Put your prices on the page

Ask any agency what to do about AI visibility and you will hear about structure. Headings. Schema. Clean formatting. Bullet points. Tables. Make the content scannable, make it parseable, make it structured.

In May 2026 three researchers set out to test that properly, and the result is the most uncomfortable finding in this field for the people selling it.

The design is the strongest anyone has attempted here. 252,000 trials across six language models. 1,440 scenarios built from a hundred anonymised product pages spanning fifty categories. Every comparison is a pair of documents differing on exactly one factor — brand names stripped out to kill familiarity bias, length held to within 5%, list positions counterbalanced so order effects cancel. Analysed with logistic mixed-effects models controlling for position.

Eighteen factors tested. Four came out as gatekeepers, significant across all six models:

Topical relevance                              odds ratio  >>10,000
List position                                              1,795 – >>10,000
Recent vs old timestamp                                     14.4 – >>10,000
Price disclosure — stated vs not stated                      6.26 – >>10,000

Three of those four are unsurprising. Being about the subject matters. Being first matters. Not being seven years out of date matters.

The fourth is the one nobody talks about. Whether the document says what something costs sits in the top tier, alongside relevance and position. We have not found a single published GEO guide that recommends it.

Now the other end of the same table. In the weakest tier — significant in three models or fewer:

Structured vs dense formatting                 odds ratio  0.79 – 1.68

Look at the lower bound. It is below 1. In some models, structuring the content made citation slightly less likely. The single most-recommended tactic in this entire industry produced an effect range that straddles zero.

The caveats are large and you should weigh them. This is an arXiv preprint and has not been peer reviewed. The corpus was generated and rewritten by GPT-4o and then judged by language models — three authors hand-checked a stratified sample of 300 scenarios, which is diligence but not independence. Brands were anonymised, so the study says nothing about how price disclosure competes against a famous name in the wild. And the authors are candid that a pairwise design is its own ceiling: production systems retrieve five to ten pages at once, where these factors would interact in ways a two-document comparison cannot show.

So this is one unreplicated result. Treat it as a hypothesis worth testing rather than a law.

Except it is not quite alone. Two independent lines of evidence point the same way.

When Kurt Fischman analysed 730 citations across 1,006 pages and found schema markup had no effect on citation once ranking was controlled for, he found one exception: pages carrying Product or Review schema with populated concrete attribute fields — real prices, real specifications — were cited substantially more. The markup was not doing the work. The facts inside it were.

And when Advanced Web Ranking hand-coded 112 passages to compare cited against uncited, they recorded a finding worth memorising: zero uncited passages contained novel claims or hard numbers. Not few. Zero.

Three different methods, one direction. Specificity gets quoted. Presentation doesn’t.

There is a plain mechanical reason why this should be true, and it has nothing to do with optimisation. A model constructing an answer to “how much does X cost in Y” needs a number. If your page says “contact us for a quote”, there is nothing to lift — you have written a page that cannot be used to answer the question your customer asked. A competitor who wrote “$400–$900 for a typical three-bedroom, same-week availability” has handed the model a sentence, and the model will use the sentence it has.

This is also the item clients resist most, and the objection is always the same: our pricing depends on the job. It usually does. But “it depends” and “we won’t say anything” are different positions, and a range with the variables named — $400 to $900 depending on roof pitch and access — is both true and quotable. The businesses that refuse to publish any number are not protecting their margin. They are removing themselves from the set of answers a model is able to give.