The Meeting That Ends At “Let’s Clean Up The Page For AI”
Every conversation about AI optimization seems to start with heading structure and how granular the bullets should be.
We burned an hour on exactly that. I doubt we’re the only building where that meeting happened.
Afterwards I went and did only the formatting, and then spent three weeks watching nothing change.
Which is the useful part, actually: if a tidy page still doesn’t get cited, the contest is being decided somewhere outside the tidying.
So where is it decided? A study published in May 2026 pulled the factors apart one at a time, running 252,000 rounds of two pages on the same topic competing to be cited first.
What Do You Actually Have To Put On A Page To Get Cited By AI?
A price, a spec, and a recent update date.
Across 252,000 head-to-heads between two pages on the same topic, a page showing a price was cited first at least 6.26x as often as one without.
A page with a recent date beat an older-dated one by at least 14.4x (Vishwakarma et al., six models, 252,000 trials, published 25 May 2026).
Whether the prose was structured or not barely moved anything: 0.79 to 1.68x.
An odds ratio here is just “how many times more likely one page is to be picked”. One means a coin flip; six means one side gets chosen six times as often.
The design is what makes this different from what we’ve covered before. Most AI-citation research is observational — collect the pages that got cited, then look for patterns.
This one hands the model exactly two candidate pages, swaps a single factor out of eighteen, and forces a choice.
Three researchers at Sprinklr ran it across Gemini-2.5-Flash, GPT-5-Nano, GPT-5-Mini, GPT-5.2, Claude-3.5-Sonnet, and Kimi-K2-Thinking.
They also flipped the order of the two candidates to cancel out the effect of position itself.
So this reads as “adding this element moves your win rate by this much”, not “pages like this tend to get cited”.
Four Gates You Have To Clear Before Anything Else Counts
Four factors produced a large effect in the same direction across all six models:
- whether the page’s topic matches the question
- whether a price is present
- whether the date is recent
- where the page sat in the presented order
The authors call these gates: fail any one of them and the chance of being cited effectively disappears.
That settles the order of work. Match the topic, put a price on the page, refresh the date. The formatting pass comes after.
The fourth gate is out of your hands, though — the order of candidates belongs to whatever is doing the retrieval. Three of the four are yours.
So which of those three is the easiest place to start?
Refreshing A Date Beats Hiding One
A recent date beats an old date by at least 14.4x.
Remove the date entirely, though, and you only beat an old date by 1.31 to 2.32x. Adding freshness and hiding staleness are not mirror images of each other.
Which means quietly stripping old timestamps buys you very little.
Peeling the sticker off the milk doesn’t make the milk younger.
Go update the content, then show the new date. That’s the page standing on the 14.4x side of the comparison.
Layout Didn’t Matter. What The Text Says Did.
Structured prose versus a wall of text came out at 0.79 to 1.68x, and the authors state plainly that formatting choice had no effect.
Does that mean how you write doesn’t matter? No.
Pages that show their evidence won by at least 2.09x. Pages written assertively won by at least 2.67x. Pages with a neutral, non-promotional tone won by at least 1.31x.
None of those describe what I spent three weeks doing.
This lines up neatly with two pieces we’ve run before — the study saying the trick is adding numbers and sources and the one where three signals lifted citations by up to 40%.
Those measured the lift on a single page after a rewrite. This one measures who wins when two pages fight. Different units.
The freshness finding sits in the same relationship. Most cited pages turn out to be from the last year and 75% of cited pages were updated within twelve months both counted pages after the fact.
Here the date is the only thing that changes, and the “no date at all” case gets measured too. The observed tendency has been pulled apart into factors.
One tension worth naming: the SOAR framework from the piece on press releases getting cited in about eight hours includes structure among the signals models use. In this controlled comparison, structure barely moves the outcome.
I’d be slow to conclude that structure is useless. An observational framework and a one-factor-at-a-time experiment are cutting at different resolutions.
Read The Caveats Before You Carry These Numbers Into Your Own Deck
Sensitivity varies a lot by model. Of the eighteen factors, the share that came out significant was 83% for Kimi-K2-Thinking, 50% for Claude-3.5-Sonnet, and 33% for Gemini-2.5-Flash.
The four gates held across all six. Everything else says you can’t check one assistant and call it done.
There’s more to hold onto:
- only two candidate pages are in play, while real AI search pulls five to ten or more (the authors say so)
- brand names were anonymized, so any advantage a famous brand gets regardless of content isn’t in these numbers (also stated)
- the material was generated with GPT-4o, and humans reviewed 300 scenarios, or 21% of the total (also stated)
- what’s measured is which page the first citation marker pointed at — not exposure, traffic, or revenue
- the three authors are at Sprinklr, which sells enterprise marketing software
Two of those deserve extra weight.
First, the material is 100 English product blog posts across 50 categories. If you’re measuring in another language, carrying these multipliers straight over is premature.
Second, the price effect assumes a page that can carry a price at all. For B2B products with private pricing, it doesn’t transfer as-is. The question becomes how much of your pricing logic and quoting assumptions you can put in writing.
Open Your Biggest Page And Check Topic, Price, And Date
There are two places to put your hands: one page, and the models’ answers.
Open the page that gets the most traffic and check three things. Does the topic match the question a reader would ask, is there a price or at least the logic behind your pricing, and is the update date stale?
The formatting pass still gets to happen. It just happens after those three.
Then confirm the change on the answer side. Because the effect runs anywhere from 33% to 83% depending on the model, measure repeatedly across several of them and watch how often your name appears.
If you want one thing to read next, take the practical checklist at the end of the paper: clear the gates, fill in missing price, date, and question terms, add spec tables and comparisons, and leave formatting for last.
So go look at one page — just the price line and the update date — and see what it says.
Sources
- Rahul Vishwakarma, Shushant Kumar, Ratnesh Jamidar (Sprinklr), “What Gets Cited: Competitive GEO in AI Answer Engines”, arXiv:2605.25517, 2026-05-25, https://arxiv.org/abs/2605.25517
