How it works

AVA measures what AI assistants say about your organisation, records each measurement as evidence you can show someone else, and compares those answers against what your site actually publishes.

The shape of a run

There are four steps and you do them once. Create a project — one site you want to be findable in AI answers. Add a domain and prove you control it, by publishing a short challenge at /.well-known/ava-site-verification or as a DNS TXT record; we then fetch it and check. Add the questions your buyers actually ask, in their words rather than yours. Start a crawl, which reads your own pages so an answer can be checked against what you publish.

After that, a sampling run puts each of your questions to an AI provider and records what comes back. Runs are yours to trigger; nothing samples on a schedule you did not set.

Why the domain check exists

Nothing is crawled until a domain is verified, and that is a control rather than a delay. Without it the tool would fetch any address anybody typed, from our network and under our name. The challenge is not a secret — you publish it precisely so that we can come and read it. Knowing it confers nothing; proving control of a domain still means being able to publish on it.

What the crawler does, and what it will not do

It fetches your pages, records each one’s address, HTTP status, content hash and robots.txt verdict, and stores the HTML as it was served. It honours robots.txt, it is confined to the exact domain you verified, and a redirect that leaves that host is refused rather than followed. Links to other hosts are dropped rather than queued — so a crawl of your main site does not wander into anything else you happen to link to.

The default is 25 pages. That is not a cost limit: crawling is cheap. It is there because a large sweep of a small server is a load event we caused, and because what a model says about an organisation is shaped by its front door, its about page and its pricing rather than by page 300 of an archive. You can raise it to 500.

Which AI systems are sampled

The providers your access permits, using their APIs. Each run names the provider and the model on the record, so a figure can always be traced back to what produced it. A provider your access excludes is reported as excluded — never silently omitted, because a number with a quietly shrunken denominator is worse than no number.

What is sampled is the API, not the consumer product. We do not measure what the chat website says to a logged-in person with a history and a memory of them. That is a real difference and it is stated rather than glossed.

The record, and what it proves

Each observation is hashed. A sampling batch is sealed as a whole: one hash covering every observation in it, signed and sent to an independent timestamping authority, with the result written back to every record in the batch. Anyone can recompute an observation’s hash, find it among the batch’s leaves, recompute the root, and check it against the stamp.

The word is tamper-evident, never immutable. Nothing we build prevents a record being altered. What the seal does is make alteration detectable by someone who does not trust us. A vendor claiming immutability is claiming something no software can deliver about data it holds.

Two grades of timestamp exist. The ordinary one is a real, independent, cryptographic timestamp and carries no legal presumption anywhere. The qualified one is issued by an accredited EU authority and carries statutory recognition in the EU — it costs money, so your access includes a set number and the dashboard shows how many remain. Where a record was not qualified-stamped, the record says so. Nothing is ever presented as qualified when it was not.

Where your data lives, and what that does and does not mean

The records are held on infrastructure in New Zealand or the EU, with no US-owned cloud in the chain, and the sealing is performed by My Digital Sovereignty Ltd as an independent custodian. That is what sovereign custody means here.

It does not mean the measurement path is sovereign. Measuring what OpenAI or Perplexity say about you necessarily means asking OpenAI and Perplexity, who are US companies. Your questions and the answers travel to them. What is held in a jurisdiction you chose is the record.

What this is not

It is not a ranking tool and it does not promise to change what a model says about you. It reports what is being said and what on your own site supports or contradicts it. It does not establish that a change you made caused a change in an answer — models change underneath any such comparison, and a before-and-after that ignored that would be a story rather than a measurement.

If something looks wrong

A crawl that stays queued means the worker process is not running — tell us, it is ours to fix, not yours. A sampling run refused with “runs exhausted” means your access allowance is spent; nothing was charged and no provider was called. An observation with no attestation carries the reason it has none, on the record, rather than an empty field.

Specific questions are answered on the questions page.