Your GEO Checklist Is Mostly Placebo

Every vendor selling AI visibility has a slide that answers how to get cited by ChatGPT. Add an FAQ block. Add a summary box. Add schema markup and an author byline. Speed the page up. I have sat through that slide more times than I want to count, and I have nodded along to most of it, because the numbers behind it looked real. They were real. They were also, it turns out, measuring something else entirely.

On October 10, Mahe Chen and Nick Koudas of the University of Toronto published the study that breaks the slide. They collected 90 days of answers from ChatGPT, Gemini and Perplexity on commercial questions, ending up with 158,099 answers and 468,037 citations, and then they asked a question the vendor research almost never asks: does this feature still predict citations when you compare two pages on the same website?

Mostly, no. And the gap between those two ways of counting is where a lot of retainer money has been going.

Our position, stated plainly before the evidence: the content half of generative engine optimization has been oversold, the retrieval half has barely been sold at all, and the correction is good news for anyone who would rather write well than decorate pages.

The GEO checklist that tests well and does nothing

A page feature that predicts AI citations across different websites can be worth nothing on your own site, because the comparison across sites is also comparing everything else about those sites: their authority, their topic, their publishing budget and the kind of team that ships an FAQ block in the first place.

This is the oldest trap in measurement and it is embarrassing how well it hides. Sites that implement schema markup properly tend to be sites with a technical SEO function. Sites with author markup tend to be sites with named authors worth marking up. Sites with fast pages tend to have engineers. Every one of those is a proxy for being a serious publisher, and serious publishers get cited more. Hold the publisher constant and the proxies stop predicting anything.

Look at the FAQ row twice, because it is the one that should sting. Plus 18.1% to minus 3.4% is not a small disagreement between two methods. It is the difference between a deliverable worth billing for and a deliverable worth nothing, and the only thing separating them is whether you controlled for the website.

I want to be careful here, because there is a bad reading of this. Nobody is saying FAQ blocks are harmful, or that schema markup is pointless, or that page speed does not matter. Schema still drives rich results. Fast pages still convert. The claim is narrower and sharper: none of those four is buying you AI citations, so none of them should be priced as if it is.

How to get cited by ChatGPT, according to 468,037 citations

The two content features that survive a within-website test are both about question matching: whether the page title matches what a person actually asked, and whether the strongest paragraph on the page matches it. Everything else on the checklist failed.

Notice what the surviving two have in common. Neither is a format. Both are a match between your words and a question somebody asked. That is not a new instruction, it is the oldest instruction in the trade, and the study's own recommendation to publishers is almost quaint: write for the questions people ask.

It also explains why the third-party numbers run higher, 25.7% and 23.6% against 15.5% and 14.1%. On a page you do not control, a listicle or a review roundup, title and paragraph match are doing even more of the work, which lines up with what we have seen in the source audits we run on AI vendor recommendations. The pages that get you named in an answer are often not yours, and the ones that are yours earn their place by answering the question in the first paragraph rather than by carrying the right markup.

There is a structural lesson in the ratio too. If the 14 page features together account for 2.8% of the variation, then a programme spending most of its hours on page-level content treatment is spending most of its hours on 2.8% of the problem. That is not nothing. It is also not a strategy.

Most of what ChatGPT reads never ranked in Google at all

Retrieval is the step where ChatGPT rewrites a question into its own sub-queries, runs them, and ranks what comes back, and it explains 32% of the variation in which pages get cited, against 2.8% for every page feature combined.

The single most useful number in the paper is this one: of the pages ChatGPT retrieves, 97% are absent from Google's top 50 for the same question. The engine is not reading the SERP you optimise for. It rewrites the question into sub-queries that often look nothing like the original, searches on those, and reads what it finds there.

Then there is the funnel. ChatGPT consults roughly 26 pages per question and cites about 3 of them, and when its own search puts a page first, that page gets cited 43.8% of the time. Being retrieved is not sufficient. Being retrieved first is close to half the battle, and being retrieved at all is the whole entry fee, which is what the authors mean when they write that every page ChatGPT cites is one its own search retrieved.

That reframes the work. If 97% of what the engine reads was never in Google's top 50, then rank tracking is a thin instrument for this, which is the same measurement problem we took apart in what rank tracking data can and cannot tell you, and the newer click data gap we covered in the AI Mode links that arrived without click reporting. Three of our own measurement complaints this year turn out to have had the same root: the retrieval layer is where the decisions happen and nobody gets to see it.

The honest caveat, which the authors state themselves: Perplexity could not be replayed, so the retrieval analysis is ChatGPT and Gemini. And these are associations drawn from observation, not a randomised trial. The controlled experiments are a separate and smaller part of the paper.

The model remembers brands your page cannot argue with

A meaningful share of what generative engines say about brands comes from model memory rather than from anything retrieved, and 46% of the brands ChatGPT names in these answers are supported by no cited page whatsoever.

This is the part the checklist was never going to touch. If almost half of brand mentions have no source page behind them, then a chunk of your AI visibility is a function of whether the model learned about you during training and from the wider web, which is earned over quarters through the kind of third-party presence we described in how credible presence gets built in the threads engines read. No amount of on-page work shortcuts it.

How to get cited by ChatGPT without buying a checklist

The work that survives this evidence is question coverage, third-party presence and retrieval diagnostics, in that order, and all three are cheaper than the template surgery they replace.

If you are shopping for generative engine optimization services this quarter, that is the question to put to every vendor on the list. Ask which of their deliverables was validated within a single website. The ones who have thought about it will know exactly what you mean, and the ones who have not will tell you about their FAQ schema.

What this should do to an LLM SEO budget

An LLM SEO budget built on this evidence spends less on page treatment and more on question coverage and third-party placement, and it reports retrieval as a condition it observes rather than a result it delivers.

The reallocation is not as dramatic as the 2.8% figure makes it sound, and I would resist anyone who uses this study to argue that on-page work is dead. Title and paragraph match are on-page work, they held up under the strictest test in the paper, and they are the cheapest thing on the list to fix. What dies is the premise that AI citations can be bought with a template change, which is a premise that was always a little too convenient for the people selling it.

There is a quieter implication for how pages get built. If what matters is whether a standalone paragraph answers a standalone question, then the structural discipline we laid out in the answer engine optimization pillar survives this study intact, while the decoration around it does not. Writing passages that make sense when read alone was always the defensible version of that advice. The FAQ block was the version you could sell as a deliverable.

And there is a budget question worth asking out loud, because the honest answer makes the category look smaller and better. If the content half explains 2.8%, and retrieval explains 32%, and retrieval is driven by question coverage plus whether the model has heard of you, then most of this discipline is editorial and reputational work wearing a technical costume. That is a harder thing to put on an invoice. It is also the thing that compounds, which is the argument we make for B2B programmes that run content and demand together rather than treating AI visibility as a separate workstream.

See where you are cited today

A free snapshot audit of your rankings and AI citations before we ever talk.