Questions

What is AVA, and why would I point it at my website?

People increasingly ask an AI assistant rather than a search engine. When somebody asks what your organisation does, whether you serve their region, or who they should use for the thing you sell, a model answers — and it answers from whatever it has absorbed, which may be outdated, partial, or about a competitor. AVA measures that. It puts the questions your buyers actually ask to AI systems, records what comes back, and compares those answers against what your site actually publishes. Where a model says something about you that your own pages contradict, that gap is the finding.

The reason to point it at your site is that this is currently invisible to you. Analytics show who arrived; they cannot show the conversations where you were described wrongly, described as someone else, or not mentioned at all. The work is to improve what is said about you, for the terms your buyers actually use — and the outcome is measured rather than promised, because nobody controls a model's answer. Each measurement is kept verbatim with the assistant and the time, and can be sealed with our key and time-stamped by an outside authority.

What do I actually get at the end of a run?

A list of findings about your own pages, each one tied to the page it came from and to the evidence behind it, ordered so the ones worth doing first are at the top. Every finding carries what was observed, how it was measured, and what to change. You can read the evidence and disagree with any of them — that is the point of reporting rather than opining.

Alongside those, a record of what AI systems said when asked your buyers' questions: every answer verbatim, which model gave it, and when — checked against what your own site publishes. Where a model says something your pages contradict, that gap is the finding.

Each finding can be turned into a draft change, and the draft comes out in the form the person applying it can use: prose for someone working in a website panel, or a brief written for a coding assistant with the repository in front of it. The quotations in a draft are checked against the captured page before it can leave the system, so a draft cannot cite something that is not there.

What sorts of things will it tell me to change?

Three kinds, and they are worth separating because they fail differently. Reachability — whether an AI system can fetch and read your pages at all. This is the one with a mechanical answer: an agent that is disallowed cannot read you, and no amount of writing changes that. It is also the finding that makes every other finding on the page moot until it is fixed.

Intelligibility — whether what you publish can be used once it is read. A title that says nothing, a page whose content only appears after scripts run, a claim buried under six paragraphs with no heading over it.

Vocabulary — whether your pages use the words your buyers search with. You tell us your brand names, the things you sell and the ideas you want to be found by; the audit reports which of them appear nowhere on a page, and which appear only in passing rather than anywhere a reader or a model looks first. It never tells you to insert keywords: if a term does not belong on a page, removing it from the list is an equally good answer.

Will this get me into ChatGPT's answers?

Nobody can promise that, and anybody who does is selling something they do not control. The work is to make you findable for the terms your buyers use, and every effort goes into it. What cannot be promised is the outcome, because what a model says is decided inside the model.

So the result is measured rather than asserted. What was true before, what was changed, what is true after — each measurement recorded verbatim with the assistant and the time; once a run is sealed, the earlier figure cannot quietly become more flattering later. Most before-and-after claims in this field rest on a number the seller could have adjusted afterwards.

What that still does not establish is cause. Models change underneath any such comparison, and a before-and-after that ignored this would be a story rather than a measurement. The chronology is evidence about what happened and when; the inference is yours to draw, with the confounders stated.

My site is on Squarespace or Wix. Is this any use to me?

Yes, and the audit adapts to the platform rather than assuming a developer is reading it. Most website owners work in a hosted editor, not in a repository. An instruction to insert a tag before the closing head element is written for somebody with access to the markup; for everybody else it describes a route their platform does not offer, which is a different sentence and needs to be a different recommendation.

The audit detects what your site is built on and routes each recommendation through a panel you actually have. Where your platform gives you no route to the page head at all, it says so and redirects you to what you can change — the title and description fields carry much of the same signal. And it asks you once whether you can reach that panel, because which plan tier includes it changes quarterly and a tool that guessed would be wrong within a season.

It also suppresses findings that are properties of your platform rather than of your site, with the reason kept so you can argue with the suppression. Telling a Wix owner their content is JavaScript-rendered, when that is true of Wix and not of them, is noise wearing the costume of insight — and a report full of impossible tasks teaches its reader to stop reading it.

Do I have to give you API keys, and what does a run cost to perform?

On the hosted service, no. You supply nothing but the questions and the site. On a self-hosted install it is your choice: point it at our inference service, or supply your own commercial API keys so it calls those providers under your own account. The second assumes you already hold keys and know what they cost, which is why the self-hosted path is aimed at agencies and developers rather than at individual site owners.

Sampling a model that you pay for costs us nothing, which is why the price does not vary by how many AI systems you sample. The one thing that does cost money per use is the qualified timestamp — each qualified stamp is priced. A record without one carries the Certum stamp and says so on its face.

What do you keep, and is any of it used to train anything?

