← Go back

Guide · CompanionGuard · Published 14 November 2026 · Reviewed 14 November 2026 · By the Ingenious AI team

Counts-Only Referral Metrics: What the Dashboard Shows and Why It Shows Nothing Else

The short answer

Referral metrics in CompanionGuard are counts only: the crisis component was shown N times, clicked M times, per period, per bot — with no names, IPs, user agents or user IDs attached at any point. The dashboard turns those counts into the operator's working view (trends, locale performance, protocol-page reach), and the annual-report CSV draws its totals from the same discipline. The counts are the most the system will ever know, and this guide explains why that ceiling is the feature.

What are counts-only referral metrics?

Counts-only referral metrics are aggregate tallies of referral events — component impressions and clicks — recorded as pure numbers with no subject attached. They answer 'how many, how often, trending which way' and are architecturally incapable of answering 'who'. In CompanionGuard they are kept from the free kit upward and feed the Pro dashboard and annual-report export.

Create the free safety kit Next guide All guides

CompanionGuard is built by Ingenious AI. A compliance toolkit — not legal advice and not a clinical service. If you are in crisis right now: US — call or text 988; UK & Ireland — Samaritans 116 123.

The two events and what each means

The component records two events: shown (the resource block rendered — on the protocol page, in the snippet context) and clicked (a resource link followed). Shown measures presence and reach — your safety surface is actually being encountered. Clicked measures use — someone wanted the resource enough to act. Read together they sketch the funnel of voluntary help-seeking around your bot: page visitors who came for the protocol itself, users routed there from the banner, and the proportion who proceed to a resource.

Keep the interpretation humble: a click is not a call, a call is not an outcome, and the metrics claim none of those. What they establish is narrower and still valuable — the referral infrastructure is alive, reachable and used. Dead infrastructure (a protocol page nobody reaches, resources nobody opens) shows up here first, as a flat line where a pulse should be.

Trends are the metric; snapshots are trivia

A single month's count says almost nothing — audiences grow, seasons turn, a mention somewhere sends a wave. The dashboard's value is longitudinal: referral counts against session counts (a ratio that should move for reasons you can name), click-through holding steady as impressions scale, the effect of a protocol-page redesign visible as a step change. Watch especially for the two alarming shapes: counts collapsing while traffic grows (your surface broke — go and look) and counts spiking without a traffic or incident explanation (suspect over-referral in replies, or a bot loop rendering the component repeatedly).

Because the underlying records carry no identity, trend analysis is the only analysis available — there is no drill-down to misuse in a worried moment. That constraint disciplines the whole team's data culture: questions get asked at the level the data can honestly answer, which for safety operations is the level that matters. The counting-definition guide explains how these counts relate to the decision-level census.

From counts to the annual report

The counts-only metrics are the free tier's gift to your future filings: they start accumulating at kit creation, before any upgrade decision, so a bot that moves to Pro in month eight brings eight months of referral history with it. On Pro, the dashboard's annual-report CSV draws the referral story together — component counts as the reach line, runtime decisions as the census line, definitions stated — the material the SB 243 report consumes from July 2027.

The practical habit: glance at the counts monthly with the same regularity as revenue, and export quarterly even before any duty compels it. A team surprised by its own referral numbers at filing time is a team that was not watching its safety surface — and the numbers, being counts of help offered and taken, are among the most meaningful your product produces.

Why there is no 'just show me this one user' — and never will be

Every counts-only system eventually faces the request: an incident, a worried partner, a legal letter — 'surely you can find this one person's referrals'. The architecture's answer is that there is nothing to find: the count incremented and no subject was written. That answer, given in a hard moment, is only credible because it is structural — and it protects the operator as much as the user, because a record that cannot identify cannot be demanded, leaked or misused.

Communicate the ceiling before the hard moment arrives: in the protocol page's data-practices block, in partner documents, in your own team's training. People who understand that the system counts help the way a turnstile counts entries — passage without identity — stop asking it for names, and start asking it the questions it is brilliant at answering. The privacy guide lays out the full architecture this ceiling belongs to.

Five actions for counts-only metrics that work

  1. Confirm counting is live from the free kit onward: open the metrics after creating a bot and watch a protocol-page visit increment the shown count.
  2. Watch ratios and trends monthly — referrals against sessions, clicks against impressions — and investigate collapses and unexplained spikes as product events.
  3. Keep interpretation humble in every document: shown is not called, clicked is not helped; the metrics evidence reach and use of the referral surface.
  4. Publish the ceiling in your protocol page and partner materials: counts only, no identities exist to retrieve — stated before anyone asks in a hard moment.
  5. Export quarterly so the annual-report habit (and its totals) are rehearsed long before July 2027.

Worked example: the flat line that found a broken page

A fictional companion, Fern, glances at its monthly counts and sees protocol-page impressions down 80% on flat traffic. No incident explains it; the counts themselves are the incident. Investigation finds a site release three weeks earlier had dropped the protocol-page link from the bot's home surface — the page lived, the door to it did not. Referral clicks from the page had fallen with it, silently, because nothing about the chat product appeared broken.

The link is restored, impressions recover within days, and Fern adds the counts to its release checklist: a step change in shown/clicked after any web release triggers a surface audit. The team's note captures the moral: counts-only data cannot tell you who needed help — but it told them, precisely and early, that help had become harder to reach. The story is invented; the metrics and their events are the product's own.

Frequently asked questions

What exactly do referral metrics count?

Two events on the crisis-resource component — shown and clicked — tallied per period and per bot. Nothing about the viewer is recorded: no names, IP addresses, user agents or user IDs, by design.

Can the metrics identify a user who was referred?

No, and that is structural rather than a permission setting: the events are recorded as counts with no subject attached, so there is nothing to retrieve in an incident, a request or a breach.

What are the counts useful for, then?

Reach and health of the referral surface: trends over time, ratios against sessions, the effect of redesigns, early warning when a surface breaks — and the reach line of the annual-report export, beside the runtime's decision census.

When does counting start?

At kit creation, on the free tier. A bot upgrading to Pro later brings its accumulated counts with it, so the reporting history is never zeroed by an upgrade decision.

Does a click mean the user got help?

No — a click means a resource link was followed. The metrics evidence that the referral infrastructure is present and used; they do not claim outcomes, and your documents should not claim them either.

Keep going