it can lift.

Now let’s walk that onto a car lot. When I worked at Mitsubishi, we used to say around 80 percent of people searched the internet before buying a car in 2020. Now, it’s more like 93 percent. And roughly a quarter of new-vehicle shoppers bring an AI into the research before they talk to a human, and the ones who arrive through an AI answer close at more than four times the rate of ordinary web traffic.

So when a dealer’s site shows up as a “source” and adds nothing the buyer actually reads, you’ve got a loss that photographs like a win.

One honest note is that the influence score is a proxy, a careful stand-in for something the engines won’t show you directly. And it’s built from one big pile of prompts. It’s the best measurement we’ve got, but it’s still not the final word. Treat it as a strong signal and not a commandment.

Most AI-search reporting counts selection, because selection is easy to count. Your name appeared. Ding. Almost nobody checks absorption, because absorption is hard, and because the honest version of the number is smaller, and nobody enjoys handing a client a smaller number.

I run these audits. The mention’s the cheap half. The question worth asking, and frankly the only one that moves metal, is whether the machine kept a single word you gave it. And here’s the part I’ll put my name on. I don’t just test what the machines keep. I write the pieces and the copy built to get kept.

Count that. All those other numbers are just applause from people you thought were in the market, but they weren’t listening.

Source: Zhang Kai, He Xinyue, Yao Jingang, “From Citation Selection to Citation Absorption,” arXiv 2604.25707 (April 2026). Vehicle-shopper AI use: Cox Automotive Car Buyer Journey Study. Online research share of 92–95%: Urban Science / Harris Poll and ConsumerAffairs, 2026. The 4.4x conversion figure: dealer-analytics reporting (Autosweet; C-4 Analytics). Influence is a proxy score, not a direct read of the models’ internals, drawn from a preprint.