AI Foundations — Your First AI Marketing Stack

Source: YouTube auto-generated captions, lightly cleaned (repeated caption lines removed, misheard brand names corrected).

0:00 — Wasted spend on zero conversions

One live account spent $925 over the last 30 days. $234 hit search terms with zero conversions. 93 distinct terms are negative keyword candidates. Average CPA sits at $66. A real search query report hands you that waste every cycle if you keep sorting it by hand. Opening the CSV, sorting by cost, and skimming for junk terms feels productive. It is slow, inconsistent, and easy to miss patterns across 93 candidates. A buyer under time pressure flags the obvious terms and leaves money on the table.

The fix is not a longer review session. It [clears throat] is a tiny AI stack

0:44 — Three pieces only

that reads the report the same way every time. The stack has three pieces and nothing more. An API key so the model can answer. One controlled API call so you trust the path. A search term CSV so the model works on your real queries, not a demo. When those three connect, the first useful job is simple. Mark the zero

1:07 — Why this order matters

conversion terms that should become negatives before the next spend cycle. Order matters because each step removes a failure mode. A key without a working call leaves you guessing at auth errors. A call without a real CSV teaches nothing about your account. Analysis before the path is proven wastes the report and your trust. Key, then one call, then the CSV keeps the first win small, measurable, and

1:32 — Step 1 — Get an API key

tied to the $234 already lost. Step one is getting an API key from the model provider you already chose. Create the key in the provider console. Copy it once and store it in an environment variable or a local secrets file your script can read. Do not paste the key into the CSV, a slide, or a shared doc. The key only exists so the next step can authenticate without drama. You get the key first because every later failure is clearer when auth is already solved.

If the call fails after the key works, you debug the request body or the file path, not permissions. If you skip this and jump to analysis, a bad key looks like a bad prompt. Senior buyers isolate variables. The key is the

2:21 — Step 2 — Make one API call

first variable you lock down before touching spend data. Step two is one API call. Send a short, plain language request through the official SDK or a bare HTTPS post using the key from step one. Keep the prompt trivial on purpose, like ask for three bullet themes in a sample string. You are not optimizing marketing yet. You are proving the wire works and that a response comes back in a shape your script can print.

One call teaches you latency, error format, and token cost before real account data is in play. If this call fails, fix headers, model name, or billing on a harmless prompt. Only when the response prints cleanly do you point the same client at a search terms file. That discipline is why the $234 of waste

3:12 — Step 3 — Load search terms

gets reviewed by a working path instead of a half-broken script. Step three is the search terms CSV from the account. Export search terms for the last 30 days with spend, clicks, conversions, and conversion value if you have it. Drop the file next to your script. Your code should read the rows, not a screenshot. This is the same report that showed $925 in spend and 93 negative candidates in the live account. Now make the first real analysis call on that CSV.

Pass a compact summary of high spend, zero conversion terms, and ask for negative keyword candidates with a short reason each. Tap the list so you can review it. The model should rank obvious junk above edge cases. You stay the buyer. The model drafts the short list

4:02 — Step 3 — First analysis call

from the same numbers that produced the $66 average CPA. Read the output like a media buyer, not like a dimmer audience. Keep terms that clearly miss intent. Park brand collisions and ambiguous queries for a second pass. Match each kept term back to spend so you know how much of the $234 you are about to cut. The stack is working when the short list is tighter than your last manual skim and still tied to cost. Turn the kept list into negatives in the platform with match types you already trust.

Start exact or phrase where the term is clearly wasteful and log what you added. The point of the stack is not a longer chat. It is a repeatable path from export to candidates to live negatives so the next 30 days do not repeat the same $234 of zero conversion spend.

4:54 — What the stack gives you

What you have now is a minimum AI marketing stack. A key, a proven call path, and one workflow that turns a search query report into negative keyword candidates. It is narrow on purpose. Breadth comes after this loop is boring and reliable. When the first call on a real report takes minutes instead of an afternoon, you earned the right to add creative or bidding jobs later. Before you widen the stack, lock three guardrails.

Never send full customer PII in the prompt. Never auto apply negatives without a spend threshold you set. Never skip the manual read on the first few runs while you learn the model's blind spots. These rules keep the 93 candidate list useful instead of dangerous and they keep the buyer in charge of live account changes.

5:43 — Recap

Recap the path in the order you built it. Get a key and store it safely. Make one harmless API call until the response prints clean. Export the search terms CSV, run the analysis call, and review negative candidates against spend. That sequence is how a $925 a month with $234 of zero conversion waste becomes a controlled first AI workflow. Run this on the next search query report before you add more tools. The first win is a clean negative list tied to real spend, not a bigger model.

When the loop is stable, extend it. The blog is on ahmeego.com. Buddy by Ahmeego, a company of It All Started With an Idea.