Prompt Engineering Tips for SEO Keyword Discovery: The Five Principles That Actually Work

I used to think keyword research meant typing “keywords for [topic]” into ChatGPT and copying whatever came back. I did this for almost a year before I noticed something. The keywords always sounded fine, they were grammatically correct, they matched the topic, and almost none of them ever ranked or drove real traffic. That is when I started treating prompt writing as its own skill, the same way you would treat writing a proper brief for a freelancer.

This article covers two things I get asked about a lot from readers of Peaceworlai.com. First, the actual prompt engineering techniques I use to pull real SEO keyword ideas out of an AI model, the kind that reflect what people genuinely type into Google, not generic filler. Second, the five basic principles behind prompt engineering itself, the rules that make any AI model, whether that is ChatGPT, Claude, or Gemini, give you sharper and more useful answers on any task, not just keyword research.

If you run a small business in Lagos, Nairobi, Accra, or Freetown, or you are a student trying to get more out of free AI tools on a tight data budget, this is written for you as much as it is for anyone in London or Toronto. The tools are the same everywhere. What changes is how you talk to them.

Why Most Keyword Prompts Fail Before They Even Start

The mistake almost everyone makes, myself included in the beginning, is asking the AI for keywords the way you would ask a search engine. “Give me 20 keywords for a bakery in Abuja” produces a list that reads like a list, generic phrases like best bakery, bakery near me, cake shop, things you already knew before you typed the prompt. The AI is not wrong, it is just answering a shallow question with a shallow answer.

What actually works is feeding the AI context first and asking it to think like a searcher, not a thesaurus. Here is the difference in practice.

Prompt Engineering Tips for SEO Keyword Discovery That Actually Work

1. Paste your competitor's headings before you ask for anything

Before asking for keywords, I copy and paste the H1 and H2 headings from two or three competitor pages that already rank for my target topic, then ask the AI to compare them against my own page outline and point out topics I have not covered. This one habit alone has found me more usable keywords than any generic brainstorm.

Swipe file prompt: “Here are the headings from three articles currently ranking for [topic]. Compare them to my planned outline below and list the subtopics or angles I am missing that searchers clearly care about. Then turn each missing subtopic into three to five keyword phrases someone would actually type into Google.”

2. Ask the AI to think like a worried customer, not a marketer

Generic keyword prompts get generic marketing language back. Specific, emotional framing gets you the actual words people search when they have a real problem. Instead of keywords for a laundry detergent business, I ask something closer to a real situation.

Swipe file prompt: “A mother in a small town is doing hand laundry and her clothes are not coming out clean. Write 15 things she might type into Google to solve this, in plain everyday language, not marketing language.”

This produces phrases like “why is my laundry still dirty after washing” and “how to remove sweat stains from white shirts by hand,” things a real person searches, not things a copywriter would write.

3. Ask for clusters organized by buyer stage, not a flat list

A flat list of 30 keywords is hard to use. What I actually need is keywords grouped by where the person is in their decision, someone just discovering a problem searches differently than someone ready to buy. I now always specify this in the prompt itself.

Swipe file prompt: “Group these keyword ideas into three stages, people who just realized they have a problem, people comparing solutions, and people ready to buy. Give five keywords per stage.”

4. Layer in local language and context for African search behavior

This is the one most guides skip, and it is the one that matters most for readers outside the US and UK. Search behavior in Lagos is not identical to search behavior in London, even for the same product. People search in a mix of English, pidgin, and local phrasing, and price sensitivity shows up in the keywords themselves.

Swipe file prompt: “Rewrite these keywords the way someone in Nigeria, Kenya, or Sierra Leone might actually search, including any common pidgin or informal phrasing, and add price conscious variations like cheap, affordable, or cost.”

I ran this exact prompt for a client selling phone accessories in Freetown and it surfaced “cheap phone case Freetown” and “affordable charger that no go spoil quick,” phrases their previous keyword list from a paid tool never once suggested.

5. Turn questions into keywords with the reverse question method

Ask the AI to generate the questions a total beginner in your niche would ask, before you ask for keywords at all. Questions convert into long tail keywords almost automatically, and they tend to match what people actually type when they are early in their search.

Swipe file prompt: “List 20 questions a complete beginner would ask about [topic]. Do not explain the answers, just the questions, in the exact words a beginner would use.”

