The agentic revolution is upon us, and our AI assistants really are doing most of the heavy lifting of our daily research.
At this point, we are handling the keys over to algorithms to find the best solutions (and deals) for us, and business are adapting to it fast.
In fact, research from G2 shows that more than 70% of businesses now rely on AI agents for software research. But the key question here isn’t how many people or businesses use AI agents to search pricing, it’s how do those AI agents look for and perceive pricing?
Companies have spent decades perfecting corporate pricing pages. But those pages were directed to humans. Now AI agents are the new audience.
And as they take over the research phase, the traditional pricing page is quickly losing its grip on how a product's price is actually communicated to the world.
together with Hard Skill Exchange:
If agents now do the buyer’s research, your GTM can’t run on fixed sequences and static playbooks. It has to learn from every account, interaction and outcome, and adjust what happens next on its own.
That’s the idea behind Recursive GTM, a free virtual session on October 15:
▫️ How self-improving systems are being built across marketing, sales and customer success
▫️ Live demos and 30-minute practice sessions with leaders from Dialpad, ZoomInfo, 1mind, Vivun and Aligned
▫️ How to compare AI-native GTM operating models and find the one that fits your org
Table of Contents
1. How Do AI Engines Understand Pricing?
2. What Crawlers See
3. The Three Domains Pricing Your Product
4. Documentation Outranks the Pricing Page
5. When the Agent Starts Buying
1. How Do AI Engines Understand Pricing?
So, how do AI engines answer a pricing query now?
Kyle Poyar and Nikolas Laskaris at Growth Unhinged ran a prompt against the Cloud 100 with the objective of evaluating the pricing of companies using ChatGPT, Gemini, Copilot, Perplexity, and both of Google's surfaces. They’ve done this 7675 times. The prompt was:
Evaluate Company X on pricing.Although the results revealed that the companies’ own pricing pages appeared in 46% of the answers, they appeared first in only 12%.
This creates an unpredictable mechanism where customers researching the same thing may come to completely different conclusions.
According to the study, ChatGPT has the most influence and leads the company’s own page by 38% and that’s because OpenAI’s model defaults to company-owned pricing pages, as opposed to Google that - according to the author - buries those.
The price a buyer reads therefore depends largely on which assistant they happened to open. The buyer is never told whether the figure came from the vendor or from someone estimating on its behalf, and nothing on the vendor's own site has any say in which assistant was chosen.
2. What Crawlers See
A web crawler is an automated bot sent by search engines and AI tools to scan, read, and index the content of web pages across the internet.
Ultimately, a web page cannot be cited if the machine reading it encounters an empty shell, which happens to be the condition a meaningful share of the Cloud 100 is in.
The authors report that, out of the 77 public pricing pages in the sample, 57 were fully readable to crawlers, while 10 hid at least 40% of their body content.
Here’s the thing. A webpage may appear normal to a human eye while remaining entirely incomprehensible to a machine when critical information hides inside tabs and interactive elements.
That’s because information can load through JavaScript or hide behind restricted paths in robots.txt, rendering it invisible to raw HTML crawlers.
The Fix Is a Deployment
Profound solved a similar issue by moving the content to server-side rendering which helped the pricing page being more visible to raw HTML crawler and rose citation bot traffic rose 13% week over week.
What complicates the obvious conclusion is that publishing exact numbers is not what earns the citation. Companies such as Personio, Pendo, Carta and Gong display plans without a single price point, even though all 4 sit above the median concluded by the study.
The difference is that readability determines how many gaps the AI agent is compelled to fill, whereas transparency determines how much detail a company is willing to reveal.
3. The Three Domains Pricing Your Product
Perhaps the most interesting finding of that study is that the same 3 external sites kept surfacing across the full dataset, and none of them belong to the companies being priced.
That’s because when the engines skip the vendor’s page, they go looking elsewhere. And apparently, they have their own biases.
Vendr, which sells procurement and negotiation data, appeared in 18.7% of runs. Reddit appeared in 18.6% and G2 in 15.9%.
The truth is that vendors themselves have vastly different levels of control over these spaces. While a company can edit its own G2 profile, it has zero control over anonymous Reddit threads or procurement logs, severely limiting vendor influence over how pricing is presented.
One vertical AI unicorn in the sample publishes no pricing whatsoever. Across 76 responses pulled mainly from Reddit, engines quoted a monthly range from $100 to $2,400, representing a 24x difference for the exact same product.
Now reddit is interesting because they did sign a data licesning deal with Google in early 2024, and OpenAI followed a few months later. The two biggest answer engines pay for structured access to Reddit threads, so a 2-year-old complaint about a renewal quote can carry more weight than the pricing page it complains about.
4. Documentation Outranks the Pricing Page
The best result within the Cloud 100 came from a page most pricing teams never touch.
Plaid had its own page cited first 42% of the time. Its billing documentation was cited in 70% of answers about its pricing, ahead of the pricing page itself at 64% and FAQs at 50%.
That fits what the crawler data would predict. Developer docs are usually static, server-rendered HTML, which every crawler can read.
Marketing pricing pages are where the JavaScript widgets and toggles live. Docs are also written for someone who has to implement the billing and will get blamed when the invoice is wrong, so they define edge cases and units that a pricing page glosses over.
Fireworks AI’s docs expose an llms.txt index and a copy-for-LLM button. It also publishes scheduled price changes side by side with their dates, so an engine knows which number is current.
5. When the Agent Starts Buying
Everything so far assumes a person reads the answer and decides, but that’s already starting to change.
In a study of AI-to-AI negotiations, researchers matched 9 models as buyers and sellers. Swapping in a stronger buyer moved the final price by 2.6%. Swapping in a stronger seller moved it by 14.9%. Buyer agents also broke hard budget limits and gave away their own ceilings despite being told not to.
Despite these changes, the checkout layer hasn’t caught up.
OpenAI shut down Instant Checkout in March 2026 after about 6 months, with Walmart reporting in-chat conversion 3x lower than click-out. Buyers researched inside the assistant and bought somewhere else.
That leaves one number nobody tracks. Companies measure citation share and sentiment, but not the gap between the price they publish and the price the engines report.
Let’s call it price variance.
For that unicorn it’s 24x. For Plaid it’s close to zero. For most software companies it’s unknown, because nobody has run the prompt and logged the answers against the page.
It takes one prompt, a handful of engines and a week of logging. The price tag is still on the product, but the engines decide what it says.
price variance is the metric i'd actually steal from this. every team tracks citation share now, but the gap between the price you publish and the price the engines report is the number that costs you deals. the plaid detail is the funniest part: boring static billing docs outperforming the pricing page your marketing team spent a quarter on. all because crawlers can't click tabs.
Price variance is a useful number, but I'm not sure the 24x figure measures what it's being used to show. For a product with no public pricing, a $100 to $2,400 spread pulled from Reddit could just be a small team on one seat and a mid-size org on fifty, both quoting accurately what they pay. Unless the logging normalises for plan, seats and contract size, you can't tell engine error apart from genuinely different deals, and that changes whether the fix is publishing a price or publishing a pricing model.