Growth playbooks
How to lower CAC, fix a landing page and test ads without wasting budget. Each one opens with the short answer.
The playbooks
Fix the page before you buy more traffic: bookings went from 1.8% to 5.0%
Traffic fine but bookings aren't? Fix the page before you buy more traffic. At Gaba Healthcare only 1.8% of visitors booked a consultation. I built a page for each condition, tested versions focused on trust, and moved part of the budget into search. The booking rate reached 5.0%. Several changes ran at once, so the proof is the pattern. I can't split the credit.
Test 120 ad variations without wasting budget
At The Cirqle, my team tested 120 ad copy variations for Color Street. We built them from a few hooks with many rewrites of each, and moved budget to the ads that sold. My mistake was spreading spend too thin across variants. Write fewer hooks, test them in smaller batches, and let purchases per spend decide.
Google and Meta run your ads now. Audit what they turned on
In late 2026 both big ad platforms do more of the buying by default. Google's AI Max matches searches beyond your keywords, rewrites ad text and picks landing pages. Its older features are auto-upgrading: campaigns using automatically created assets or campaign-level broad match from September 2026, Dynamic Search Ads from February 2027. Meta's Advantage+ automates audience, placements and budget. Meta's chief executive has said advertising could need little more than a goal and a payment method. What stays with you is the inputs (conversion signals and creative), the guardrails (knowing what's switched on) and the proof (experiments, because platforms report on their own work). A monthly audit covers all three, with AI doing the sorting and a person making every decision.
AI cold email that survives the spam filter
AI cold email works in 2026, but the rules flipped. Gmail and Microsoft now run AI spam filters trained on billions of emails that reliably catch templated, mass-generated outreach. Use AI to send more generic email and you get filtered, not replies. What works has three layers. First, set up the basics: SPF/DKIM/DMARC authentication, warmed domains, and roughly 35-40 sends per day per address. Second, target small checked lists tied to real buying signals (funding, hiring, leadership changes). Third, use AI for research and timing, not volume. Vendor-reported benchmarks suggest signal-based personalisation beats generic blasts several times over. Average reply rates sit around 3-4%, while disciplined signal-driven teams report 10%+.
How I cut weekly research from 4 hours to 1
I cut weekly competitor and account research from four hours to one by writing the steps down first, then letting Claude, n8n and Make do the reading. I checked every brief and struck any line with no source. The first version was too long to read, so I cut it to one page.
Your first automation: a weekly research brief
Automate a weekly research brief in four moves: a trigger, a list of sources, a prompt that writes one page, and a place where the team reads it. You can build it in n8n or Make. Write the steps down first, and check every line for a source before anyone acts on it.
How to test an AI agent before you trust it
Test an AI agent in layers. Rule-based checks on tool calls. Scoring of each step in the run. End-to-end task completion on a dataset built from your real work. An LLM judge for screening (checked against humans, never trusted raw). And expert review on the high-stakes cases. This playbook gives you the five steps, the judge rubric and the ways it goes wrong.
Deploy your first support-triage agent, human in the loop
Start with one narrow ticket category. Ground the agent in your own help content. Decide honestly what 'resolved' means before the vendor decides for you. And put a human review gate on every action that touches money or account state. This playbook gives you the six steps, the prompt and the ways it goes wrong.
How to use AI to review a contract before you sign it
To review a contract with AI, upload the full document to a capable model. Ask it to summarise the deal in plain English. Then run targeted passes for risky clauses, missing protections and anything that differs from a standard agreement. Treat every flag as a question to check, not a verdict. AI is a great first-pass reviewer that finds issues fast, but a human still makes the final call. Never sign based on AI output alone. Never paste a confidential contract into a tool without checking its data policy.
Use Claude to cut first-draft time with four prompts
Use AI for the mechanical parts of writing (outline expansion, transitions, headline variations, fact-checking). Keep the parts that matter (voice, argument, original reporting) human. The 4-prompt sequence below does this without the obvious AI tells most workflows produce.
Running ads now? Check your ad account against six ratios: type your numbers or load a Meta or Google Ads export, and nothing is uploaded.
More on using AI in marketing: all lessons and case studies.
Questions, answered
How do you lower CAC in telepsychiatry?
Start with the page, because that is where most paid visitors leave. At Gaba Healthcare, a telepsychiatry company, only 1.8% of visitors booked a consultation. I built a landing page for each patient condition, ran A/B tests on trust, and moved 40% of the Meta brand budget into high-intent search. The booking rate reached 5.0% and qualified calls went from 12 to 20 a day. The changes ran at the same time, so I can’t split the credit between them.
How long does it take to lower CAC by 20%?
For my consulting clients it took nine months. The cost to win a customer went from USD 100 to USD 80, on budgets under USD 1,000 a month. The funnel was fixed first, and spend was added after.
What is a good booking rate for a clinic landing page?
Unbounce’s 2024 Conversion Benchmark Report puts the median for medical practices at 2.4%, and the best medical pages at 20.4%. Gaba Healthcare went from 1.8% to 5.0% in 2023.
How do I know the page is the problem and not the traffic?
Look at who arrives and what they do. At Gaba paid ads brought plenty of visitors and only 1.8% booked. When people arrive and few act, suspect the page first. Check the targeting too, so you're not fixing the wrong thing.
How many ad variations should I test at once?
Fewer than we did. The team ran 120 over the campaign. Spread across that many, each one had too little budget to read. Start with a small batch, kill the losers, and write the next batch from the winners.
Do I have to use Google's AI Max?
It's an optional layer for Search campaigns. Google's help page says its features can be switched off one by one. But old settings are being upgraded into it: campaigns using automatically created assets or campaign-level broad match from September 2026, and Dynamic Search Ads from February 2027, per Google's June 2026 update. Check which features are on after any upgrade.