The research says hiding your prices keeps you out of AI answers

Almost every local business we audit hides its prices, and almost always on purpose. Clinics treat "book a consultation" as the price conversation. Tour operators quote per group after a phone call. Hotels keep rates inside the booking engine where they can move them. The reasoning is the same everywhere: a number on a page is a number a competitor can undercut and a customer can anchor on.

A peer-reviewed experiment presented at SIGIR this July puts a cost on that decision, and the number is larger than we expected.

Does putting prices on your website help you show up in AI answers?

Once an assistant is comparing you against someone else, yes, substantially.

The paper is What Gets Cited: Competitive GEO in AI Answer Engines by Rahul Vishwakarma, Shushant Kumar and Ratnesh Jamidar, posted to arXiv in May 2026 and published in the Proceedings of the 49th International ACM SIGIR Conference, held 20 to 24 July 2026 in Melbourne. It is five pages and the full text is open. All three authors work at Sprinklr, and the paper says the resulting checklist was piloted internally there. Worth knowing before you read the practitioner section. It does not change what the table reports.

The design is a two-document retrieval-augmented generation testbed. For each trial the model receives exactly two candidate sources that are identical except for one attribute, and the researchers record which source the first citation marker points at. They tested 18 content factors across six models: Gemini 2.5 Flash, Claude 3.5 Sonnet, Kimi K2 Thinking, GPT-5-Nano, GPT-5-Mini and GPT-5.2. Brand names, product models and publisher names were replaced with fictional aliases so the model could not fall back on recognition, and source order was counterbalanced to separate content effects from position bias. That produced 1,440 base scenarios, three query paraphrases each for 4,320 scenario-query instances, five runs apiece, and 252,000 trials in total.

The abstract states the headline finding: "topical relevance and list position are the biggest drivers of being cited first. Including explicit price information and a recent timestamp also helps consistently. Completeness and trust cues add smaller gains, while formatting-only edits have little impact."

The numbers

Table 2 reports odds ratios per model, where an odds ratio above 1 favours the variant carrying the attribute. Here are the three rows that matter for this argument.

Two of the six models put price beyond the measurement ceiling. The lowest reading on any model is 6.26. The paper groups price with topic mismatch, stale timestamps and list position as one of four "gatekeepers" that "were unanimous across all six models with large effects," adding that "failing on any one can eliminate citation odds regardless of other content strengths."

We will flag one inconsistency rather than paper over it, because this site's whole position is that you should be able to check us. The paper characterises those gatekeepers as having odds ratios above 100, and the price row reports 7.82 and 6.26 on the two smallest GPT models. Those are still large effects and still statistically significant, and they are not above 100. Read the gatekeeper label as the authors' summary across models and the table as the evidence.

The formatting row is the more useful half

Look at the bottom row again. Structured versus dense formatting scored between 0.78 and 1.68, and three of the six models landed below 1, meaning the structured version lost. The paper's own conclusion is blunt: "Formatting choices (Content Structure, Scattered Information) had no impact."

Restructuring content is a large share of what the GEO industry sells. Break the wall of text into sections. Add hierarchy. Put a summary box at the top. It is visible work, it photographs well in a before-and-after slide, and in a controlled experiment across six models it is indistinguishable from leaving the page alone.

Meanwhile, the change that moved the needle costs one sentence. The paper's practitioner guidance says as much: "Quick wins are usually editorial: add concrete pricing, update timestamps, and close keyword gaps."

That contrast is the whole article. We have argued before that most of what is sold as GEO does not survive scrutiny. This is the rare case where a controlled study points at something cheap that most businesses are deliberately not doing.

What the study does not show

This is where most coverage of a paper like this goes wrong, so we will be explicit.

The testbed injects exactly two documents. The authors say what that means: "Production RAG often retrieves five to ten or more pages, so real citation pools are larger than our testbed," and "the estimates are therefore pairwise preferences over a controlled slate, not full multi-document competition."

So this measures what wins a comparison you have already entered. It says nothing about entering it. The paper is direct on this point too: "if the brand is absent from citations entirely, the bottleneck is retrieval." Publishing a price on a page that ChatGPT's crawler cannot fetch changes nothing. Crawler access and rendering come first, every time, and the order is in our AI crawlers and robots.txt guide.

Second caveat. The seed corpus was 100 product review blog posts across 50 B2C product categories such as consumer technology, home goods and fitness equipment, capturing 2026 market prices and specifications. No local service pages were tested. A dental implant quote is not a robot vacuum, and we would treat this as a strong prior for service businesses rather than a measured result for them.

Third. Brands were anonymised on purpose, so the study deliberately removes the effect of being a name a model has seen before. The authors note that production systems "may still favor trusted domains or strong brands." Real answers carry that effect and this experiment does not.

Anyone citing this paper as proof that publishing prices gets you into ChatGPT has read the abstract and skipped section five.

What to publish when you cannot publish a price

The factor is defined in the paper's taxonomy as "Product price information absent." The test is presence, not precision. That leaves more room than most owners assume.

A clinic can publish a band per procedure with what moves it: the consultation fee, the range for a single implant, and the two or three variables that decide where in the range a case lands. A tour operator can publish the per-person rate at minimum group size, what is included, and what a private booking adds. A hotel can publish a from-rate by season on the page itself instead of leaving the number locked inside the booking widget, which is also a rendering problem since a widget is usually invisible to the crawler anyway.

What does not count is the sentence most of these sites currently run. "Contact us for pricing" is the absence of price information wearing a call to action.

Pair it with the other cheap gatekeeper from the same table. Recent versus old timestamps produced odds ratios from 14.4 up past the ceiling, comparing content dated 2026 against 2019. A dated review line on a pricing page, updated when the prices change, does two jobs at once.

Our guide to schema for AI search lists price among the attributes worth marking up, and the tour operator piece puts total price in the set of filters that decide a match. Both asserted it. This is the evidence underneath them, with its limits attached.

Go and look at your services page. If the answer to "what does this cost" is not on it in any form, you are failing the cheapest of the four gatekeepers, and no amount of restructuring will cover for it.