A 500-Run Study Shows Where Agents Abandon Your Site, and Where They Go Instead

We argued in Agents Read Your Pricing Page Before Any Human Does that the first reader of your commercial terms is a machine, and that most machines cannot parse the terms you publish. That was a direction with an anecdote behind it. Now there is a measurement. Kevin Indig ran 500 agent queries across 100 B2B products and logged where the agent got its answer, and the numbers say the failure is concentrated in exactly one place: pricing.

Here is the shape of it. Across topics like integrations and security, the agent found its answer on the vendor’s own site 92% to 93% of the time. On pricing and features, that first-party answer rate dropped to 79%. Pricing is where the agent leaves your site to find what it needs.

The Fallback Is Not Rare, It Is the Default for Price

The number that reframes the problem is 77%. Of every third-party citation the study recorded, 77% came from a pricing query. When an agent cites G2, Vendr, or a review roundup instead of you, it is almost always because it was trying to answer a question about price and could not get the answer from your page.

Then comes the mechanism. Access errors (a page the agent could not load or parse) showed up in only 7% of runs. Small number. But those 7% of runs behaved completely differently from the rest. In a clean run, third-party fallback sat around 17%. In a run with an access error, fallback jumped to 77%. A rare technical failure carried an outsized consequence. It flipped the agent from reading you to reading about you, four times out of five.

That is the cost line we could only assert before. A pricing page that renders for a browser and returns nothing to a parser does not simply go unread. It reroutes the agent to whoever else is willing to state your price.

Silence Does Not Buy Discretion

The instinct behind an unreadable pricing page is often deliberate. Keep the number off the machine-readable surface, force the conversation into a sales call, control the framing. The study measures what that instinct actually produces.

In 45% of runs, the agent cited a third-party source even when the vendor’s own page stayed silent on the question. Withholding the number did not stop the agent from producing one. It changed who supplied it. The agent went to a review site, a procurement marketplace, or a competitor comparison, and returned with a figure you never wrote, attached to framing you never approved.

Indig’s phrase for this is the one worth keeping: AI agents turn websites from showrooms into barcodes. A showroom is a place you curate, where presentation carries meaning and a visitor lingers on what you chose to show. A barcode is a lookup key. The agent scans, resolves your product to a set of extractable facts, and if your surface cannot yield those facts, it scans a different barcode that will. Opacity leaves your product on the shelf. It just means the price attached to it gets read off someone else’s label.

Why Pricing Specifically

Integrations and security clear 92% because those answers tend to live in documentation, and documentation is written to be read. It is text, structured, stable, addressable. Pricing is the field most likely to sit inside a rendered widget: a plan toggle, a slider, a calculator, a number that only exists after a script runs. The agent asking about your SOC 2 posture finds a page written for reading. The agent asking about your price finds a page written for clicking.

That is why the same buyer’s agent succeeds on one question and fails on the next, on the same site. The security answer was authored as a document. The pricing answer was authored as an interaction. Only one of those survives a fetch.

The consequence compounds because pricing is also the question that decides the shortlist. An agent that cannot rank you on price cannot place you in the comparison, and a comparison it fills from G2 is a comparison where your number, your framing, and your caveats are all someone else’s to write. The 45% is an authorship problem. More content will not resolve it, because the fact circulates regardless. The only variable you control is whether the source is you.

What One Study Can and Cannot Tell You

This is one dataset, proprietary, from a single credible analyst. Kevin Indig has spent his career on this surface, formerly at Shopify and G2, so the source knows the terrain. Treat these as one strong measurement, not a settled census. The exact percentages will move with the models tested, the products sampled, and the quarter.

The direction, though, matches what the earlier post predicted, and the mechanism is now legible. Opacity was supposed to protect the number. Instead it relocates the number to a source you do not control, and it does so most reliably on the one query that decides whether you make the list.

Do This Now

Run the pricing query against yourself the way a buyer’s agent would. Ask a general-purpose agent for your product’s price, plans, and overage terms, and read where it gets the answer. If it cites a review site, a marketplace, or a competitor comparison instead of your own domain, you have located your barcode, and it is on someone else’s shelf.

Then close the access error first, because that is the 7% that becomes 77%. Fetch your pricing page over plain HTTP and confirm the numbers are in the response, not locked behind a script. Publish price, unit, included limits, and overage rate as structured text the agent can read on the first pass. You will not stop the fact from circulating. You decide whether it circulates in your words or in G2’s.

---

This analysis synthesizes Where AI agents get stuck on your site (Kevin Indig, growth-memo, July 2026).

Victorino Group helps teams make their commercial surface legible to the agents that read it first. Let’s talk.