Infrastructure for agentic commerce
AI agents now read your catalogue before your customers do. Most platforms block them without knowing it, serve them an empty page, and measure nothing at all. Metavizor diagnoses, fixes and instruments.
An engineering firm led by practitioners: our consultants work in production, not in workshops.
A page served to the visitor, refused to whoever looks for it on their behalf. That is what the test measures on your address.
The layers where agent accessibility is decided : Varnish, Cloudflare, Datadome, Akamai, Fastly, NGINX, LiteSpeed, Magento, Shopify Plus, PrestaShop, Salesforce B2C, Next.js, Kubernetes, Terraform, Grafana, Elasticsearch.
Services
Two areas make your catalogue readable by agents; two others keep the production that serves it standing. They sell separately and reinforce each other.
Find out whether agents can see you — and why they can't.
Produce the number that unlocks the budget.
Hold up on the days that matter.
Use AI where it actually delivers: operations.
Platforms
Diagnosis and fixes depend on the stack. These are the ones we work on in production.
Magento / Adobe Commerce
Server rendering, Varnish cache, product feeds
Shopify Plus
Themes, Liquid, structured data, third-party apps
Salesforce B2C Commerce
SFRA, PWA Kit, catalogue exposure
PrestaShop
Rendering, robots.txt, sitemaps, performance
Next.js & client rendering
SSR, ISR, hydration, what an agent sees without JavaScript
Bot protection & CDN
Cloudflare, Datadome, Akamai, Fastly, NGINX, LiteSpeed
Resources
Technical analysis written by the people doing the work.
·7 min
The neighbouring-rights agreement has just been reached: Google will roll out AI Mode and AI Overviews in France on 22 July. For a retail site, the consequence is not a drop in rankings: it is a shift in where the purchase decision gets made.
·6 min
An llms.txt file takes an hour to put in place and costs nothing. It is also the subject with the most false claims in circulation — in both directions.
·8 min
Most retailers who discover they are absent from generative answers do not have a content problem. They have a WAF rule written in 2021, before these crawlers existed, that nobody has re-read since.
Method
We open your logs, your WAF rules and your templates.
We execute the plan: access rules, server-side rendering of critical templates, structured data, feeds, instrumentation.
The dashboard runs, alerts catch regressions, and a monthly review decides what comes next.
22 July 2026
AI Mode and AI Overviews rolled out in France by Google
26%
of France's top 100 retailers have deployed an AI assistant
79%
of those assistants do nothing but redirect to a product page
The diagnosis
None of them show up in a marketing report. All of them are visible in your logs within the hour.
ChatGPT-User, Claude-User, OAI-SearchBot, GPTBot and the rest fall under the same Datadome, Cloudflare or Akamai rules as malicious scrapers. The costliest is the first: a customer named you in an AI, and the agent sent to read your page is turned away. You are absent from generative answers, and nothing in your dashboards says so.
Do not assume an agent sees what a browser shows. Googlebot has rendered JavaScript for years; training crawlers do not — and behaviour differs by crawler and by retrieval mode. Where a browser shows a complete product page, they often receive an empty container. So you can be perfectly cited by Google and absent from ChatGPT, Claude or Perplexity. Our test measures the page as it reaches them.
Varnish, CDN, purge and variation rules: depending on your URL patterns, your headers and your cache variation strategy, agent traffic can degrade the hit ratio your infrastructure was sized on and drive origin load up with nothing in return. Measuring it is the only way to know.
LLM-driven traffic is segmented neither in your analytics nor in your CDN logs. With no numbers, no budget decision is possible — and the topic stays a corridor conversation.
Free test — 30 seconds, no sign-up
We query your page live using the user agents of GPTBot, ClaudeBot, PerplexityBot and Google, then show you exactly what they receive. Not an estimate: your server's actual response.
We keep the address tested and its score to follow how the tool is used, and each test reaches our internal mailbox with the technical connection data — described in our privacy policy. Nothing is published or shared with any third party.
Your server refused our reference request, made with an ordinary browser user agent. The refusals then given to AI agents prove nothing: it is not them being turned away, it is our address.
This is how IP-range filtering behaves — common with Cloudflare, Datadome or Akamai, which turn away datacentre traffic. It can only be observed in your own server logs.
That is exactly what the full audit examines.
Request an auditAccessibility score
This test queries your server from our own infrastructure, announcing each agent by name. Two things escape it, and they cut both ways. IP-range filtering may refuse us while your visitors get through. And some firewalls verify a bot's real identity by its address: they will turn our request away while letting the genuine crawler in, making us report a block that is not one. Only your server logs settle it — that is what the full audit is for.
Want to know what this test cannot see?
The full audit opens your server logs and WAF rules, template by template, and puts a cost on every fix.
A diagnostic takes five to ten days. It tells you whether AI agents reach your catalogue, what stops them, and what the fix costs.
Request an auditWe would like to measure this site's audience with Google Analytics, which requires setting cookies. You may refuse: the site works exactly the same, and nothing will be sent. Our privacy policy