AI Interview Usage Analytics

Method & caveats

Partial period

Marked *. It will grow. Never read it as a decline.

Qualified

Qualified means approved. The candidate cleared a technical session with a human, or cleared the AI assessment. That is the north star, and chart 00 plots it: 207 in this window, 626 all time. It is a different question from how much talent gets contacted and sits an AI interview, which is throughput and is what charts 02 and 03 plot.

Approved vs assessed

Assessed counts people who completed an evaluation on any of the four evidence paths: 848 in this window against a 1,761 pool, roughly four times the approval count. The two populations do not nest in either direction. The week of 25 May approved 12 against 10 assessed, because someone assessed earlier can be approved later. Never difference them and never read assessed as a denominator for approval.

Chart 00 vs chart 03

Chart 00 counts people. Chart 03 counts AI-interview rows, of which 123 carry a qualified outcome. Different unit and population: never difference the two.

Nia attribution

Charts 04–07 use operator accounts: 359 assigned. The causal sourcingOrigin tag gives 47. Chart 06 shows both.

Completion rate

Completed ÷ created, including invitations never opened. Not a started-session rate.

Model cost

$0.023 per audio minute against measured duration. A billable ceiling, excludes text and thinking tokens, not shown before 22 May.

Growth operator cells

Co-owned rows count once per owner: 213 attributions over 199 rows. Do not sum operator cells back to people.

Counting unit

Prospect rows, not globally distinct people. Dashboard, funnel and analytics scopes are not interchangeable.

Sessions vs pageviews

The landing-source annex counts sessions; every other traffic figure on this page counts pageviews. Never difference or add the two. Internal there means Niuro’s own hosts, Cloudflare preview deployments, developer machines and the Google sign-in callback, which PostHog files as Referral and Organic Search; all are excluded from the plotted sources and published per week in the table.

Known gaps

282 interviews await a start, so recent completion and abandonment rates are not final. Historical external-list membership cannot be reconstructed: identity is retained only in an expiring cache.

