Field notes · SEO & Content · Mar 27, 2026 · 6 min

How I use ChatGPT for SEO: intent, metadata, audits

ChatGPT researches and edits. I write. The rewrite prompt failed first. What stuck: intent checks and one metadata framework across 3 regions. The final call stays mine.

3 regionsone metadata framework
6 slotsmetadata per page

Most conversations about “AI in SEO” start with the same idea: use AI to write more content. I tried that version first, and it failed. Anyone who has run a multi-region site knows the bottleneck is clarity and consistency, page after page. The site I ran as PingPong's Senior Digital Marketing Manager from 2023 to 2025 spanned North America, Europe, and Southeast Asia. Three regions, one editorial brain (mine), and never enough hours. That's where ChatGPT earned a place in my workflow, as a tool that works on the plan behind the page.

The version that didn't work

My first attempt was the obvious one. I pasted an underperforming page into ChatGPT and asked for an “SEO-optimized rewrite.” The output was fluent and interchangeable with every competitor page on the same query. It kept the keywords and dropped the substance, the compliance phrasing EU readers need, the payment-terms detail SEA buyers search for. I shipped none of it.

One rule survived that experiment: the model never writes the page. It checks intent, tightens structure, drafts metadata, and flags suspects.

Start with intent

When I optimize an existing page, my first step is never rewriting. I ask the model one question: what is the user behind this query trying to accomplish? That forces alignment, and pages miss the query more often than they read badly. The pattern I keep finding: a page written as a product announcement sitting under a query where people are asking “how do I.”

So I fix the mismatch first, before writing a single new paragraph.

Query vs page
Query intent“how do I …?”
Page deliversannouncement
Verdictmismatch
Re-align the page first
Answer the “how do I” first
New paragraphs added0
The most common failure is a product announcement parked under a “how do I” query. Re-aligning the page to the question lifts it without adding a single paragraph.

A structural editor

Long-form content fails on logic: transitions break, arguments repeat, the page never delivers what the headline promised. I use the model to surface these issues early, where a concept jumps too quickly and where a claim needs proof.

I still rewrite everything myself, now with a clearer map of what broke.

Metadata across three regions

Metadata was my biggest time sink, and the arithmetic explains why: every page needs a title and a meta description tailored per region. 3 regions × 2 fields = 6 pieces of metadata per page, before a word of body copy. I used to write each one from a blank cursor. The fix: I defined the structure once, then had the model draft inside it. This is the actual title pattern from my framework:

// one title framework, three regional slots [primary keyword]: [outcome the region cares about] | [brand] NA: outcome = speed ("paid in days") EU: outcome = compliance+FX ("SEPA-ready, no hidden FX") SEA: outcome = market access ("collect from US & EU buyers") meta: 1 sentence matching intent + 1 sentence of proof, ≤155 chars

The model fills the slots; I fix the nuance by hand. Every page then reads like one editor wrote it, and I drafted faster.

NAspeedEUcompliance + FXSEAmarket accessone framework, three slots, six fields a page
Six slots, one template, and the model only fills in the blanks.

Diagnosis: the content auditor

The audit habit started with a dip I misdiagnosed. An EU page slid, my first theory was links, and I spent a week looking in the wrong place. When I finally pasted the page plus its target queries into the model and asked where reader expectations weren't met, the answer was more boring: the H1 promised a comparison and the page never delivered one. Readers bounced; rankings followed.

That's the routine now. Traffic dips, the model flags the suspects, questions the page never answers, places it stays shallow, points it repeats, and I verify which suspect is real. The loop shortens the distance between spotting a dip and knowing why, which is where most audit time goes.

The audit loopfrom noticing a dip to naming the causeTraffic dipan EU page slidesPage and queries inpasted with the target queryModel flags suspectsquestions the page ignoresHuman verifieswhich suspect is realRoot causeH1 promised a comparisonthe model lists suspects and never decides which one is trueThe first theory was links, and a week went into the wrong place. The confirmed answer took an afternoon.
The model lists suspects. I check them one by one. Judgment stays at the verification step.

Systems are the compounding asset

The biggest long-term advantage was the libraries I built with the model: outline frameworks, metadata patterns (above), how each region talks, internal linking rules, and editorial rules for tone. Two examples, so this isn't abstract:

  • The internal-linking logic is 3 rules the model applies as a checklist: every new post links up to 1 pillar page, sideways to 2 siblings on the same intent, and never to a page targeting the same query. It is mechanical, so writers skip it at volume unless a checklist enforces it.
  • The most-reused prompt in the library: “Here is the page and the query. List every question a reader with this query expects answered, then mark which ones the page answers, partially answers, or ignores.” That one prompt is most of the audit.

My part was the structure; the writers owned the words. Those assets kept the pages aligned as volume grew, without flattening the nuance each market needs.

The model shortens the audit loop; matching reader intent is what moves the rankings.
the whole thesis, in one line

What stays human

I own positioning, market context, compliance-sensitive wording, final editorial quality, and any claim that needs business validation. The model drafts and flags. I decide what the page says.

Safe to hand the model: outlines, title and meta variations, internal linking ideas, content audits, competitive gap reviews. Not safe: anything a regulator, a customer, or a subject-matter expert would need to sign off on.

The order I keep on any site: pick 1 underperforming page. Ask the model what the user behind that query is trying to accomplish, compare that against what the page delivers, and fix the mismatch before touching anything else. Then write the metadata framework down and make the model draft inside it. The body copy comes last.

Suggested posts