Blog / AI & search /

AI-assisted keyword research: where it goes wrong

Ask a chatbot for “20 keywords for a handmade soap shop, with monthly search volumes” and you’ll get a tidy table in seconds. Every row has a number. Some of them look very precise. “Organic lavender soap: 2,400.”

Where did that number come from? Nowhere you can check. And that’s the whole problem with AI-assisted keyword research in a sentence: the tools are brilliant at producing ideas and unreliable at producing facts, and they present both with the same confidence.

I use ChatGPT for keyword work, and I’ll keep using it. But after a year and a half of trying it on real sites, I have a fairly clear list of where it goes wrong. Here it is, with the checks I’d put around each one.

1. It invents search volumes

Unless a tool is connected to a real keyword database, a language model doesn’t know how many people search for anything. It has read a lot of text about keywords, so it can produce numbers that look like search volumes. They aren’t.

Some AI features inside SEO tools do pull real data from the tool’s own database, and that’s a different thing. The issue is the general chatbot asked to “include search volume”.

The check: never take a volume, difficulty score or CPC from a chatbot. Use it for the list, then put the list into Google Keyword Planner, Ahrefs, Search Console or whatever data source you trust. If you have no paid tool, keyword research for a small website with no budget covers the free options.

2. It guesses at intent instead of looking

Ask a chatbot whether “best running shoes” is informational or commercial and it’ll give a sensible answer. Sensible isn’t the same as correct for your market, today.

Intent is decided by what Google shows, and that changes. A query that showed blog posts a year ago might show product grids now, or a row of forum threads, or an AI answer at the top in the countries where those appear. A model working from old training data can’t see any of that.

The check: for any keyword you’re going to build a page around, look at the actual results page. What type of page ranks? Product pages, category pages, guides, videos? That decides what you build. I wrote about why this matters in search intent: why the same keyword needs different pages.

3. It clusters by meaning, not by results

This is the subtle one. AI is very good at grouping keywords that sound related. “Waterproof hiking boots” and “hiking boots for rain” obviously belong together. So do “how to clean suede boots” and “suede boot care”, probably.

But the right question for clustering isn’t “do these mean similar things?” It’s “does Google show the same pages for them?” Sometimes two phrases that sound almost identical return different results, because one is a shopping search and one is a how-to. Sometimes two that sound different share most of their top results.

The check: use AI for the first rough grouping, then spot-check the clusters that matter by comparing the top results. If most of the top five URLs overlap, one page can target both. If they don’t, you probably need two. My earlier workflow in using ChatGPT for keyword clustering still holds; this is the step I’d add to it now.

4. It doesn’t know your market

Ask for keywords for a furniture store and you’ll get a list that leans towards the US, because that’s where most of the training text comes from. “Couch” instead of “sofa”. “Dresser” where a British shopper might search “chest of drawers”. Prices in dollars.

For a store selling in Nigeria, the gap can be wider. People search in a mix of British and American English, use local product names, add city names, and ask about payment and delivery in ways a generic list won’t predict. I touched on this in local SEO in Lagos and Abuja: what the textbooks miss.

The check: tell the model the country and the language variant up front, and then trust your own data more than its suggestions. Your site search, your customer messages and your Search Console queries are written by your real customers. The AI list is written by an average of everyone.

5. It hands you a list that’s too clean

AI keyword lists tend to be tidy. Sensible phrases, nicely formatted, no typos, no odd questions. Real search data is messy: misspellings, half-sentences, brand names spelt three ways, questions you’d never think to ask.

That mess is where a lot of the opportunity sits, especially for small sites. The long, specific, slightly strange queries are often the least competitive and the closest to a purchase. A clean AI list tends to steer you back towards the obvious head terms everyone else is already chasing.

The check: pull real queries before you ask the AI anything. Search Console, Google’s autocomplete and “People also ask”, your store’s internal search, forum threads. Then use AI to organise and extend that list, not replace it. Keyword research with Reddit and forums covers one of the richest sources.

Where it genuinely helps

None of this means “don’t use AI for keyword research”. The places it earns its keep:

  • Brainstorming angles. “What problems might someone have before they buy a standing desk?” is a great prompt. You’ll get ten ideas, and three will be worth checking.
  • Expanding a seed list with synonyms, modifiers and related questions, which you then check against real data.
  • First-pass grouping of a long export, so you’re refining clusters rather than starting from nothing.
  • Mapping keywords to page types once you’ve checked intent yourself. It’s quick at the boring sorting.
  • Spotting gaps when you paste in your current page list and ask what’s obviously missing.

All of those are tasks where a wrong answer costs a few minutes of checking, not a page built around a keyword nobody searches.

A workflow that keeps it honest

  1. Gather real queries first (Search Console, site search, autocomplete, forums).
  2. Use AI to expand and roughly group them.
  3. Check volumes in a real data source.
  4. Check intent by looking at the results for the keywords that matter.
  5. Decide which page each cluster belongs to, and whether that page exists yet.

Steps 3 and 4 are the ones AI can’t do for you. They’re also the ones that decide whether the work pays off.

Next time a chatbot gives you a keyword table, pick the three rows you were most excited about and check them in a real tool. How many of the numbers survive?