Revision notes: scope cut to the decision set

• The Nia, acquisition and outcome layers were refreshed 20 Aug 2026 from read-only SQL replicating each endpoint (the interview and qualified-added layers were refreshed the same day from the Niuro API; see the notes below). The API routes read the same Postgres tables; each replication was validated by reproducing the published 12 Aug figures at a 12 Aug cutoff before refreshing, with mutable fields (pending, escalated, qualified, advanced) drifting by single digits because Postgres keeps no historical row state. Headline moves: Nia share of owner-attributed assignments 15.0% to 15.6% (399 of 2,558), completed-AI-interview funnel 705/145/48 to 775/174/64, advanced 54 to 66 (hired 8 to 9), Aimfox invitations 675 to 804 and reply conversations 61 to 80 across the Nia campaign cohort (now 50 campaigns), with every Aimfox send bucket reproducing exactly (675) at the 13 Aug cutoff. The Growth talent-map section is deliberately not refreshed: its source PDFs remain the 14 Aug exports.
• Nia’s real reach was refreshed on 20 Aug 2026: 758 of 1,561 (48.6%) became 808 of 1,602 (50.4%) on the same scrape-presence basis. Cause: the rolling 12-week window moved, 25 May–13 Aug to 1 Jun–20 Aug, and matching now runs against the live sourcing database instead of the 1 Jul and 1 Aug snapshot files the 13 Aug figure used. A frozen re-run of the 13 Aug pipeline returned 826 matched of 1,658 attribution rows against the published 815 of 1,659; email (24) and Nia-tagged (27) matches reproduce exactly, and the surplus comes from two prefilter false negatives in the original matcher. The card also gains a stricter read: 311 of 1,602 (19.4%) sit on an actual platform Nia run (33,710 of 33,713 prospect rows resolved to an identity). The two reads are not comparable and are labelled.
• Six decision charts remain. The 23 diagnostic charts were removed, and with them every PostHog-derived chart. Landing source returns as an annex, outside the twelve numbered sections, because it names the channels every acquisition read on the page rests on. The page now covers volume and cost, acquisition, sourcing and Nia attribution.
• Campaign-tag coverage was replaced by landing source on 19 Aug 2026. The retired annex published 79,063 pageviews and 0.67% campaign-tag coverage over 13 weeks ending 9 Aug; the annex now publishes 24,636 sessions over 14 weeks ending 16 Aug, of which 5,447 (22.11%) carry an identified source. The unit changed from pageviews to sessions and the window gained a week, so the two sets of figures are not comparable. Cause: coverage measured only whether a utm_source existed, and only 374 of 24,636 sessions arrive tagged at all, spread over nine values (linkedin 249, chatgpt.com 58, juicebox 38, whatsapp 14, email 6, hs_email 4, adwords 2, social 2, nia 1), so it could not name a channel. Campaign tagging is retained as a table column, on the new session basis.
• Two PostHog channel buckets are corrected, not republished. Referral is 3,311 of 3,391 sessions from Niuro’s own hosts, and Organic Search includes 688 sessions from the accounts.google.com sign-in callback. Both are labelled internal and excluded from the plotted sources; the excluded count is published per week.
• Identified sessions revised down on 20 Aug 2026: 5,473 to 5,447, 22.22% to 22.11%. Cause: the internal rule matched Cloudflare *.workers.dev previews but not *.pages.dev, and matched no developer host at all, so 26 sessions from preview deployments, localhost, 127.0.0.1 and a LAN address were published as identified acquisition. All 26 are now internal. Other referrers moved 82 to 56; no other series changed.
• The remaining reads are operational. Chart 03 adds unit cost beside volume; Growth and charts 03–05 show weekly movement; scorecards link only to retained charts.
• Chart 00 now plots approval, not assessment. Qualified means the candidate cleared a technical session with a human or cleared the AI assessment, and that is the north star. The card previously plotted first completion of an evaluation under the heading “how much qualified talent are we adding”, which overstated the north star by roughly four times: 848 → 207 in this window and 1,761 → 626 all time. Cause: the report read dashboard.journey completion events, while the platform records approval separately as a pass badge on the candidate profile (passedTechAt, with passedTechEvaluationMethod naming human or AI). The assessment series is not deleted: it is the Assessed column in the table and a second scorecard tile, labelled as throughput.
• The approval mix inverted in mid-July. Technical sessions with a human ran 11–22 a week to the week of 6 Jul, then 1–5. AI approvals start the week of 22 Jun and run 3–13 a week after it. Window split: 165 human, 42 AI. All time: 584 and 42. The two series partition each bar, because a profile carries exactly one evaluation method.
• The AI interview layer was re-pulled on 20 Aug 2026. Charts 02 and 03 closed on 10 Aug and were stamped 12 Aug while metric zero already ran to 17 Aug and cold email to 19 Aug; both charts now carry 17 weeks to 20 Aug. The week of 10 Aug was published at 3 of 7 days elapsed and is now a complete week: 96 → 159 created, 21 → 66 completed, 1 → 14 accepted. The week of 17 Aug is new and partial at 4 of 7 days. Window totals go 1,616 → 1,773 created, 711 → 795 completed, 110 → 123 qualified, and the completion rate 44.0% → 44.8%. Seven earlier weeks moved by one or two rows on advanced and qualified, and the week of 3 Aug on seven of nine fields. Cause: a recruiter writes those two fields when they act, not when the interview is created, so a settled week keeps accruing them. Those earlier values were correct on the date they were pulled.
• Measured session minutes for the week of 10 Aug are corrected: 23.188 to 19.447 minutes, over 28 sessions to 74. Cause: the mean was measured while the week was 3 of 7 days old, and the sessions that ran in those first three days were longer than the ones that followed. A partial-week mean is not a smaller sample of the same number, so the old figure priced that week's cost per completed interview too high. Every settled week from 18 May to 3 Aug returns its published value unchanged.
• The cold email row in the source table was eight days stale. It read 26 Jun to 12 Aug as of 12 Aug; the series behind it was re-pulled on 19 Aug and runs to 19 Aug. No published figure moves. The table row was missed when the series was refreshed.
• Cold email figures revised, 19 Aug. Re-pulled from one endpoint. w/c 3 Aug: unique opens 1,275 → 1,316, clicks 120 → 131 — opens and clicks count by event date and were still accruing when the 13 Aug fixture was captured. w/c 10 Aug is now a complete week (1,685 → 1,907 sends) and w/c 17 Aug is new and partial. Bounce for w/c 22 Jun goes 2 → 3 and w/c 29 Jun 2 → 10: the 17 Aug bounce pull used a second endpoint that disagreed at day boundaries. The nine buckets now sum to exactly the 11,683 sends and 935 bounces a single window pull returns.
• Qualified talent added re-pulled to 20 Aug 2026. The week of 10 Aug was published partial at 3 of 7 days and is now a complete week: 33 → 56 added, pool 1,696 → 1,719. The week of 17 Aug is new and partial at 4 of 7 days: 42 added, pool 1,761. Window total 783 → 848 over 17 weeks, active in window 801 → 866, requalified holds at 18. Cause: the series is live-sourced and was carried from the 12 Aug as-of to the report date; every one of the 16 previously published cumulatives reproduced exactly, so nothing before the week of 10 Aug moved. The anchor moved 13 → 20 Aug 2025 because the endpoint caps a window at 366 days; it reproduces the published 913 and 1,696 exactly. Chart 00 now ends 8 days after every fixture-derived chart, which is frozen at the 13 Aug export and closes on 12 Aug, so the 4W and 8W buttons cover a later span here than there and the two are not comparable. The unqualified-supply tile is not part of this re-pull and stays the 12 Aug inventory of 12,242, against 12,872 reported live on 20 Aug.
• The document is the source of truth. Sections are authored in markup, not hidden by a runtime allowlist. All-time distinct people sourced by Nia is still unknown (bounded at ≤ 25,599) and needs one SQL query.