What we changed, and whether it worked.
We install an AI operating system for other people's marketing, so this website runs on one. This is its record: each change, the reason, the measure, and the result, including the ones that proved nothing.
How the loop works.
Traffic here is small, tens of visits a week, so we do not run A/B tests. We judge each change by checks we can run and counts we can read, and we say plainly when that is not enough.
- 1. MeasureThe site records anonymous visits and where people leave the Diagnostic. A weekly self-score checks structure, discoverability, content, design and measurement.
- 2. DiagnoseThe lowest-scoring pillar is the bottleneck. The same Hexagram we use on clients names it, and the Monday digest reports it.
- 3. ChangeOne fix at a time. Low-risk fixes such as link targets and alt text merge on their own. Copy, layout, claims, prices and legal text always wait for a person.
- 4. CheckEach change records what would show it worked. When the data is too thin to say, the entry says inconclusive, not kept.
Changes, newest first.
Privacy: consent, unsubscribe and an accurate policy
- Problem
- Analytics and the Meta Pixel loaded before anyone chose, completing a form quietly joined the mailing list, follow-up emails had no unsubscribe link, and the privacy policy described AI scoring and human review that the code did not do.
- Change
- A cookie choice with equal Accept and Reject, a separate unticked box for marketing email, a signed one-click unsubscribe, a rewritten privacy policy for India, the EU and UK, the US and the UAE, automatic deletion on the stated schedule, and a script to export or erase one person's data.
- Measure
- Four self-score checks that fail if a tracker loads without consent, a form lacks a privacy link, the mailing list is joined without consent, or a marketing email lacks an unsubscribe link.
- Result
- All four checks pass in the repository, and each was shown to fail when the control was removed. It takes effect in production once the database migration is run and the unsubscribe secret is set. Whether the policy meets each country's law is a question for a lawyer.
The site now measures itself
- Problem
- Visitor behaviour lived in Google Analytics and the Meta Pixel, which an agent cannot query. Nobody could say where people left the Diagnostic, and the site's own health was checked by hand.
- Change
- An anonymous first-party event record (no cookies, no IP address, no query strings), a Diagnostic drop-off section in the Monday digest, a weekly self-score against the six Hexagram pillars, and this page.
- Measure
- Whether the Monday digest names a specific step where Diagnostic visitors leave, and whether the self-score's bottleneck leads to a change that fixes it.
- Result
- Needs a few weeks of events before the first drop-off reading means anything.
One client count everywhere
- Problem
- The homepage said 30+ brands while four other pages said 47+ clients.
- Change
- All five mentions now say 50+, the figure Abhiraj confirmed.
- Measure
- The self-score check that compares every client-count claim on the site.
- Result
- The check moved from failing to passing. It is a consistency fix, so no visitor metric applies.
Links that went nowhere now go somewhere
- Problem
- Four homepage case-study cards, the agency comparison link, the specialist-bench link and the blog share buttons pointed at "#", so they looked clickable and did nothing.
- Change
- Each now opens its case study or post, the LinkedIn and X share pages, and the Copy link button copies the post address.
- Measure
- The self-score check for links that point at "#", and the check that every internal link reaches a real page.
- Result
- Both checks pass. The check found five places in the code before, and finds none now.
The header stays on one line at laptop widths
- Problem
- Between 1300px and 1440px wide, nav links and both buttons broke onto two lines. The logo file is a 2000px square with the wordmark in its middle third, so the 240px logo column showed a mark about 53px wide.
- Change
- Tightly cropped logo files, a logo column sized to its content, and narrower gutters below 1440px.
- Measure
- Header height and line count at 1301, 1366, 1440 and 1920px wide, measured in a browser.
- Result
- The header is 86px tall on one line at every width tested, down from up to 127px on two lines.
FAQ questions use the site font
- Problem
- Buttons do not inherit the page font by default, so every FAQ question rendered in the browser's Arial.
- Change
- Form controls now inherit the font family from the page. Sizes and weights are unchanged.
- Measure
- The computed font family of every button and input, compared before and after on nine pages at two widths.
- Result
- Only the font family changed, on every control that had been falling back to Arial.
One design system, written down once
- Problem
- DESIGN.md and the design-checker data still described a retired system with no rounded corners, no shadows and gold actions, while the site and the published design system did the opposite.
- Change
- Both files were rewritten from the published ADG Advisory design system.
- Measure
- Whether the written rules match the code, checked by rendering the design file's own component examples.
- Result
- The examples render like the published system. Drift in the code itself is now tracked by the weekly self-score.
Want this loop running on your marketing?
Every engagement starts with the Hexagram. The Diagnostic scores your six pillars and names the bottleneck, the same way this page does for ours.