Once you have the questions, each one becomes a keyword phrase and often a subheading for your article at the same time.

The Five Basic Principles of Prompt Engineering, for Any AI Model and Any Task

Keyword research is just one use case. These five principles are the ones I lean on for everything, writing product descriptions, drafting emails, building lesson plans, whatever the task is.

1. Say exactly what outcome you want, not just the topic

“Write about productivity” gets you an essay. “Write a 200 word Instagram caption encouraging small business owners to plan their week on Sunday night, in a warm and slightly cheeky tone” gets you something you can actually post. The AI is not guessing your intent, you are removing the guessing.

2. Give it a role and an audience

Tell the model who it should sound like and who it is talking to. “You are a patient teacher explaining this to someone using AI for the first time” produces a completely different answer than no framing at all. I learned this the hard way after getting three overly technical explanations in a row for a client who just needed simple, friendly copy.

3. Show it an example of what good looks like

If you have ever written or seen a piece of content that hits the tone you want, paste a short sample into the prompt and say match this style. This is called few shot prompting, and it consistently beats describing the tone in adjectives. Telling an AI to write casually is vague. Showing it three casual sentences is not.

4. Break the task into steps instead of asking for everything at once

Asking for a full blog post, keywords, meta description, and social captions in one prompt usually gives you a rushed, average version of all four. I get better results asking for one thing, reviewing it, then asking for the next thing based on what came before. Slower, but the quality difference is real.

5. Treat the first answer as a draft, not the final answer

The biggest mindset shift for me was accepting that the first response is a starting point, not a finished product. I almost always follow up with something like make this shorter, this section feels generic, or remove the corporate tone. Two or three rounds of refinement consistently beat trying to write the perfect prompt on the first try.

Anthropic's own documentation on prompt engineering makes a similar point, that being specific and giving the model room to work through a task in steps tends to outperform a single vague instruction, which matches what I have found running this on real client work. You can read their official guidance at docs.claude.com if you want to go deeper on this.

What This Actually Looks Like for Small Businesses and Students in Africa

Data cost and time matter here in a way that guides written for a US audience rarely account for. A few adjustments I make for African users specifically.

• Keep prompts tight and specific the first time, fewer back and forth messages means less data used on mobile

• Ask for outputs in a format you can copy straight into WhatsApp Business or a Facebook post, since that is where a lot of small business marketing actually happens here, not on a polished website

• If you are a student, use the five principles above for assignments too, giving the AI your course level and the exact marking criteria produces far more useful study answers than a bare question

• A trader in Freetown selling fabric does not need SEO keywords in the technical sense, she needs the exact words her customers type into Facebook search or Google when looking for ankara or lace, and the prompts above work the same way for that

A Few Mistakes I Have Made So You Do Not Have To

I once ran a keyword prompt without specifying location and built an entire content plan around search terms that were popular in the US but barely searched at all in the market I was actually writing for. Always specify location, even when it feels obvious.

I also used to accept the first batch of keywords without checking search intent, only to write an article answering a question nobody who searched that phrase actually had. Now I ask the AI directly, what is the searcher likely trying to accomplish with this exact phrase, before I commit to writing anything.

For search volume and competition numbers once you have a shortlist, a free option worth knowing is Google's own Keyword Planner tool, which is still one of the more reliable places to sanity check whether a phrase is actually searched enough to bother targeting.

Frequently Asked Questions

Can I use these prompt engineering tips with any AI tool?

Yes. These principles work with ChatGPT, Claude, Gemini, and most other AI chat tools, since they are about how you structure the request, not a feature specific to one platform.

Do I still need a paid keyword tool if I use AI prompts like this?

Not always. For small businesses and students starting out, AI prompts can get you most of the way there for free. Paid tools become more useful once you need actual search volume numbers to prioritize between keyword options.

How many keywords should one prompt generate?

I usually ask for 10 to 20 at a time, grouped by stage or intent, rather than one long list. Smaller, organized batches are easier to actually use.

Is prompt engineering something only tech people can learn?

No, and this is worth saying clearly. Prompt engineering is closer to learning how to give clear instructions than it is to coding. Anyone who can write a clear WhatsApp message to a colleague can learn this in an afternoon.

Will these keyword prompts work for local languages like Pidgin, Swahili, or Twi?

Yes, as long as you specifically ask the AI to respond in that language or dialect. It will not do this automatically unless you tell it to.