Slide 2 · Method

How the analysis was run

• Pre-registered: the three hypotheses and what each one predicts were written down before analysis, so the test is deduction, not narrative.
• Blind: the prompt contains only verified data and the hypotheses. No conclusions, no opinions, no company names.
• Two independent systems: the identical prompt was run verbatim in Claude (max reasoning) and ChatGPT (deep reasoning), in fresh sessions with no prior context.
• Falsifiable: the prompt explicitly asks each system what evidence would prove each hypothesis wrong.

Anyone can re-run the same prompt and check the result. That is the point.

Slide 3 · Hypotheses

Three explanations, three sets of predictions

Slide 4 · The data

Three years, same report, same window

Views tripled. Completions rose by 519 in two years — just 21 of them in 2026. GA4 shows the entire 2026 traffic increase was paid: 12,577 paid landings on the form page versus 0 in 2024, while direct and organic traffic stayed level.

Sources: 700Credit QuickQualify reports (pulled by Mark), Google Ads billing records, GA4. Staff test accounts excluded.
Source reports: QuickQualify Results Analysis, pulled for Jan 1–Aug 6 (2024/2025/2026), Jul 28–Aug 6 (2024/2025/2026), and Jul 24–Aug 6 (2026).

Slide 5 · Finding one

Ads on vs. ads off — same form, same week

Ads were paused July 27. The two-week report minus the ten-day ads-off report isolates the four final ads-on days:

Same form. Same defect. Same week. A four-fold difference in completion rate, statistically decisive despite the small window. A platform defect applies to both periods equally — it has no mechanism to produce this step.

Slide 6 · Finding two

A defect suppresses a percentage. It can't cap a number.

If a broken form loses some share of everyone's submissions, then tripling the traffic still roughly triples completions — at the depressed rate. That is not what happened:

The incremental traffic converted at 0.15% — one completion per ~680 added visitors. Both AI systems reached the same verdict: no defect-only story can produce flat completions under tripling views. Only near-zero-converting incremental traffic can.

Slide 7 · Finding three

The form-side decline is real — and it started first

In matched windows with essentially no ad spend in any year, the rate still fell:

Ad traffic cannot explain this — there was none in these windows. The 2024→2025 drop is mobile-only while desktop held flat: the signature of a mobile-side platform problem, consistent with what 700Credit has confirmed. This side of the decline is real, it is the larger share, and it began before the ad spend existed.

Slide 8 · Verdict

Both systems, independently: it's both

Each AI system rejected both single-cause stories and decomposed the 17-point decline the same way, within a few points of each other:

Claude: 58/42 (≈50/50 under symmetric attribution). ChatGPT: 65/35. The difference comes from one methodological choice — whether ~81 residual paid sessions are removed from the 2026 clean window — and is shown here rather than smoothed over.

The platform side is the larger share and owns 2024–2025. The dilution side owns the 2026 collapse — and it is the share the money is attached to: roughly $42k of spend for on the order of 21–63 attributable completions.

Slide 9 · A mechanism finding

The confirmation-page bug can't be the driver — but the SSN prompt can

Both systems independently flagged the same detail: the blank-confirmation defect happens after a successful submit. The completion is already counted by the time the customer sees the blank page. A post-submit bug therefore cannot lower the reported completion rate — if anything, resubmissions inflate it.

What can lower it: the form's SSN demand when a customer's name doesn't exactly match bureau records. Anyone who types "Bob" instead of "Robert" is asked for an SSN — on a product marketed as "No SSN required" — and abandons before submit. That is pre-submit friction, it is device-agnostic, and its onset date is unknown.

The non-ads decline is real either way. This finding redirects which platform mechanism to investigate — and it makes the SSN behavior a marketing-accuracy question as well as a conversion one.

Slide 10 · Open items

What would change this conclusion

• Deploying the built fix is the single most decisive test both systems identified: if the ad-free rate rebounds toward 27%, the platform share grows. If it moves under 2 points, it doesn't. The fix is built and, as of Aug 7–8 testing, not deployed.
• QuickQualify's view counting is unverified across years. QQ shows 72% more 2026 form views than GA4 shows landings, the date-range correction's retroactivity is unknown, and one internal inconsistency exists in the 2025 web figures. The decomposition is denominated in QQ views — its precision is capped until this is reconciled.
• Source-tagged completions (UTM/gclid passthrough) would replace every inference in this deck with a direct measurement.

Slide 11 · For Tyler / 700Credit

Four requests, in priority order

Also open: UTM/gclid passthrough or a completion event we can capture (thread of Jul 27 — five integration options listed). Any of the five works for us.