Nothing you give us trains anything. What is kept is what an audit is made of: the pages as captured, the answers as returned, the findings drawn from them, and the sealed records of each measurement. They are held on infrastructure in New Zealand or the EU and no US-owned cloud holds it.

⚠️ One thing to be clear about, because it is a real limit and not a small one. Finding out what a US model says about you means asking a US model: your questions and its answers travel to that provider under their terms, not ours. The measurement path is not sovereign and cannot be. What is held in a jurisdiction you chose is the record.

What happens to my records if I stop paying?

The records you have already sealed stay verifiable. A registrar that took verification down on non-payment would have made every seal it ever sold worthless, and a seal that can be switched off was never evidence in the first place.

What a subscription buys is new measurement — further audits, further sealed records, and the working system. What it does not buy is continued permission for what already happened to have happened.

What runs on my own hardware, and what does it need?

That depends on which of the two deployments you are on, and the answer names both. On the hosted service, nothing but a web browser: the crawler, the analysis, the database and the sealing run on infrastructure we operate in New Zealand or the EU, there is no client to keep patched, and the minimum specification is whatever runs your browser.

There is also an install kit for running AVA on your own infrastructure, and that is a design requirement rather than something bolted on later. It carries its own schema and migrations and does not need a service of ours to crawl a site or analyse what it finds. The crawl then reaches your site from your network rather than ours.

Two parts stay with us, for different reasons. The sealing does, because a seal you generate yourself proves nothing to a third party — the whole value of the record is that an independent party made it. And the models you sample are your choice: either you point the install at our inference service, or you supply your own commercial API keys and it calls those providers under your account.

How you get it: the kit is issued, not published. There is no trial and no free tier — that is an access ruling rather than a technical limit, and the two should not be confused.

How long does an audit take?

On the hosted service it does not run on your laptop, so the figures below are about our timing rather than your hardware. On a self-hosted install the same work runs on whatever you put it on, and the crawl is still bounded by the politeness delay rather than by processing power.

Measured on 21 September 2026, auditing mysovereignty.digital: eight pages fetched in 42 seconds end to end, about one second per page including the pause between requests. Page sizes ranged from 6 KB to 53 KB. A 50-page site is minutes, not hours. You start it and come back; nothing is tied to your machine staying awake.

That figure is one measurement of one small site over a fast connection. It is an order of magnitude, not a service level.

Does an audit produce a timestamped record I can rely on later, and how would I test it?

The word is tamper-evident, never immutable, and the distinction is the product. Nothing we build prevents a record being altered. What the seal does is make alteration detectable by someone who does not trust us: the record is hashed, signed, and stamped by an outside time-attestation authority, so a later change breaks the chain and anybody can see that it did. A vendor claiming immutability is claiming something no software can deliver about data it holds.

What the stamp is today, precisely. Sealing is available on request. Where it is used, records are time-stamped by Certum, an RFC-3161 authority in Poland. That is a cryptographic time stamp from an outside authority — and it is not eIDAS-qualified, so it carries no statutory presumption in any jurisdiction. The application says so in the record itself rather than leaving it to be discovered.

A qualified authority is contracted, built and tested: Aruba PEC, an Italian qualified trust service provider, checked against its live endpoint. It is an opt-in, switched on per customer, and each qualified stamp is priced.

✅ The seal has now run on a real observation — 21 September 2026. An answer was recorded, hashed, and sealed against the live attestation service; the receipt came back stamped by Certum. The whole chain rebuilds from the stored record alone — the answer, its hash, the batch root, and the digest the authority signed — with no network and no key needed to check it. Change one character of the answer and every check refuses it.

Checking a record without us: a sealed run can be downloaded as a self-verifying bundle holding the signed receipt, the observation it covers and the steps to check it entirely offline. Hand it to anyone; they can verify it against the signer’s DID document with no account here and no call to our API. The digest is recomputed from the observation in the bundle, so any alteration shows.

Where does the sealing run — on my hardware, or yours?

Ours. SSST — sovereign, sealed, signed and timestamped — is operated by My Digital Sovereignty, and is one of two assets intended to remain permanently owned by it, the other being the Governance API.

The name is the mechanism, in order. A record is sealed so its contents are fixed to a hash; signed so the seal is attributable and cannot be forged; timestamped by an outside authority so the moment it existed is attested by someone other than us; and sovereign because all of that happens on infrastructure in a jurisdiction you chose, and no US-owned cloud holds the record. Take any one away and the record stops proving what it claims: unsealed it can be edited, unsigned it can be fabricated, unstamped it has no date anybody else will accept, and without sovereignty it is subject to a legal regime you did not pick.

It is the primitive underneath AVA and the conduct record, and AVA is one application of it. Its value depends on the sealing happening somewhere the customer does not control — a seal you could regenerate yourself would prove nothing to a third party.

More

What it does, and what it does not · Pricing · Ask something that is not here