By Andrew Pierce, Founder of TheSaaSDir

AI Agents on Your Signup Form: What 800 Submissions in Ten Days Showed Us

About a quarter of the people filling in our submit form are now AI browser agents. How we spotted them, what they do at the payment step, and what we changed.

About a quarter of the submissions to our directories now come from AI browser agents, not people. Between 7 and 16 November 2026, 221 of roughly 800 submissions came from one Linux desktop Chrome build with no referrer, a pattern that barely existed before late September. Those agents paid at about the same rate as people on a Mac: 2.7% against 2.5%.

Data pulled 16 November 2026 from our server request logs, submission records and Stripe. Spam is excluded.

TL;DR: If you run a SaaS with a signup or checkout form, some of your "users" are already agents acting for a person. In our data they fill forms competently, they default to the free option, and they stop at a payment page unless their operator says yes. Bulk tools submit a lot and almost never pay; individual agents pay like humans. The changes that mattered for us were a plain-text page telling agents what to ask their operator, a saved state that survives a stopped checkout, and form guards that stop crude scripts without blocking real agents.

Where the numbers come from

We run eight product directories. Founders submit a product on a one-page form with no account. We matched every submission between 7 and 16 November 2026 to the request that sent it, which gave us the browser and operating system for each one.

Browser and OS at submit Share of submissions Paid Paid share
Mac 39% 8 of 324 2.5%
Linux (desktop Chrome, no referrer) 27% 6 of 221 2.7%
Windows 27% 0 of 225 0%
Headless Chrome 4% n/a n/a

Two caveats. First, "Linux desktop Chrome with no referrer" is our inference for "cloud browser agent", not a label anyone gave us. Almost nobody browses our site from desktop Linux, and almost every one of these requests used the same Chrome build. Second, the payment counts are small. Six against eight is not a difference; the point is that agents are not worse.

How to tell agent traffic from people

The pattern appeared quickly. Our analytics show 5 direct Linux visitors to the submit page in July 2026 and 261 in October. Nothing about our site changed in between; the agents did.

Three signals did most of the work for us:

  1. The user agent at the moment of submission. We now store it on every submission. Your analytics script may never see an agent, because many agents load the page without running third-party scripts. Our analytics counted about 30 visitors a day on the submit page while the server logged about 3,000 requests a day to it, most of them crawlers and link checkers.
  2. The email domain. Several agent products give the agent its own inbox. Since mid-October, 252 submissions came from Grok Bot's built-in mail domain, from 118 different inboxes for 146 different websites. Another 49 came from AgentMail inboxes. Every status email you send to those addresses is read by an agent first.
  3. The timing. Three of the nine people who paid for orders in that window went straight to the submit page and paid within a couple of minutes, with no earlier visit. That is what an agent holding a link looks like.

What agents do at the payment step

Agents are cautious with money, and they are told to be. The most-installed directory-submission skill for Claude Code instructs agents to use free tiers only, to stop at payment pages, and to hand the step back to a human. In a test where we gave five research agents five different founder briefs, all five defaulted to free options and stopped at logins, CAPTCHAs and payment pages.

That matches what we saw on our own form:

  • Free is the default. Agent submissions mostly choose the free plan. Between 62% and 73% of their free listings went live, because the agent did what the free plan asked (in our case, adding a small badge to the product's site). Agents with access to the product's code repository did this without help.
  • Paying needs a person. We have not seen an agent buy a listing on its own initiative, and when we researched how agents spend money more widely, we found none doing so either. An agent spends from an account a person already funded, after it has shown that person a price.
  • Stopping is not abandoning. When an agent halts at checkout, the order isn't dead. The person behind it may pay hours later, if the agent passes something useful back to them.

What we changed, and why

1. A page that tells agents what to ask. We publish a plain-text file at /llms.txt on every site. Its main job is not search. It tells a visiting agent, in fixed words, the question to put to its operator before choosing a plan, and what to report back afterwards. An agent will not volunteer "this costs $19, do you want it?" unless something prompts it.

2. A saved state before the money. A submission is saved before checkout opens, and the agent gets a link to finish payment later. If the agent is not allowed to pay, the operator still ends up with a sentence like "your listing is saved but not live; to finish, pay here", plus the link.

3. Guards that stop scripts, not agents. A hidden form field and a minimum time between page load and submit catch crude scripts. Real browser agents fill only the visible fields and are slower than a script, so they pass. We tell them both rules up front in /llms.txt, including what to do if the form says to wait.

Bulk tools are not customers

The submitters who paid least were the busiest. Email addresses with five or more submissions sent 328 submissions in the period and paid once. Addresses with two to four submissions paid most often, at 4.3%. Volume tells you a submission service is working through a list. It doesn't tell you anyone wants to buy.

If your signup numbers jumped this autumn and revenue did not, check how much of the jump comes from a few addresses or one browser build before you celebrate or panic.

What this means for your own signup form

  • Log the user agent on signups and orders on the server. Client-side analytics undercount agents.
  • Put your prices and plan differences somewhere an agent can read without running JavaScript.
  • Assume an agent will stop at your payment page. Make sure what it saved survives, and give it a link to hand back.
  • Write the first line of every transactional email so it still makes sense when an agent quotes it alone.
  • Judge agent traffic by what it pays, not by how much of it there is.

Questions Founders Ask

How do I know if AI agents are using my signup form?

Store the user agent and email domain with each signup on your server. Look for a single desktop Linux Chrome build with no referrer, headless browsers, and inboxes from agent mail providers. In our data these patterns went from almost nothing in July 2026 to about a quarter of submissions in November.

Do AI agents pay for software?

Yes, at about the same rate as people. In our data, agent submissions paid at 2.7%, against 2.5% for Mac users. We have not seen an agent pay without a person's go-ahead: agents spend from an account a person funded, after showing that person the price.

Should I block AI agents from my signup form?

We don't. Agents acting for a real founder behave like careful users and pay at a similar rate. A hidden honeypot field and a minimum time before submit stopped crude scripts on our form without blocking browser agents. Bulk tools are worth limiting with per-email and per-site caps, because they submit a lot and rarely pay.

What is llms.txt used for on a SaaS site?

On our sites, llms.txt is mainly instructions for agents that are already on the page: which fields to fill, what each plan costs, which question to put to the person they work for, and what to report back. We do not yet have evidence that it changes how often AI search engines cite us.

List your SaaS — free dofollow backlink

Submit yours →