What AI agents do to your conversion funnel
The conversion funnel assumes a person visits your site and does each step themselves. When an AI agent does part of the shopping, some of those steps move somewhere you can't see. This post explains what changes at each stage, why your reports may start to mislead you, and what to measure instead.
The funnel you report on today
Most e-commerce teams watch the same chain every week: sessions, product page views, add to cart, checkout started, order placed. Each step is a share of the one before it, and each drop is somebody's job to fix.
That model rests on one assumption: the person is on your site, in a browser running your analytics tag, for every step. Agent-assisted shopping breaks that assumption at the top, in the middle and at the end.
Person shops alone
- Searches, lands on your site
- Browses several product pages
- Compares tabs and reviews
- Adds to cart, checks out
Person shops with an agent
- Asks an assistant, which reads your pages
- Gets a shortlist with your product, or without it
- Lets the agent compare and build the cart
- Steps in to pay
What changes at each stage
Discovery happens inside the assistant
A shopper asks ChatGPT, Claude or Perplexity "what's a good waterproof jacket under $200?" The assistant fetches pages, reads them and answers. Your store is either in that answer or it isn't, and you never see the question. There's no impression, no ranking report and no session unless the person clicks. How AI assistants decide which products to recommend covers what decides it.
Fewer page views, each one doing more work
An assistant fetching your page to answer a question usually reads a page or two, not twenty. It doesn't run your tag, so the visit doesn't appear in your analytics at all. The browsing that used to fill the top of your funnel happens as a handful of server requests.
The agent does the comparing
The long middle of the funnel, where a person opens tabs, reads reviews and checks delivery, is exactly what agents are good at. If your price, stock or returns policy is hard to read, the comparison happens without you. The page doesn't lose a session; it loses the shortlist.
The handoff comes at payment
Browser agents can search, choose a variant and build a cart. Most then hand back to the person to sign in, confirm and pay. Agent payment methods aim to change that, but they're early. See payments for AI agents. Either way, the person often arrives late in the journey, straight to a full cart or a checkout, with very little browsing behind them.
- Question Asked inside an assistant you can't see
- Reading A few server requests, no analytics tag
- Comparison Done by the agent, across stores
- Handoff The person steps in, often at the cart or payment
Why your numbers may start to mislead
None of this shows up as a clean new line in your reports. It shows up as familiar numbers moving for unfamiliar reasons:
- Sessions can fall while demand doesn't. If more research happens inside assistants, fewer people browse. That looks like lost interest when it may just be research moving elsewhere.
- Conversion rate can rise for the wrong reason. Visitors who arrive after an agent has done the comparing tend to be further along. Fewer, later sessions can make a conversion rate look better without anything on the site improving.
- Some sales arrive with no label. People who copy a link from an assistant app can show up as direct traffic. Your AI referral number is a floor, not a total.
- Failures don't show up at all. An agent blocked by bot protection, or stuck behind a cookie banner, never ran your tag. Your funnel can't drop at a step it never saw.
Be careful with comparisons: a change in conversion rate or traffic mix isn't proof that agents are involved. Check what else changed (campaigns, prices, releases) before drawing conclusions, and look at the measures below together.
What to measure instead
Keep the funnel you have. Add measures that see the parts it can't:
| Measure | What it tells you | Where it comes from |
|---|---|---|
| Agent visits by kind | Which assistants, AI search crawlers and browser agents read which pages | Server or CDN logs |
| How far agents get | Product page, cart, checkout, order: the agent's own funnel | Logs, grouped into visits |
| Agents turned away | Blocks, CAPTCHAs and errors at each step | Status codes in your logs |
| AI referral visits and orders | People who arrive from an assistant's answer, and what they buy | Your analytics and order data |
| Share of answers | Whether assistants name you when shoppers ask | Asking assistants your customers' questions, regularly |
| Journey pass rate | Whether an agent can actually finish the job | Running AI agents through your journeys |
Our guides go deeper on each: measuring AI agent traffic from logs, tracking AI referral traffic in your analytics, tracking how assistants describe your brand, and agent readiness KPIs for putting it on one page each month.
- Before the visit Share of answers: are you on the shortlist?
- During the visit Agent visits and how far they get, from logs
- When it breaks Agents turned away, and journey pass rates
- After the handoff AI referral visits and orders
In the Ghost Agent Lab app, Agent traffic reads your server or CDN logs, and its Journeys view groups each assistant's and agent's requests into visits and follows them through product page, cart, checkout and order confirmation, with the pages where they give up. Ghost Agents run real AI agents through your journeys, stop at payment and replay each step. AI visibility, in beta, shows how Claude answers your shoppers' questions and whether it names you.
Fix the funnel for both audiences
The good news is that most fixes help people and agents at once. Work through the funnel stage by stage:
- Discovery: let assistants and AI search in. Check robots.txt and bot protection, so assistants can read your pages when someone asks. See agent traffic isn't bot traffic.
- Comparison: put prices, stock and policies in the page HTML and in structured data, so the agent's comparison includes you. See your product page is now an API.
- Cart and checkout: no CAPTCHA for assistants at the cart, labeled fields and guest checkout. See checkout for AI shopping agents and guest checkout.
- Handoff: make the moment the person takes over easy. A cart that keeps its contents, a clear total and a payment step that works first time.
New questions for the weekly review
You don't need a new dashboard to start. Add four questions to the meeting where you already review the funnel:
Ask each week:
- Did assistants read our key product pages, and did any get turned away?
- How far did agents get toward an order, and where did they stop?
- How many visits and orders came from AI assistants' answers?
- Did anything we shipped change what agents see?
The funnel isn't going away. Part of it is moving into conversations you're not part of, and the businesses that measure that part will know where they're losing customers before their competitors do.