Simulated Agent Buyer

Real AI agents try to buy from you. You see where they fail and why.

Browser agents with buyer intents and personas run your purchase flow while every step is recorded, producing an agent conversion funnel and the reasons agents stall or choose someone else.

Early access. Diagnoses are run by us and reported within five working days while the self-serve version is built.

Request a diagnosis Get your free score first

Pricing

Diagnosis
£950
Three buyer intents, 60 instrumented runs, the agent funnel and the failure table with the agent's own reasoning at each stall.
Agent regression
£299/month
The same runs on every deploy or weekly, so a release that breaks agent checkout is caught before customers' agents find it.

Prices exclude VAT and sales tax where applicable. Monthly plans are billed monthly and can be cancelled any time. Refunds and cancellation.

Is this for you?

Your Agent Conversion Score is fine but agent-referred sessions do not convert, and nobody can say why.
You ship changes to product pages or checkout regularly and need to know whether each release still works for an agent.
You want to see the exact reasoning an agent gives when it picks a competitor instead of you.

Probably not for you if…

Your checkout legally requires a human in the loop at every step, as in regulated gaming. Use the score and monitoring instead.
Your score is capped at 40. Fix the blocker first; a diagnosis will only tell you what you already know.

What it measures

Landed: The agent reached your page without a 403, a challenge or a consent wall that stops it.
Found product: It located the item that matches the shopper's brief.
Confirmed offer: It read price, delivery and returns correctly, checked against the true values.
Added to cart: It picked the right variant and the cart updated.
Reached checkout: A guest path was available, with no login wall or captcha.
Ready to pay: It reached the payment step with the correct total. The run stops here; nothing is bought.

How it works

Two layers
Live assistants (ChatGPT, Perplexity, Gemini and Claude through their APIs) answer "which do we pick for this brief, and why". An instrumented browser harness then runs your site to find the friction.
Intents and personas
Three buyer intents drawn from your catalogue, across budget, mid-range and premium personas with shallow and deep exploration styles.
Sixty runs, every step recorded
Each run logs the page, the action, the agent's reasoning and the outcome. Failures are classified as product not found, offer misread, no-buy decision, failed to add, stuck in a loop, blocked by site, or price rejection.
The funnel and the reasons
You get the drop-off at each stage, the failure table, and session replays with the agent's reasoning at the exact point it stalled.

What you get

The agent funnel with counts and drop-off at six stages
A failure table by type, with frequency and example reasoning
Which competitor the live assistants chose instead, and on what grounds
Session replays with the agent's reasoning at the point of failure
A fix list ordered by funnel stage, with expected recovery per fix
On the regression plan, a pass or fail on every deploy with a diff against the last run

What it does not do

Questions

How is this different from the score?

The score inspects your site for the conditions an agent needs. The Simulated Agent Buyer puts an agent through the flow and records what happens. The score is the X-ray; this is the stress test.

Which agents do you simulate?

The harness uses browser agents driven by current frontier models, with session memory and guardrails, following the approach Shopify validated against real human outcomes in its SimGym research. The live-assistant layer calls OpenAI, Perplexity, Google and Anthropic.

Do we need to give you access to anything?

No. Everything runs against your public site. For the regression plan you add a webhook from your deployment platform so runs trigger on release.

What about personas grounded in our real customers?

Available on request. We build personas from a GA4 export or clickstream sample, which research shows predicts outcomes better than generic personas. Generic personas are enough to rank your failure modes.

Next in the suite

Recommendation Monitoring — When a shopper's agent is choosing, how often is it you?

Request a diagnosis.

Tell us your site and platform. We reply within two working days with next steps and a start date.

Request a diagnosis