Buyers used to click through to your pricing page and squint at it themselves. Increasingly, an agent reads it for them, decides what it means, and reports back. If the agent gets stuck, it does not wait. It goes to a directory that already has the number and cites that instead.
Kevin Indig published the data on Growth Memo on July 13. It is the most specific look yet at what happens when AI agents, not just AI answer engines, try to complete a real B2B buying task on your site, and it is the clearest evidence yet that AI agent optimization is a distinct discipline from getting cited in a generated answer.
What the study measured
Indig ran 100 B2B products through three buyer-relevant tasks: pricing and features, integrations, and security and compliance. Each task ran five times per product, so the numbers reflect repeat behavior, not a single lucky or unlucky pass. For every run, he logged whether the agent answered from the vendor's own site, first-party, or fell back to a third-party source, and where that fallback came from. It is a research design closer to the technical audits we run than to a typical citation-tracking study: it measures whether the site actually functions for the task, not whether it theoretically could.
Where AI agent optimization breaks down
| TASK | FIRST-PARTY ANSWER RATE | FIRST-PARTY CITATION SHARE |
|---|---|---|
| Pricing & features | 79% | 84% |
| Integrations | 93% | 99% |
| Security & compliance | 92% | 99% |
Integrations and security questions are close to solved. Agents find the answer on the vendor's own site almost every time. Pricing is the outlier, and the gap is not small. When a vendor did not disclose real numbers, agents cited a third party in 45% of runs. Even when the vendor published a clean numeric price, agents still reached for a third-party source in 18% of runs. That second number is the one that should worry a marketing team: publishing a number is necessary, but on its own it is not sufficient.
Why pricing breaks first
Indig frames the shift plainly: sites built as showrooms, full of narrative and nurture sequences, do not serve an agent that just wants the fact. An agent wants the barcode, not the pitch. That framing lines up with what we've argued in the anatomy of an AI citation: extractable structure beats persuasive copy every time an engine, or now an agent, has to lift a fact under time pressure.
The variance problem
One well-known B2B SaaS pricing page scored 73 out of 100 on Indig's agent-readiness scale on one run, then 92.5 out of 100 on a second run in the same hour. Same page, same hour, 20-point swing. If your own measurement of agent readiness is a single spot check, you are measuring noise, not signal.
Where the 580 third-party citations in the study came from, by source type
Editorial and directory sources split the fallback almost evenly. Neither one is you. Every point of third-party share on your own pricing question is a point where a review site or a comparison blog gets to frame your price for the buyer, in their words, with their spin.
Agentic search changes the unit of measurement
This data lands at the same moment agentic AI adoption is compounding across B2B. Roughly 60% of companies now run agents in production, and three out of four are actively investing further, context Indig cites alongside the study. Salesforce has pointed to 20% of its own sales volume as agent-originated. None of that traffic shows up in a channel report the way organic or paid does, which is exactly why the reporting and analytics work around GEO has to expand to cover it, not just citation counts in a generated answer.
It also matters because of what's on the other side of that citation. Seer Interactive's June 2025 data put ChatGPT-referred visitor conversion at 15.9% and Perplexity-referred at 10.5%, against a 1.76% average for organic Google traffic. When a pricing question falls through to Capterra or G2 instead of your own site, you are not just losing a citation. You are losing one of the highest-converting touchpoints available in a buyer's research process, and handing it to a directory that monetizes the click either way.
Directories win this fallback for a structural reason, not a content-quality one. G2, Capterra, and Vendr publish pricing tiers in clean, consistent, machine-parseable tables across thousands of listings, the exact format an agent is built to extract with confidence. A vendor's own pricing page competing against that has to match the same clarity, not just publish a number somewhere on the page. AirOps' 2026 State of AI Search work, citing Kevin Indig, found only 30% of brands maintain consistent visibility across AI sessions at all, and inconsistent pricing pages are a large part of why: the agent that got a clean answer today may hit a redesigned calculator tomorrow and bail to the directory it already trusts.
The fix isn't guesswork, either. A KDD 2024 study out of Princeton, Georgia Tech, and IIT Delhi, one of the earlier pieces of hard research behind the whole GEO discipline, found that adding statistics to a page lifted AI visibility by 41%, adding direct quotations lifted it by 28%, and citing external sources lifted visibility by as much as 115% for pages that started out lower-ranked. Combined, these techniques delivered up to a 40% visibility improvement. None of that is agent-specific research, but it maps directly onto the pricing-page problem: a number alone is opacity's opposite, but a number backed by a source, a quote, or a comparison stat is what actually survives an agent's confidence check.
What to fix first in AI agent optimization
There's also a compounding relationship between this data and the placement debate the industry is having on the answer-engine side of GEO right now. As engines start experimenting with where a citation sits in a generated answer, agentic buying tasks make that placement question moot in a different way: an agent completing a task doesn't necessarily surface a citation to a human reader at all, it just acts on the fact it retrieved. That makes the underlying page even more important, and the visible citation even less of a complete picture of your actual influence on the purchase.
Demand for this kind of fix is already visible in search behavior, not just in agent logs. Ahrefs data puts monthly US search volume for 'generative engine optimization' at roughly 7,900, with a difficulty score of 59, evidence that the category of problem this study describes, being found and trusted by AI systems, has moved well past early-adopter curiosity into a search term buyers and practitioners are actively typing. The demand for agent-readiness fixes is not hypothetical. It's already showing up in keyword data before most enterprise sites have caught up.
Publish a real number, even a starting-at number, in plain HTML near the top of the page, not inside a calculator or a gated PDF. Then check that AI agent user agents can actually fetch the page: no soft-block, no aggressive rate limit, no login wall in front of the number. Run the check more than once. A single clean pass tells you almost nothing, given the variance Indig found in the same hour on the same page. Pricing is the one page on your site an agent visits on every single evaluation. Treat it like the highest-traffic page you own, because for agents, it already is. It's the same starting point we use on our GEO engagements, and it's what we led with on a recent cybersecurity B2B SaaS engagement where machine access, not content volume, was the first-quarter fix.
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Josh leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.