Product updates

What we have shipped, newest first. If it is listed here it is running in production — nothing on this page is planned, in review, or nearly done.

  1. When nothing changed, we now say nothing changed

    We measured how much our own score moves when nothing about a site has changed. Running the same brand through the same code three times produced 40, 25 and 25 — a 15-point spread, all of it from a single query answering differently one time out of three. Nobody had touched the site between those runs.

    That number matters more than it sounds, because the smallest change a single real improvement can produce is 10 points. The noise is larger than the signal. So a month-over-month movement of 10 or 15 points is not evidence that anything got better or worse, and until today we would have shown it to you as though it were.

    From this release, any month-over-month movement of 15 points or less is reported as "no detectable change" rather than as a number. The underlying figures are still recorded and still available — we are not hiding them, we are refusing to draw a conclusion the measurement cannot support.

    We would rather tell you we cannot yet detect your improvement than take credit for our own variance. The real fix is a wider set of questions, which is what we are building next: the variance is between queries rather than within them, so asking more of them is what narrows it.

  2. Every score now records how it was calculated

    Each score carries the version of the scoring method that produced it, and which AI models actually answered the questions. Both were previously invisible.

    This closes a real gap. When we improve how the score is calculated, every number moves — and without a version stamp, that improvement is indistinguishable from a change in your brand's standing. Scores computed by different versions are now marked as not comparable rather than being subtracted from one another.

    The second half matters for the same reason. We ask several AI models the same questions, and if a model is unavailable the request can fall back to another one. A run that quietly answered with a different model is a different measurement, and we were not recording which one answered. Now we are.

  3. The monthly re-audit now proves it ran

    Your score is refreshed on the first of each month. On 1 August that job reported success and did nothing at all — it re-audited none of the brands it tracks, and every system we had said it was fine.

    The cause was mundane: the job hands its work off to a background process and never checked whether any of it arrived. It reported success at the moment of handing off. We found it by reconciling internal request records that outlive the logs, not because anything alerted us.

    It now keeps a record of every re-audit it starts, checks each one, and fails loudly if any brand was missed. The August re-audit has been run and completed for every tracked brand. We are stating this plainly because a monitoring system that reports success while doing nothing is worse than having none, and we would want to be told.

  4. This changelog, and the facts a model reads about us

    This page. Until today /updates/ was blocked in robots.txt and returned 404 — we were hiding a page that did not exist, on a site whose entire argument is that machines should be able to read you. It is now a real changelog, and the second indexable page on this domain.

    We also started stating who runs the company. An audit of our own domain found that AI answers about Converg3nce were sourced entirely from third-party profiles, and got the founder list wrong, because this site asserted nothing about its own leadership. The founders are now named in the structured data, in agents.json, and in llms.txt.

    And /llms.txt is now generated from the same source of truth as every other machine surface, rather than maintained by hand. A hand-written copy of a fact stays correct exactly until the fact changes.

  5. The verification record — /v1/verify

    An endpoint an AI shopping agent can call to ask whether a business can be trusted, before it recommends or pays. It returns one machine-readable answer the agent can act on directly. Access is by API key, so we know who is calling and can revoke per caller.

    Two design choices worth stating plainly. If we have never audited a domain we say so explicitly rather than returning an error — an agent that gets an error cannot tell 'Converg3nce is down' from 'this merchant is unknown', and those need opposite responses at a payment step. And if our own measurement was too thin to stand behind, the record says so rather than publishing a confident number.

    Everything else we sell is auditing: measure a brand, say what is wrong. This is the other half — the record a machine checks before it transacts.

  6. API and dashboard on our own domain

    DNS moved to Vercel and two branded subdomains went up: api.converg3nce.com for the machine-facing API, and dashboard.converg3nce.com for the customer console. An agent-facing product that hands out a *.vercel.app URL does not read as infrastructure.

  7. Operator console rebuilt, and the first client report from the live pipeline

    The internal tool we run engagements from was one long scrolling page. It is now a routed console with six working surfaces — audits, fixes, reports, leaderboard, activity log, and chat. The light theme was reworked so its status colours are legible, which sounds cosmetic and is not: those colours are how an operator tells a passing check from a failing one.

    In the same window, the first real client report was produced end to end by the live pipeline rather than assembled by hand. That is the difference between a demo and an operation.

  8. AI crawler controls on client reports

    We host client audit reports on this domain. Those are confidential, and must not end up in a search index or a training set. The named AI crawlers are now explicitly blocked from that directory — including OpenAI's and Anthropic's — and the rules are in a file anyone can read in ten seconds.

    It is the first question a serious prospect asks about handing us their data.

  9. Machine-readable agent layer, regenerated from one source

    Most websites are written for people. We publish a parallel set of files written for machines, so an AI reading this site gets clean facts instead of inferring them from marketing copy. Ours had drifted badly: it described a differently-spelled company, and carried four links pointing at a developer laptop.

    Every one of those files is now generated from a single source of truth, so they cannot silently disagree with the site again. This is the exact failure we charge to find — two surfaces stating different facts is what makes an AI describe a brand wrongly. We had it, and fixed it on ourselves first.